AI on the Ground: Realistic Use Cases for Market Systems in Fragile Contexts
Event: 19 August 2026
Editorially reviewed transcript, 16 October 2026

AI disclosure and limits: generated locally with MLX Whisper large-v3-turbo, then reviewed with AI assistance for names, obvious recognition errors and speaker turns. Selected unclear passages were re-transcribed locally with the turbo and medium models. This is an editorial review, not a complete listening-based verification or a fact-check of the speakers' historical claims. No human approval is implied.

Speakers: Karri Goeldner Byrne (labelled Karri Byrne), moderator. Thomas Byrnes (labelled Tom Byrnes) and Emil Heidkamp, speakers. Blerta Peta, organiser and Q&A contributor.

Speaker labels are editorial attributions based on introductions, handovers, content and selected on-screen name tags. Brief overlaps that cannot be assigned confidently are labelled Speaker uncertain. Square brackets mark editorial clarification, tentative readings or unresolved wording. Original spoken claims and figures are retained, without presenting them as verified facts.

Names and spelling were checked against the organiser's email, circulated event summary and the video's on-screen labels. Terminology and British spelling have been normalised. The closing section was recovered from an isolated recheck because the first pass had produced repeated thank-yous. Original automatic TXT, JSON and SRT files are preserved alongside this version.

Source: https://marketsincrises.net/all_events/ai-on-the-ground-realistic-use-cases-for-market-systems-in-fragile-contexts/
Detailed correction record: 2026-08-19_MiC_AI_on_the_Ground_Editorial_Review.json

[00:00:02.700 - 00:00:57.440] Karri Byrne
Hi everyone. Welcome. Glad to have you joining us today. We're going to give just a few minutes for everybody to get settled in the room. While we're waiting, feel free to introduce yourself in the chat box. Let us know where you're calling in from or calling, you know what I mean, zooming in from today. But we'll give just a few minutes. We had quite a large sign up. So as people make their connections and I see the numbers counting up, I just want to give time for that to all happen. But we'll get started in just about one minute. So if you haven't already gotten a cup of coffee or a glass of water to keep you company for this, go ahead and do that now. There's probably still time.

[00:01:12.600 - 00:02:02.940] Karri Byrne
All right. I said I was going to give it another minute, but I feel like the numbers are maybe slowing down now. So I'm going to go ahead and get started. I can see that many of you have introduced yourself in the chat. Thank you. That's great. It helps us learn where everybody's coming from. So thanks for joining us, everyone. My name is Karri Byrne and I am a member of the Markets in Crises Advisory Committee. So we're here today to talk about using AI in market systems in an appropriate and effective way. Before we get very far in, I really do want to say a thank you to PRO-WASH & SCALE.

[00:02:03.600 - 00:03:00.640] Karri Byrne
PRO-WASH & SCALE is a U.S. government initiative to strengthen the design and implementation and overall effectiveness of food security and wash interventions. And they're co-hosting this event today with MiC. So we're really grateful for their support. And I am super excited by our speakers today. I think all of us are probably feeling a little overwhelmed by AI and maybe under-informed on how we can use it responsibly. So what I'm looking forward to today in our discussion is talking about that some more. We know at MiC that people want to use AI to make their work more efficient. But we also have a responsibility to use it well, right, to make sure that the information that we're gathering about vulnerable populations is protected.

[00:03:00.640 - 00:03:50.800] Karri Byrne
And that our market assessments, that the data is correctly represented. And it's not just this like pretty report that kind of misses the key issues. So our speakers today have some great examples and experience. They are using AI for market systems work in fragile contexts. And they're going to tell you a bit first about what they do. And then we'll go into some of the questions that you all submitted. Thank you. Great questions, guys. And then towards the end, we will have some Q&A. So feel free to put your questions in the chat as we go along. But honestly, we're not going to get to those questions until towards the end of this session.

[00:03:50.800 - 00:04:01.320] Karri Byrne
We'll focus first on the ones that you all pre-submitted. And just to say, this is for our speakers, too. I am anticipating there will be a lot of questions.

[00:04:01.560 - 00:04:05.200] Karri Byrne
So [early warning], I'm going to be strict about some of our time limits.

[00:04:05.200 - 00:04:58.900] Karri Byrne
I know both our speakers have just so much experience and information to share. Everyone forgive me for cutting people off in advance. I'm going to try to be strict on time. So with that, let me introduce the speakers. Tom Byrnes is the CEO of MarketImpact Digital Solutions and runs AidGPT, which is a responsible AI training programme for humanitarian and development professionals. He's been working for 15 years in humanitarian operations. So he knows the sector well, including work founding the Cash Consortium in Yemen and working with the Danish Refugee Council as a regional economic recovery coordinator. Emil Heidkamp is the founder of Parrotbox, which specialises in coaching, performance support,

[00:04:58.900 - 00:05:40.060] Karri Byrne
and simulation-based training. And he has developed some really interesting, both human-led and AI-enabled learning and support solutions for training humanitarian organisations. He has worked with IFRC, PSI, and the CGIAR Consortium. So you can hear there's so much experience across so many different elements of this area of work. As I suggested, I'm going to give our speakers a chance to kind of talk about their work first. Tom, why don't you kick us off? Tell us what you've been up to.

[00:05:41.120 - 00:06:16.680] Tom Byrnes
Thank you very much, Karri. And it's really great to be here. And I'll try and keep it as concise as I can, though. Anyone knows me, I'm not very good at that. So what have I been doing? One of the key things at my company and MarketImpact we're doing is still continuing to support market analysis, crisis analysis. We recently did work for Mercy Corps Crisis Analysis Unit on the Strait of Hormuz, which you may have seen I shared in a LinkedIn post and to the MiC group earlier this week, an example of that there. And we've also been working and providing training and advice to agencies and governments

[00:06:16.680 - 00:06:49.320] Tom Byrnes
on how to use AI, when not to use AI, what are the opportunities and the capacity that they need to do so. But I want to be very clear and upfront about one thing. I am not actually pro AI. If I personally, I think AI is terrific, as in it's terrifying. And if I could close the Pandora's box and put it back in again, I probably would. But we don't get to get that choice. And that's a key thing. If we were, these tools are out there, they are available, our teams are using them. And we also don't get the choice.

[00:06:49.420 - 00:06:56.320] Tom Byrnes
If I was working in a Walmart and I was responsible for just putting baked beans on a shelf, then I could make an ethical decision, I'm not going to use this.

[00:06:56.500 - 00:07:01.420] Tom Byrnes
But for someone in [place name unclear] or someone in the field who needs to make a very quick response,

[00:07:02.340 - 00:07:41.020] Tom Byrnes
choosing not to use this very powerful tool is like choosing not to use Excel. And that's part when we do trainings. That's one of the key things that we do. ChatGPT, Claude, Gemini, these tools are like Excel. They're incredibly powerful tools, but useless unless someone shows you to know what a formula is. So you need to learn different skills to engage that. And also like Excel, you know, the entire world's financial system runs on Excel. But unless you know how to use it, you can cause major issues. Most of the crises I've been in operationally are because someone hasn't updated a formula correctly,

[00:07:41.500 - 00:07:53.200] Tom Byrnes
or has pulled information incorrectly into a sheet. So that's what we're trying to go there. We also want to talk about the risks. I mean, through that we work to, and the risk is not the invented citation.

[00:07:53.500 - 00:07:54.620] Tom Byrnes
[Albeit] it is a true risk.

[00:07:55.040 - 00:08:27.460] Tom Byrnes
The risk, and particularly for Markets in Crises and market analysis, is the fifth village. It's when your teams went and assessed 20 villages and identified five as needing life-saving assistance. That got put into an Excel spreadsheet. That then, or a long report, that was synthesised by someone at an area office using AI to quickly turn that around. And when it did it, it dropped one of those villages off. And that means that that time of that field team is wasted, but also that village will never get assistance. Because now the official record is that village doesn't need assistance and that they are fine.

[00:08:27.780 - 00:09:03.200] Tom Byrnes
So this is where harm can get into the system, not just inclusion, but also omission. And that's what we try and explain here. And what we try and do, and what I've been trying to try and push, is good ways of doing this, but also building the practice. We need to concede and be honest and disclose when we're using AI, so that when we are using these tools to do work, which we need to do, that it's clear for the next person in line, how did I use AI just to format this? It's perfectly fine for a national NGO or someone who is English is not their native language to use AI to help write this better.

[00:09:03.400 - 00:09:36.300] Tom Byrnes
But not if did the AI make the decisions? Did it come up with the ideas? And that's also very hard skills that we're going to have to try and learn. We need to learn because we can't do it all ourselves. We need in the JMMI and other tools to trust that the data that's coming from multiple sources that we bring together helps us understand that full picture and understands what's happening in the market systems. And we know that people will be using these tools and they should be using those tools. And so that's the key challenge that we're trying to work on by getting information out there and by doing trainings on the topic.

[00:09:36.560 - 00:09:40.280] Tom Byrnes
I'll go a lot longer, but I'll throw it back to you now, Karri, on that point there.

[00:09:42.560 - 00:09:46.720] Karri Byrne
Great. Thanks, Tom. Emil, what have you been up to?

[00:09:47.780 - 00:10:23.640] Emil Heidkamp
Well, I suppose we have a nicely balanced panel here because I am very much in favour of the use of AI for humanitarian and development purposes. I think there's, yes, there are lots of legitimate risks. And we've probably heard a lot about those risks and, you know, especially within our own sectors, you know, conversation. But we always have a positive ethical obligation to use whatever tools we have available, whether it's medicines. I mean, I ride a bicycle everywhere, every day around town. I do not fault the IFRC for deploying helicopters and pickup trucks in an emergency. So it's really about how we use AI.

[00:10:23.920 - 00:10:53.380] Emil Heidkamp
And I think we need to make a distinction because a lot of people, when we say, I do agree with Tom about one thing. Just going to a generic chatbot like you have on the ChatGPT or Claude website and asking it questions, you know, critical questions and expecting it to give you an oracular answer is irresponsible. And I actually do have an example later that speaks to the fifth village, like what if the AI overlooks something. But I think part of this, though, is like a lot of us, we honestly have the wrong idea about what AI is and what it does.

[00:10:54.240 - 00:11:35.800] Emil Heidkamp
People imagine because it gives such eloquent answers that when they ask a, you know, a general purpose chatbot, you know, a question that it's going to go like a librarian into the stacks, do thorough research, come back to you with a pile of books and give you something for definitive sources. When, in fact, it's more like a generally knowledgeable colleague that reads at light speed and studied an absurd range of subjects at university. But when you go and say, hey, colleague, should farmers in country X be producing sunflower oil, you know, as supplemental income? And then, you know, you basically get it's the same thing like walking up to them at a party and they go, well, I vaguely remember an FAO report saying seed oil prices are down.

[00:11:35.800 - 00:12:12.100] Emil Heidkamp
And I don't know where, but a couple of places I read that sunflower, though, they are an excellent rotation crop for farmers. So and then it's going to give you an answer, which might be qualified and measured or might be overconfident. And so the question is, well, does that make AI a party trick? No, you can take that same colleague. Forget the chatbot context. Take that same engine, that AI model. Bring it to the library and say, hey, knowledgeable friend, I got some data for you to analyse here. Put the data in front of them and then say, use your light speed reading and synthesis powers to, you know, according to this, you know, evaluation framework I'm giving you, you know, produce some guidance.

[00:12:12.280 - 00:12:33.960] Emil Heidkamp
So I think we need to think beyond the lazy, oh, hey, the free chatbot option and start thinking about building proper systems. I mean, maybe you've heard of the company Palantir. I don't share their politics at all, but I do agree with them that that it's not just about the model. It's about building systems that can bring the proper data to models to give better answers. And we've got to build those kind of things in a humanitarian development context.

[00:12:35.040 - 00:13:25.240] Karri Byrne
I like that. I like that. Well, if it's okay with you guys, I'm going to just dive into some of the questions here. I can tell we're going to have a lively discussion, and I'm excited about that. Tom, I'm going to kick off with you because you've had a couple LinkedIn posts that I think were really interesting. One of them explained your experience when the output, the market systems report, looked really good. It was beautiful, but was not grounded in reality. And you kind of gave an example just a moment ago that touches on that a little bit. But what would be your top three tips for getting it right, for going from I've got the data to when I'm doing a market systems assessment report,

[00:13:25.820 - 00:13:26.220] Karri Byrne
get right?

[00:13:27.260 - 00:13:33.820] Tom Byrnes
Okay, so I'm going to give three tips and then one uncomfortable tip or comment that goes with it there.

[00:13:34.040 - 00:13:34.380] Karri Byrne
All right.

[00:13:34.720 - 00:14:05.940] Tom Byrnes
The first key tip is you need to ground it before you start. It's what Emil was saying. But even when you're creating a chat, you need to give it the context. You need to first tell it who it is, but you need to give it the context. So don't open a blank chat. We put the entire EMMA guidance into a workspace first and told it it was role playing as an expert of the EMMA guidance. This way it knew. In the same way that a coding agent would know what the language is, it said this is the task. It knows all the different steps.

[00:14:06.020 - 00:14:38.460] Tom Byrnes
And we have a really detailed frame there. This was really powerful. Even when I wasn't doing a full EMMA because it was working out, it was taking the lessons learned and the framing from that. So it knew what role it was playing. It knew what rules it could not break. And EMMA already sets out good assessment contains. So the machine never had to invent the shape of a document. The second key thing was to make every claim carry the evidence as it is written. That was the key instruction and key working there, which is no claim gets written without a tag.

[00:14:39.460 - 00:15:17.100] Tom Byrnes
Every claim, if it finds the claim, you need to go and then find two independent sources to back that up. You know, telling it in that way and running it so that you make sure that every single thing is cited, citable, and ideally from multiple angles. This gives reasoning rather than evidence. And it says, you know, make sure that what is being inferred is very clear. So that when I read it through, I can quickly go and see what it is actually coming from and I can pick out the issues there. And that's the key thing. And not just one source actually saying, go find evidence of this from multiple sources.

[00:15:17.260 - 00:15:40.540] Tom Byrnes
And we're only treating this as credible if we can have multiple independent sources picking that up. The third thing is check it. And check it twice with different questions and not just in a session that wrote it. What I mean by that is, first of all, especially as a final output, you need to read it. If you can't be bothered to read the output of your AI, why should anyone else read the output of your AI?

[00:15:41.000 - 00:15:44.340] Tom Byrnes
Give it to a different [model]. Don't ask [it] to mark its own work.

[00:15:44.560 - 00:16:18.780] Tom Byrnes
Give the, say, go, your job is, you're a red team agent. Your job is to take this and break it. Find where this is wrong. That is important. And that allows you to go and find where you've made mistakes. And that's what we do to refine our methodology. Because we go, why did we break this? Why did it not work? How can we fix that by doing that? And the fourth uncomfortable point is that you have to know the shape of the answer. You can prompt an AI to be an expert. I can prompt an expert AI to say, you are a global expert in public health.

[00:16:19.440 - 00:16:45.500] Tom Byrnes
But the problem is, I'm not an expert in public health. And so I can't actually validate its work. So at some point, you've still, I was doing this as a desk review. But I still had to go and check it with someone actually in Somalia. Like, hey, does this make sense? Because I'm not actually an expert. Or at least, you know, I can only treat it with a certain level. In this case, I was working with Mercy Corps and we had a country team. So it was only needed to be a first draft for them and that they would then be able to come back on.

[00:16:45.660 - 00:17:11.380] Tom Byrnes
But that's the important thing. Sound like you're an expert, but you're in an area you're not. But you can't do verification, truthfully, on something that you aren't actually an expert on. You can get the AI to help teach you that topic if need be. But you still have to use the Mark 1 eyeball and the Mark 1 brain, ultimately, before you put your name onto something and give credibility that you understand what it's done. Anyway, I'll stop there and throw it back to you.

[00:17:12.280 - 00:17:42.360] Karri Byrne
Great. Those are great tips. So ground it in advance, giving it that context. Do you have a source or two behind it? Check it. Check it twice. That's where I feel more confident that humans are still going to be needed. And I feel like three and four have a lot of similarities because there's checking it, but there's being knowledgeable when you're checking it, right? Have the right people check it. Anything you'd add to that, Emil?

[00:17:43.500 - 00:18:19.920] Emil Heidkamp
Yeah, I guess there's also, it's interesting, Tom brought up Excel before. I mean, so I'm sure we've all been in a situation where either maybe we are that person sometimes or you've seen that person who uses a spreadsheet to do the job of a database. You're like, oh, my goodness. This is, no, no, no, no, no, no. This is way too structured a data set. You can't, like, analyse an entire, you know, country's, you know, livestock feed supplies on an Excel sheet. Your goodness, you know? Or maybe your data person, your information management person said that to you. So, it's kind of, imagine, like, the public chatbots, like, you know, where you go in and you prompt it and you give it documents.

[00:18:20.160 - 00:18:55.320] Emil Heidkamp
Imagine those are like a spreadsheet. They're great for really quick ad hoc analysis. And if you have a novel problem and you don't, but it just needs solved, then absolutely give it the guidance, give it instructions, give it supporting documents, have it give you answers. But if you're going to be doing something repeatedly at scale, like, say, we have to analyse focus group discussions from 300 villages, you know, like we were saying. Well, then you might want to start going, okay, maybe I need a better tool. Maybe I need to collaborate with some people to build a system. And maybe that system comes, you know, infused with some of the expert criteria.

[00:18:55.380 - 00:19:28.680] Emil Heidkamp
So, like, we built a thing that helps with social service case management. And, but it's like, it's like, these are things where there's a horrible backlog in the social service system. Social workers would love to spend more time with directly with their clients. We are helping them eliminate their, you know, paperwork to do that. But when it came to interactions, when I was like, oh, someone, you know, you know, it has a quick job interview and it says, hey, how did it go? We, what we did is we actually gave it like all the criteria a social worker at an employment services office would use to evaluate someone's, you know, interaction.

[00:19:29.180 - 00:19:56.380] Emil Heidkamp
And then we fed it a lot, but then we independently had human social workers doing it. And then we got to a level where it was achieving 95, 98% alignment and none of the deviations were catastrophic. Then we're saying, okay, this is a system we can trust to do a large volume and then have human spot check it. So it's different if you're doing a bespoke paper, you know, one time versus you actually have to automate a huge process. So I guess.

[00:19:56.660 - 00:19:56.720] Speaker uncertain
Yeah.

[00:19:56.920 - 00:20:48.380] Karri Byrne
And that's the, and that becomes one of the issues where the funding tends to be one-off, one-off project led. And so I don't know that we, you know, with the recent news about FEWS NET having to, to dial back it's FEWS NET and JMMI and some of these others that you would expect to have maybe longer term analysis structures. I, I worry that we don't have the funding resource to make that happen in our, our current context. Um, because what, what I see, we've, we keep talking about Excel and, um, and one thing I, I see is that I feel much more confident analysing, using AI to analyse a lot of numbers.

[00:20:48.380 - 00:21:32.180] Karri Byrne
And then I analyse what the numbers mean, then I do having them analyse qualitative data. And I have used it to, um, uh, to, to analyse qualitative data and it did save me time and that was great. But even as I read through the analysis, I could see what it missed. And I felt like I had to go back very, very much to Tom's point of like, see how you can break the report. Um, what do you. Um, are there ways to reduce the burden of analysing large amounts of qualitative data without losing that nuance or, or is that just the human kind of value add?

[00:21:32.460 - 00:21:38.620] Karri Byrne
What do you, what do you, because context is so important in our work. Um, yeah, I, I don't know.

[00:21:38.660 - 00:22:14.220] Emil Heidkamp
I just sent a share request. I have actually have a chart that speaks to this and to the fifth village problem that was raised earlier. Um, is that, um, you know, actually LLMs, large language models actually excel at analysing language and qualitative data. And, and relative to that capability, they're actually kind of dodgy at math. I mean, you can hook them up to calculators. Which they do in the background of the public websites. But, uh, but, you know, but what I'd say is like, the problem is how we use it. So what we do is we say, Hey, take a huge volume of data.

[00:22:14.440 - 00:22:51.260] Emil Heidkamp
Like, you know, if we took actually 300, I don't know if it'll allow you. If you took 300, uh, notes from focus group discussions. And dump them into, you know, just the input box of ChatGPT and said, okay, review these in aggregate. Remember we said it plays a connect the dots game between your input. It's going to kind of try to take in that big aggregate data set, uh, connect some of the recurring themes. And then go, uh, I don't know. And synthesize an answer, you know, according to whatever effort level you said to have it take. And that's where you get the big general, the general answer that misses all the nuance drops off the fifth.

[00:22:51.840 - 00:23:23.040] Emil Heidkamp
Well, that drops off the fifth village. Now, a more responsible thing to do would be, okay. Um, have it go one at a time for every one of those 300 things. Have a, have a framework distilled. I want this criteria. I want to know what the, if we're doing access to healthcare, uh, what were the issues the community members raised or the points of agreement? What were the dissenting opinions? Repeat 300 times. And then have it tagged by village, by region, summarise. And then eventually you get this pyramid of summarisations, but with the audit trail preserved going down.

[00:23:23.240 - 00:24:01.100] Karri Byrne
So I'm going to stop you, Emil, because I feel like everybody on this call, at least I have, everybody's sort of freaking out in this moment. Like, I don't know how to do that. I don't know how to do that. the thing you just described. Tom, do you have an example of the market assessment that you did recently? Did it do, did it do something? So again, what I asked is how can we structure it? Where, how can we, are there ways to reduce the burden of analysing qualitative data? Because that's something we do in all these market systems of it without, um, losing that nuance.

[00:24:01.100 - 00:24:08.240] Karri Byrne
Did you have experience from yours that doesn't scare me maybe as much as the one, the example Emil just came up with?

[00:24:09.840 - 00:24:13.380] Tom Byrnes
Well, yes. I mean, not from that direct example.

[00:24:13.480 - 00:24:16.300] Karri Byrne
Not that you're not right, Emil. I know that's the right thing to do.

[00:24:16.400 - 00:24:17.340] Emil Heidkamp
We do it. We do it this way.

[00:24:17.340 - 00:24:18.740] Karri Byrne
A lot of us are scared and not there yet.

[00:24:18.920 - 00:24:28.460] Emil Heidkamp
We do it this way. We do it with interns and volunteers and sometimes overpaid experts. So we're doing it with humans right now. If we're doing it. Or other than that. I mean, who's summarising the focus group data? So, you know.

[00:24:29.140 - 00:24:34.380] Tom Byrnes
Yeah. So I think the key thing. So the first key step is not to just give it one goal.

[00:24:34.620 - 00:24:35.100] Speaker uncertain
Right.

[00:24:35.200 - 00:24:38.400] Tom Byrnes
Break the project out into the actual individual steps.

[00:24:38.720 - 00:24:39.120] Speaker uncertain
Right.

[00:24:39.180 - 00:25:16.480] Tom Byrnes
Because if it's just read all these transcripts, make me a summary and then create a report. You're asking it to do three or four different things. And you're not necessarily prompting. Now, you could probably do that with a orchestration system that Emil is indicating, which you built. All and even if you've set up a project and setting, you still want to break it down to steps. And the first step would be, let's review all these transcripts and first identify what's the analytical frame or how we would want to go about doing this. You know, let's do it once. Let's do it once manually and try and work it through and say this.

[00:25:16.800 - 00:25:47.360] Tom Byrnes
OK, this works. This way of asking it. This way of coding it. These are the key things I want to build out. You prototype a little bit. You know, try a few chats back and forward. See the output of what it's done. And then you get a level of confidence that, yes, OK, this is actually now working. I am. I'm asking it the right questions. I'm not assuming it's got knowledge. And now I've got an approach that I can replicate. And at that point, you can replicate that and you can use it multiple times. This is often being used now more for in third party monitoring.

[00:25:47.460 - 00:26:19.280] Tom Byrnes
I know a lot of companies or UN agencies are now asking their third party monitors. I want you to give me all the transcripts and all the focus group discussions because I want to be able to sense check that your final report actually matches what is being said in the report that you're doing here. You know, I want to be able to hold you accountable and I want to be able to see those things. Also, though, crucially here, if we're talking about focus group discussions, key informant interviews or unstructured data, did the person who give you that data give you the consent for you to use it, to pass it?

[00:26:19.280 - 00:26:40.400] Tom Byrnes
And are you using a free tool or a paid tool when you're doing that? So, again, I would also bear in mind that is PII and it is there potentially in their data. So that's also a point that needs to be brought into that part when rather than just throwing it quickly into these tools. So at least take a beat. And do you actually have permission to use this data in that way before you use the AI to do it?

[00:26:42.380 - 00:27:38.500] Karri Byrne
OK, here's a here's a question for you. So, given that EMMA is, you know, kind of has its steps and even has questions, you know, recommended questions for when you're talking to traders or recommended questions for when you're talking to, I don't know, other community leaders. I suppose you could not just having fed EMMA in, but also saying, OK, now I'm talking to traders. Read this transcript. The question was, you know, are what are the biggest infrastructure barriers? I'm making up that question. But what are the biggest infrastructure barriers for traders in this market? And then you ask it to analyse for just one of your questions, your EMMA survey, not survey, because that's not quite how EMMA does it.

[00:27:38.600 - 00:27:48.520] Karri Byrne
But you know what I mean. One of your survey questions at a time. That might be a better way of analysing the qualitative data as well. Am I am I right?

[00:27:48.620 - 00:27:51.560] Speaker uncertain
More likely to be able to see what's not there.

[00:27:52.220 - 00:27:52.460] Karri Byrne
And.

[00:27:53.320 - 00:27:57.740] Tom Byrnes
Yeah. And I think simply put, how would you as a human do that task?

[00:27:58.200 - 00:27:59.140] Speaker uncertain
Right. Right.

[00:27:59.680 - 00:28:07.800] Tom Byrnes
Because you want the AI to. People seem to assume AIs are either incompetent or literally God. And you should.

[00:28:07.960 - 00:28:15.640] Tom Byrnes
I always describe it as an overly confident new intern who's just come out of studying at the University of Nairobi or [university name unclear].

[00:28:16.000 - 00:28:32.760] Tom Byrnes
They're very smart, but not contextual. So how would you want to tell that new intern who's just got off the plane to go and do that? And if you don't know how to do that, then why would you think the AI is going to be able is not God. It's not going to solve fill in the gaps.

[00:28:33.000 - 00:28:33.120] Speaker uncertain
Right.

[00:28:33.160 - 00:28:50.240] Tom Byrnes
If you don't. So, again, that would be, again, a key step here. The key thing is that the AI can help make you go faster. But if it's something that you don't know how to do at all yourself, it's not going to unblock that issue. Because even if it does it, you won't know if it's lying to you or if it's made a mistake.

[00:28:52.160 - 00:29:16.920] Karri Byrne
Yeah, exactly. I can see coming up in the chat a little, and it was in the pre-submitted questions as well, is really what are the risks that we should be considering as we're designing programmes and how we'll use AI in it? And how can we mitigate them a little bit? Emil, did you want to add something before?

[00:29:16.920 - 00:29:52.860] Emil Heidkamp
I saw one pop-up that ties this. I saw a pop-up question that actually tied in what was talked about. Someone asked about how to make sure it reflects local knowledge and localisation. And so you're saying, oh, hey, couldn't we take the EMMA framework and question guide and just say, hey, it's for traders and country X and ask it about infrastructure. Well, again, those two connections are related because if you pose that question to a generic AI chatbot and then ask it to connect the dots, it's going to connect the dots, but it might connect the dots to data it has that assume a different context, a different country context.

[00:29:53.000 - 00:30:40.940] Emil Heidkamp
So we created a tool that actually helps bank staff in low- and middle-income countries write loans for energy efficiency and renewable energy upgrades at local businesses. So if you just ask a generic chatbot, hey, there's a produce distributor in Honduras and they have a large fleet of old diesel vans, what's a good way for them to improve their energy efficiency? If it's just connecting to generic data, it might go upgrade to electric. And then the problem, though, is there's a little detail that might not come up and if it's just skimming the world, it's unfortunate there is bias. The folks in California and in Shanghai building these models aren't necessarily focusing on the challenges of some of the places that we work.

[00:30:41.140 - 00:31:18.560] Emil Heidkamp
And so the fact is there's a little thing. There's only two charging stations that last I checked in Tegucigalpa right now. So it's like so telling them to switch their vans, you are committing them to actually install their own charging infrastructure for that van fleet. And that's the whole thing about market systems, isn't it, is that it's our job as humans to understand the localisation issues and the correct way to connect the dots, shall we say? And so the one thing we can, the value we can add is we actually worked with our partners to create country data sheets about the current state of renewable energy infrastructure and government incentives and stuff.

[00:31:18.560 - 00:31:24.220] Emil Heidkamp
So at least it wasn't going general. It was looking up that country before it spewed an answer.

[00:31:25.760 - 00:31:37.400] Tom Byrnes
And if I can come in on that question about localised [context?], where the problem is, I wouldn't say the key use case is at the point of analysing the question of that trader that you've gone and spoken to.

[00:31:37.400 - 00:32:23.280] Tom Byrnes
But actually what we found is that one of the key points is the local, one of the most things that I changed was that going for local and regional language media first. In Sudan, we were doing work and it was getting the Telegram channels and scraping the information, which was all in Arabic, publicly shared there, and translating that to get the initial brain, our initial understanding of what is actually happening. What are the actual market systems that we need to consider? What are the key enabling environments that is happening? Or what is critical infrastructure that may not be there? But bear in mind, the step one was we used that to identify, you know, for our piece of work on Iran and Strait of Hormuz, it was what are the actual big second order impacts?

[00:32:23.400 - 00:32:56.740] Tom Byrnes
In Myanmar, we suddenly saw that these things are being reported quickly. And so that would have taken time to come into English language media if I was just doing a desk review. But crucially, the next step then was we engaged with the cash working group. We engaged with the country team. We validated that. We pushed that. That came our analytical frame. And so the question is, it's the EMMA that's targeting at the wrong value chain. It's the wrong trader who was actually reached out to. This is where these tools can help you at that early step. Maybe refine, should I be talking to a wholesaler or a local trader?

[00:32:58.100 - 00:33:27.480] Tom Byrnes
This type of person? All that. So that's also where I put it into this market analysis is a very useful way to do that, to compress that time of understanding what is actually the problem I'm trying to look at here. Make sure that I've got my analytical frame as robust as I can before I've even. But you're still going to need to have teams go and actually talk to these people. They're still going to have to go there. And there's still value in having the team who's gone there and says, hey, yeah, I went there. There's one trader and he's literally got four things on the shelf.

[00:33:27.780 - 00:33:36.100] Tom Byrnes
You know, that sort of contextual knowledge that's not in the questions, it's in the other section at the bottom, is also going to be still incredibly valuable.

[00:33:36.360 - 00:34:21.140] Karri Byrne
Oh, it's still going to be critical. And the desk review, you know, the AI is only working for that. The first step of EMMA has always been do a desk review, see what information you already have so you're not collecting unnecessary information. But I feel like that first step has expanded and changed a little bit because you can do a lot with sources that already exist and asking appropriate questions around getting the AI to review it. But then step two kicks in with looking at the gaps, cross-checking what came out. And so it's almost like step one of EMMA has really gotten bigger in its role as well as it's probably time.

[00:34:21.660 - 00:35:05.540] Karri Byrne
But it means that very much by the time you're doing what I, you know, like the main meat of the EMMA, you're actually really in the gaps and you know you're in the gaps. It's that people would be like, yeah, yeah, yeah, literature review. And nobody really did it. But maybe that's kind of where AI can be a real tool for us. I want to, I'm going to cut myself off there because I really want to, it's really important that we get to this question. It comes up again and again in different ways. And that is this one that I asked already, which is around what are the risks and how we can mitigate them, particularly because we're working in fragile contexts.

[00:35:07.440 - 00:35:11.600] Karri Byrne
Tom, do you want to, or no, Tom, you just spoke. So I'm going to go to Emil first.

[00:35:14.080 - 00:35:48.420] Emil Heidkamp
Well, you know, the, the usual risks, you know, you get because we work with, with, you know, clients who work in fragile contexts. We work with, with vulnerable populations with other, you know, people with disabilities and whatnot. And so there's, there's kind of a double-edged aspect to this. So the first risk everyone says is, you know, as Thomas pointed out, it's, it's data privacy consent, you know, and sure. And, and, and that carries an extra burden. I don't think any of us read those terms of service that we click to when we launch a thing. And that goes double for, for people who might have, you know, limited literacy.

[00:35:49.420 - 00:36:26.440] Emil Heidkamp
So they're really, so, but, but the, the challenge is the first thing you need to do with consent is making sure that you are using tools that conform to whatever level of privacy your organisation and any, any partners, you know, expect you to, expect you to maintain. So, and oftentimes that's not the public cheap chatbot because those are paid for on the same basis as social media. You're, you're getting it for free because you're paying with your data, you know, so you do have to, you do have to go get an account. And now the problem is we can't expect tons of people in fragile contexts to go pay $30 a month, you know, for accounts.

[00:36:26.620 - 00:37:05.180] Emil Heidkamp
So maybe we have to find a way to share and make those resources available. So that's one, that's one challenge. And then sort of the, the next, the next challenge is it's not that it's not that, oh, well, we're going to protect them from AI. It's, it's unfortunately, it might be a case where, you know, cause when we're working, working with people with disabilities, you know, here in the United States, folks know that ChatGPT exists. They are going to go use it. So the question is, can we actually provide tools that actually provide enough additional benefit and enough additional local relevance that people will willfully use it?

[00:37:05.320 - 00:37:08.160] Emil Heidkamp
And, and, and, Hey, we're all about market solutions here. Right. So.

[00:37:08.420 - 00:37:50.240] Karri Byrne
And I want to add on that one, which is something that came up in our planning discussions earlier. And that's to say to those of you out there who are managers, remember that you have to give your staff some guidance and talk about it because we, in our conversations, we called it shadow AI. People do know these exist. They will use these tools. Either you can provide them kind of frameworks and parameters around which to use them, or they will hide the fact that they're using them. And meaning some of these issues that we're talking about may be hidden. And as a manager, we don't, you know, you don't want that.

[00:37:50.320 - 00:38:05.460] Karri Byrne
You want your team to be able to be honest about the way that they're using tools and what they need more help with. So I hope this is raising some of, some of those questions. Tom, did you want to continue on that point of risks and how we can mitigate them?

[00:38:06.440 - 00:38:07.180] Tom Byrnes
Yeah, absolutely.

[00:38:07.180 - 00:38:26.140] Tom Byrnes
So I think I'm going to carry on that exact point that you just made there and then loop back onto this because there's two key things, which is every graduate arriving in your team right now, the new [ACTED?] intern or whoever has just spent the last four years of undergraduate or graduate using AI to write every essay they've done.

[00:38:26.720 - 00:39:11.320] Tom Byrnes
New people coming into this are AI literate in a way. We talk that you need capacity training, but actually, to be honest, your new people are going to be capacity trained. Key issues here that can fix risks. I mean, there are many different risks. We could talk. AIs generally make, or AI, a general type of AI will make mistakes, which is say something that's not true, omit something, or say something that's true but is wrong. And in particular, even when it's not true, the issue could be a hallucination or it could be a bias. It could change that, you know, there was a UK study that showed that when given exactly the same medical records, when the gender changed, ChatGPT thought that women had less pain being reported than when the...

[00:39:11.320 - 00:39:51.780] Tom Byrnes
So biases that can also fit in there. Now, the key thing here, though, is not to go into details of each of those individual points. It's, as a manager, you can't actually manage a team who's using AI unless you have disclosure. Unless you're mandatory. You say, yeah, you are going to use AI. I'm actually giving it to you. This is an important thing. But you need to tell me how you are using AI. So there's a comment in the chat about the AI watermarking that Claude is doing. But it is functionally useless as a manager because it doesn't distinguish between I've used this to rewrite and reformat my bullet points versus I never actually thought I never did my work that I was meant to do.

[00:39:52.200 - 00:40:27.360] Tom Byrnes
And if you say in my... If I write AI disclosure, your team and your field officer, I used AI to reformat this. I used AI to make sure to check against this. But I validated everything is correct. And then you find out that they didn't do that. Then that's a code of conduct violation. And that's a performance improvement plan or termination. So that's the first key thing. And the shadow AI is based on the fact that people feel ashamed to say that they are using AI. That they are using these tools. But that's like saying I'm ashamed to use Excel or Microsoft Word and I'm still using a typewriter.

[00:40:27.640 - 00:40:54.940] Tom Byrnes
It doesn't, you know, so... But the key thing here is that we need to make sure... And the problem is, as a civilization, we haven't really solved that issue. As a team, you've got to... You may not want to put it as your final report that I used AI to do this. But as a manager, you definitely need to make sure that your junior staff are making it clear so that you know, okay, first of all, that if you find out... Otherwise, you can't penalise them. So that's the key thing. Making transparency...

[00:40:54.940 - 00:40:56.660] Karri Byrne
Or reward them for doing it correctly.

[00:40:57.320 - 00:41:20.680] Tom Byrnes
Or reward them. Exactly, right? But that means unless we have transparency, you cannot resolve risk. And you can also not have any learnings. And also, if one of your team members has got a brilliant prompt and a brilliant approach that has allowed them to do their work much more efficiently, and then that person goes on R&R, God help the person who has to step in and do the work without the tools that they've been relying on.

[00:41:21.100 - 00:41:50.340] Karri Byrne
Yeah, absolutely. So again, these are organisational learnings. These are... So the approaches we need to take are the approaches that we would normally take. But disclosure is what makes that operationally viable. And if you don't have that, you're in a big trouble. Yep. Yep. Absolutely. We're quickly running out of time, and I knew that would happen. But I want to go to... Blerta, have you spotted some good questions in the chat?

[00:41:50.380 - 00:42:03.540] Blerta Peta
Yeah, I suggest maybe since we're on the topic of data safety, there is a question on data security. The last question that just came in. Just Tom and Emil, if you want to quickly provide any inputs more...

[00:42:03.540 - 00:42:06.020] Karri Byrne
Can you read out the question that you're looking at, Blerta?

[00:42:06.120 - 00:42:07.280] Blerta Peta
Yes. Yes, sure.

[00:42:07.280 - 00:42:11.760] Blerta Peta
So this is from Saeed Majid [spelling unverified], the question around...

[00:42:11.760 - 00:42:41.220] Blerta Peta
He's giving the example of Spirit Airlines that is making available all the data of their staff and so on to fit into Google AI training. And he's drawing the parallel with our work and asking how will we make sure that humanitarian data across the spectrum stays safe during and after analysis, and even after the programmes or organisations don't exist anymore in this context. So maybe just a quick add-on to what you were previously discussing.

[00:42:43.680 - 00:42:46.740] Tom Byrnes
I can come in quickly on that, Emil, and then hand over to you, if that's okay.

[00:42:47.420 - 00:42:49.160] Emil Heidkamp
Yeah. Oh, that's a good card.

[00:42:50.700 - 00:43:28.000] Tom Byrnes
So basically, my key... So first of all, AI doesn't change your data management problem. You've always had this data management. It's just maybe making it mobile. The metaphor I always use is that beneficiary data, and especially beneficiary data, any personally identifiable data is like nuclear fuel. It powers your programme. Amazing. I can do these cash programmes. I can do deduplication. But it's nuclear waste. And any data of any programme should have a lifecycle of management. Don't keep it live when there's no reason for it to be kept live on a public-facing system. And this problem here particularly would be if you were using a third-party company.

[00:43:28.140 - 00:43:57.160] Tom Byrnes
What's the end point here? Make sure that there's... Once the data is not needed for that programme activity, it's taken off Kobo, SharePoint, you know, the server, and is archived. Because it needs to be kept... Data needs to be kept, but it needs to be kept in a hard drive or in cold storage. It doesn't need to be in a place which could be accessible afterwards. So that's the key thing. Humanitarian data, we should always be cold storaging it. We should be moving it to hard store. And it should only be kept in active use as long as it's actually needed.

[00:43:57.160 - 00:44:01.080] Tom Byrnes
It's very useful, but it's dangerous. And its lifecycle needs to be treated in that way.

[00:44:02.760 - 00:44:51.860] Emil Heidkamp
Yeah, I'd expand on that to say that, yes, we've always had this problem with the... shared SharePoint folder exposing things that should have been disposed of back in 2016, but wasn't. And the unfortunate thing is now with AI, if you connect it to your Microsoft 365 instance, now you have a colleague that could actually find all that. So that increases the exposure. But, of course, when we build systems, though, we have to do it for finance, for healthcare. So oftentimes we're building things where the data resides on a database controlled by the organisation, encrypted. And then it's sent either to a self-hosted AI model or to an AI model provider with which we've signed a zero retention agreement.

[00:44:52.800 - 00:45:31.220] Emil Heidkamp
And then so it's only in the AI's consciousness for a flash, but then resides on the company or the organisation's databases. And it's all encrypted. And we do cryptographic destruction when an organisation says, destroy all the records from Q1 of 2017. We just destroy the encryption key. And at least until they invent quantum decryption, that legally means that that has been destroyed. So, again, it's just good data hygiene. It's just now you've got a user, AI, that's actually going to find and use the stuff that maybe your own staff were just too slow to access without authorization.

[00:45:32.400 - 00:45:39.460] Karri Byrne
Got it. Wow. I see a question here in here that I'd really like us to go back to.

[00:45:39.580 - 00:45:51.500] Karri Byrne
We kind of touched on it earlier, but Ria [spelling unverified] is asking about conversations around the use of AI that intersects with the localisation agenda.

[00:45:52.000 - 00:46:15.260] Karri Byrne
We had talked at, I think, our prep meetings about is there a way to, you know, have AI bring those voices closer to us, which is always our goal. Any thoughts from you guys on how AI can be used to and intersect with that localisation agenda?

[00:46:17.540 - 00:46:19.280] Tom Byrnes
Emil, do you want to go first this time?

[00:46:19.280 - 00:46:50.860] Emil Heidkamp
Yeah, I actually think that this could be tremendous for that. Because right now, I mean, I've sent some reports to, you know, I've had some organisations ask us for advice on how they can localise. And we said, hey, you're going to have to get used to the fact that power is not going to reside in Geneva and London and Washington, D.C. And it's going to reside all over the place. And now the beauty of it, though, is that a lot of these global organisations have set themselves up as gatekeepers for knowledge. So, like, knowledge has to go up in a parabolic arc, up through the authorities, through peer review, get shared down.

[00:46:50.980 - 00:47:31.200] Emil Heidkamp
Someone hears at a conference and then it lands somewhere. A learning lands in another part of the world. But the thing is, we said earlier, you can use AI to analyse the results of a focus group discussion. But if you have everyone using a system, you know, around a certain topic, around maybe, like, supporting women entrepreneurs, you can have the same AI system and go, hey, wait a minute. Someone in Manila had a really good idea, escalate it up, and then it could be just briefly reviewed and vetted. And in real time, at the speed of markets, not the speed of academic research, suddenly someone in, you know, someone in Nairobi who's running a milk stall can, like, benefit from that idea or that insight.

[00:47:31.200 - 00:47:50.680] Emil Heidkamp
So, it can actually, by allowing, just accelerating the whole process and not having to wait for some intern to read 300 reports and for that to be turned into a report and submitted to a donor and presented at the conference, we can speed it up. And that can actually, there's just, there's actually power the leveling and the eradication of hierarchies.

[00:47:52.500 - 00:47:52.960] Karri Byrne
Maybe.

[00:47:54.660 - 00:48:02.960] Tom Byrnes
Well, I don't know if it's going to change the power dynamic in that power dynamic is because of money. And that's, we always have to be open on that. But it will change the dynamic.

[00:48:02.960 - 00:48:05.380] Emil Heidkamp
It can let you do things for less money. It can let you do things more efficiently.

[00:48:06.260 - 00:48:06.660] Tom Byrnes
Indeed.

[00:48:07.460 - 00:48:22.940] Tom Byrnes
But the bigger point, I think the key point here is it will play an important role because, for the first time ever, someone in Juba can access, you know, if I pay $20 a month, I can get access to the best AI, the same AI [unclear comparison referring to Oxford University or Alberta].

[00:48:24.620 - 00:48:25.360] Speaker uncertain
Right. Right.

[00:48:25.360 - 00:48:57.780] Tom Byrnes
You know, so there's a democratisation of AI. That is one thing. These tools aren't gatekeeped to only the global north. Everyone has access to them. And the AIs work, AIs have been trained on all the world's knowledge. They're not as good in Nepalese as they are in English, but in Hindi, they are relatively well. In fact, you know, so in a lot of languages, in Swahili, you can use them in those different. So, again, that's also the barrier to using these tools is reduced because now that if you're in a language that doesn't have much digital footprint, it's still not going to do a good job in it.

[00:48:57.920 - 00:49:35.160] Tom Byrnes
But if it's in Arabic or in other ones, then there can be a key role there. The other key thing is that it, you know, we never lacked for incredibly smart people in local organisations. What they lacked for was time, capacity and the ability to access these resources. I've heard I was talking to someone and they were saying, oh, yes, what I did is I went and I got AI to go and download the entirety of ReliefWeb for its entire history. That's available for my country into a local share into a local hard drive. And then I got the AI to query it and find lessons learned from the different sectors and clusters.

[00:49:35.280 - 00:50:04.200] Tom Byrnes
And I never thought to do that. Never thought to take two gigabytes of data off ReliefWeb and then go looking for. And they were like, yeah, there's so much information there. And because what they also have is the local context of what did. Hey, I remember there was a project two years ago or five years ago or 10 years ago. Now the AI can go dig that out, can actually go, you know, an idea of like, but this really good project. I worked for this organisation a couple of years ago. Now it can go and find all of those World Bank reports or whatever and bring that in.

[00:50:04.560 - 00:50:41.460] Tom Byrnes
And so it gives that opportunity for people to go find and they can ask those questions in their own languages in some cases. So that these are areas where, but it won't naturally pass down. These tools are there, but the issue will still need to be making sure that the space is given. And crucially, the issue here is, OK, but what happens if that person then if there's a stigma about AI usage and then that local voice? No, that person doesn't know anything. That person just got relief, read ReliefWeb and made something up. It's not really there. You know, the biases, prejudices that from the global north to global south aren't going to be fixed.

[00:50:41.820 - 00:50:54.140] Tom Byrnes
And AI could easily reinforce that. So, again, these are things that we have to be aware of. And it's a tool. It's not a good tool or a bad tool. It is a tool. And how we bring our back, what we bring to it is what the issues could be.

[00:50:55.380 - 00:51:41.960] Karri Byrne
Yeah. And, you know, as this conversation continues, I'm just so aware of the biggest conversation that I think we need to continue is how anybody who manages a team, whether it is a temporary team like an assessment team or a long, you know, a project team or a country programme or whatever, anyone who has a team really has to think about what is the conversation I want to be in with my team about how I use these tools? Because I hear almost I can't it's like I can't see it through the fog, but I almost can see an entirely new role for programme officers and technical advisors and that kind of thing.

[00:51:41.960 - 00:52:23.140] Karri Byrne
But because we don't quite understand what AI can do for us. We're tiptoeing right now when actually I think there are a couple of like one big solid step we need to take, like the transparency of of set some rules, you know, say on our team we use AI to do this. That's fine. Please note it when you do anything other than X, Y, Z, you know, translating it from another language or whatever. And I just I feel like we've got some big changes ahead of us. We know we have big changes ahead in the industry anyway because of funding.

[00:52:24.720 - 00:52:57.260] Karri Byrne
But this feels like yet another step change. And we all probably want to be on the front of it instead of on the back of it. We are very close to time. I want to give each of you one minute to just tell our audience, like, hey, if you only remember one thing from this webinar, I want you to remember this one thing. So who wants to go first? Emil, please go ahead.

[00:52:57.260 - 00:53:31.120] Emil Heidkamp
OK, if you only remember one thing, it's that whatever you've experienced using, you know, the generic chatbots is not the end of what AI can do. And there are many ways it can it can help our sector and our partners and on a local level, on a global level that go beyond. So so there. So, yes, absolutely. Use the generic chatbots for the things are good at like an Excel spreadsheet. But be aware there's more and it might take a little investment, but it might pay off. And so that's the number one thing I would tell people.

[00:53:32.260 - 00:53:32.700] Karri Byrne
Tom.

[00:53:33.500 - 00:54:06.020] Tom Byrnes
I think my number one thing is it's here. Any investment in time that you make in learning how to use these tools correctly is going to be valuable because it's not like it's going to vanish off out of the radar. These tools are only going to get better and they only get cheaper and more available. And in two years time, they could be open source and running on everyone's laptop, maybe on that side. So the key thing here is it's an investment in time. I think it's also not just how you use it, but how your teams use it, but how your beneficiaries, how the community members, how the government are using it.

[00:54:06.480 - 00:54:40.360] Tom Byrnes
This is this democratisation goes in all different angles on that side. And crucially, the last key thing is Emil made a good point: no one in Shanghai Or San Francisco has thought about the humanitarian community or us at all. We don't know what other rakes in the grass that we're going to walk into. We don't know what things are really safe to do and what are not safe to do. And unless this community shares what's worked and what hasn't worked, we will continue to not know and we will continue to make mistakes. And as you say, we don't have much money.

[00:54:40.360 - 00:54:45.020] Tom Byrnes
Our needs are growing, and [unclear climate reference] is going to hit us next year.

[00:54:45.360 - 00:54:47.100] Tom Byrnes
[Unclear: reference to Iran and the World Food Programme.]

[00:54:47.360 - 00:55:04.060] Tom Byrnes
So we got to we got to use these tools and we need to know how to use them safely. And the only way we can do that is by breaking the shame of saying that we're using these tools, sharing what's working, what's not working. Because they're going to be with us and that's that will pay forward.

[00:55:04.920 - 00:55:42.480] Karri Byrne
Such a good point. Such a good point. I want to I'm going to build on that in a second. But everybody who is on here, every 44, all 44 of you, I've got two one minute tasks for you. So this will only take two minutes. The first one is put in the chat right now what your one thing that you would take away from this session. It helps us know. The second thing is in a minute there's going to be a poll coming up and do make sure you you fill that out. And I'll talk about that again in a second.

[00:55:43.860 - 00:56:23.220] Karri Byrne
I just want to say thank you, guys. This I mean, I did a bunch of prep work for this, you know, to be able to to have a meaningful conversation today. But wow, I feel like I was I have I thought about new things today and I always like a day when I am like, oh, man, I got to think about that some more. So so thank you. Thank you for this really rich and interesting discussion. Thank you also again, PRO-WASH & SCALE for your support. I want to build on that last point that Tom had there about we've got to keep this conversation going.

[00:56:23.220 - 00:57:04.460] Karri Byrne
So if you feel like there were questions that weren't answered, you have more that you want to ask new ones, go to the MiC discussion board. Blerta will drop the link in the chat and post your questions there because I 100 percent agree with what's being said today. We need to keep this conversation going, particularly about what doesn't work and why and what does work. It's hardest to talk about what doesn't work. And many of you will remember the fail forward kind of conversations that we had for years about being willing to say, hey, this just didn't work. And this is why and this is what we learned.

[00:57:04.920 - 00:57:44.200] Karri Byrne
This is a really good time to to bring that forward around the A.I. space and what we did do and didn't do. Because it's just a little bit easier right now to to fix our problems before we get them out into the communities and make make bad decisions. Thank you also all of you for all of you participants who joined today. Your questions were really great. And I can see there's a lot of interest in this topic. I really do hope that you'll you'll jump onto the discourse site. And like I said, fill in the poll.

[00:57:44.940 - 00:57:45.660] Karri Byrne
Tell us [sentence breaks off].

[00:57:45.840 - 00:58:27.000] Karri Byrne
MiC wants to keep having the conversations that are important to you guys. What works and what doesn't work in our webinars. So we really, really appreciate when you when you answer those polls. Thanks, guys. I realise we're sort of two minutes ahead. Blerta, did I miss anything? It felt like a good good place to wrap up, though, because we could go on forever with this stuff. Fantastic. Thank you, everyone. Thank you, especially our panelists. You, yeah, like I said, you really gave me a lot to think about today.

[00:58:27.000 - 00:58:33.000] Karri Byrne
And I hope everyone else got some good [wording unclear] too.

[00:58:33.000 - 00:58:35.000] Karri Byrne
So have a great day, everyone.

[00:58:35.000 - 00:58:39.000] Speaker uncertain
Thank you. Thank you all very much. [Wording unclear.]

[00:58:39.000 - 00:58:46.360] Speaker uncertain
Thank you very much, Karri. It was great to, especially the questions at the end. Thank you.