The programs that help people earn a living from their own business: business training, poverty graduation, youth and producer programs, run by governments, NGOs and donors for millions of people.TheLivelihoodsCollectiveA shared, open-source foundation. We help each program build on top, run it, constantly improve it and push out the frontier. Fully transparent with no lock-in, usage passed through near cost, evidence built and lessons shared across the members.
Technology for livelihoods programs to continuously support small businesses and the self-employed as they learn, build habits, adopt new tools and grow.
Bailey KlingerNovember 2026
02
Summary
Make continuous support the norm in livelihoods programs
Buttons along the bottom of each slide open the detail: partners, how it works, sources and more.
03
The problem
Livelihoods programs reach millions, but in-person delivery limits their impact
Governments, NGOs and donors support entrepreneurs and the self-employed through three types of programs, reaching tens of millions of beneficiaries, and spending at least US$1 billion a year on business training alone. The evidence shows that support works better when it is continuous and personal, and that is the most expensive part to deliver at scale. IPA lists soft-skills training for business owners and consulting for small businesses as 2 of its 14 best bets for poverty reduction.
Business training
Poverty graduation
Youth livelihoods
Who
Existing micro and small businesses
Extremely poor households, mostly rural
Young people in self-employment and informal work
Typical delivery
A classroom course, such as the ILO's SIYB, fairs, mentorship, and self-directed courses
A productive asset, consumption support, savings and frequent (usually weekly) coaching visits
Cash grants, vocational and soft-skills training, mentoring
What the evidence says
Profits and sales up 5 to 10% on average. Larger when delivery is personal: Personal Initiative training raised profits 30% in Togo, three times standard training; one-on-one consulting in Mexico raised employment about 50% five years later
Lasting gains: positive returns in 5 of 6 countries, US$1.33 to US$4.33 in benefits per dollar; in India, consumption still 20% higher after ten years
Cash grants raised Ugandan youths' earnings 38% after four years, but the control group caught up by year nine. Persistence is the open question
Continuous support
Usually ends with the course. Consulting works but cost US$10k to 30k per firm in a Colombia trial
Coaching is one of the costliest components; the field is now testing group coaching to bring that down
Rarely continues after the program ends
04
Why now
AI can make continuous support affordable and impactful if done right
An AI coach on WhatsApp fills the gaps in-person delivery can't: personal, constant, open-ended, for about US$10 a year per active owner.
But it isn't easy to do right
A Q&A chatbot isn't enough
In a randomized trial with 640 Kenyan entrepreneurs, a GPT-4 Q&A chatbot on WhatsApp had no average effect. Owners already doing well gained about 15%; those struggling did about 8% worse. The gap came from how owners chose among the advice and acted on it. So this takes more than a chatbot: an actual coaching tool, designed to help all entrepreneurs, not only the strongest.
Building it alone is costly, and the effort is repeated
A program takes 9 to 18 months and US$80k to 200k to get a coach live, then US$80k to 200k a year to run it well, or it degrades. Most governments and NGOs don't have the engineers. Those that do rebuild the same stack, repeat each other's mistakes on engagement and safety, and measure results their own way.
Chart. The shaded area marks hybrids: AI alongside human coaches, at any mix, between the two ends. Costs as each source reports them. Classroom: provider cost per trainee across the trials McKenzie and Woodruff reviewed (US$20 to 740) and the Togo trial (US$756). Graduation: the full program per household (BRAC, US$300 to 2,000). Consulting: per firm in the Colombia trial (US$10,000 to 30,000). AI coach: MAIA's variable cost in Peru, about US$10 a year per active owner (US$0.70 to 1.00 a month, depending on configuration).
Q&A chatbots. Otis, Clarke, Delecourt, Holtz and Koning, The Uneven Impact of Generative AI on Entrepreneurial Performance. A pre-registered randomized trial of a GPT-4 business assistant via WhatsApp with 640 Kenyan small business owners. The authors could not reject a null average effect. Sub-sample analyses, split at the baseline median, show high performers gaining just over 15% and low performers doing about 8% worse.
05
The product
A business coach on WhatsApp, not a chatbot
‹MAIA MYPE Asesor IAonline
day 0
Hi MAIA!18:21 ✓✓
👋 Hi! I'm MAIA, your free assistant to earn more and spend less.
To tailor my answers, tell me your business 👇 1️⃣ Bodega/Mini 🛒 2️⃣ Restaurant 🍽️ 3️⃣ Market stall 🧺 ... 8️⃣ Other business 💼18:22
118:23 ✓✓
For your bodega: what's more urgent, selling more or spending less?
Or try one of these quick actions, ideas other bodegas have tried: 1️⃣ 🎨 Create your logo: describe your business in a voice note (name, style) and I'll send you options. 2️⃣ 📸 Photo of your shop: send a photo of your front or shelves and I'll give you 3 ideas. 3️⃣ 💡 Products at eye level sell up to 35% more.18:24
day 63
TemplateHi! For your bodega, I have an idea for recurring WhatsApp sales: a weekly product bundle your customers come back for, without you starting from scratch each time. Interested?9:15
↩ Tell me more↩ Not now
Tell me more9:16 ✓✓
The idea is to turn WhatsApp into a buying routine, not a one-off post: 1. Each week, pick 2 or 3 high-rotation products, like eggs and sugar. 2. Send a short message to customers who already bought from you. 3. Note who asks or buys.9:17
day 63
Ok, let's try it out this week9:22 ✓✓
Here's this week's post with your logo. Send it to your customers today, and I'll check in tomorrow.9:23
23 hours later
How did the bundle go? How many orders came in?9:24
3 orders the first day9:25 ✓✓
Good start. Next step: send it to the customers who haven't answered yet, and we'll see how many come back next week.9:26
day 80
What should I work on next?16:41 ✓✓
Your WhatsApp list is going well: 14 customers now. Where you can gain most is pricing. You told me you price by feel.
Want to work out the real margin on your 3 best sellers this week? Send me a photo of a supplier receipt and I'll start the sums.16:42
day 92
I want to do a delivery service, my neighbours keep asking.12:06 ✓✓
Great idea! Two things from before: in July you told me lunchtime is when the shop is busiest and you're alone. Also, your WhatsApp routine already has 14 customers on the list. So one way to get started instead of "delivery all day" is the eggs-and-sugar offer as a delivery-only option, orders by 11am, dropped after the lunch rush.12:07
Yes, my son helps in the afternoons, it would be easy to make deliveries then. Let's do it12:08 ✓✓
day 110
A customer asked for the complaints book and I don't have one. Do I really need that for a bodega?20:13 ✓✓
Yes. In Peru, every business that serves the public has to have a Libro de Reclamaciones, and a notice saying so in a visible, easy-to-reach place. INDECOPI, the consumer authority, oversees it.
Keep the notice next to the till: it also tells customers your products are backed.
Want me to write out what you need to get it set up?20:14
Yes please20:15 ✓✓
days 60 · 150
Template💸 Hi! Two months working together. Have you managed to reduce your monthly costs?10:31
↩ They didn't go down↩ Down < S/150↩ Down > S/150
Reported execution, read from the chat
Did those extra products at the counter make any difference to the day's total?10:32
Yes, they saw the small products and a few sold10:33 ✓✓
Month 5
TemplateFive months together 💪 Would you recommend MAIA to another owner?10:34
↩ Yes, definitely!↩ Maybe↩ No
Message🎤
Starts in one tap
How it worksNo name, no form, no app to install: just the WhatsApp they already use. One tap on the type of business, and something useful by the second message. The coach learns the rest as the conversation goes.
Why it mattersIt's easy, and it delivers value right away: an idea to sell more before they've done anything. Friction at the start costs owners who never come back, and the gains in sales and costs go to those who stay. Long forms, bot-like menus and intimidating open questions each lose a different kind of owner.
Reaches out first
How it worksA weekly message with the next idea for this business, a touch the day after a conversation, and a follow-up on the day a step falls due. Limits are hard-coded: it stops when an owner pauses, opts out or goes quiet.
45%of conversations are started by the coach
Why it mattersIt works like a human coach, who checks in, brings the next idea and comes back to see how it went. A question-and-answer bot waits to be asked, and that isn't what owners need: most don't know what to ask an AI. Owners who keep coming back report higher sales about twice as often as those who don't.
The learning loop
How it worksOne small step at a time, toward a goal the owner chose. The coach does the first part itself, comes back when the step is due, asks for the number, and changes the plan when something stalls.
Why it mattersTraining teaches 20 to 30 good practices and owners adopt one or two, and those are where the profit gains come from. The gap is doing, not knowing, so the loop is built to get practices into the business: people work hard when they can see today's step moving them toward a goal they chose. The design draws on Personal Initiative training, Fogg's behavior model, and experiential and situated learning.
The diagnostic
How it worksFrom the conversation, the coach scores the business on eight dimensions and reads that score when it writes a reply or a proactive message. Scores are built from what the owner says, with no questionnaire.
Customers and marketOffer and qualitySales channelsOperations and supplyFinancial recordsCosting and pricingPeople and productivityPlanning and risk
Why it mattersOwners get what they ask for, and also what they wouldn't think to ask. The diagnostic balances the topics they choose against the change most likely to raise this business's profit.
Structured memory
How it worksNine kinds of memory, each with one job: the business now, what the owner has settled, what was tried and how it went, the current goal and steps. Facts and outcomes are kept only when the owner's own words back them, and code checks the words.
Why it mattersA coach who forgets makes the owner repeat themselves and pushes what they already ruled out. One who remembers gives advice that fits this business's real numbers and builds on what already worked, so it gets used. General AI tools compress a chat for a different purpose; this one is compressed for business coaching.
Local and co-created
How it worksEach coach draws on a knowledge base of local manuals and rules, built with the institutions that know them (the good-practices manual for Peru's bodegas here, INIA for crops, ILO curricula) and checked against the source, plus web search for what changes.
Why it mattersA generic chatbot can miss a rule like this one, and for a small business that can mean a fine or a closure. Local advice is advice an owner can act on. And taking part in the build creates ownership: the coaches and frontline staff who present this to owners help make it and test it, so they see it as theirs and as an ally in their work.
Data that feeds testing
How it worksMeasures are read from the chat itself: every step ends in a record (done with the owner's words and number, stalled, dropped or expired), and reported execution (REX) picks up what owners say they did, with no survey question. Short in-chat surveys at months 2 to 10 add sales, costs and satisfaction.
Why it mattersAll of it is collected automatically as people use the tool, with no external survey to field, and the chat-based measures ask nothing of the owner. The same data runs pre-registered A/B tests, partner reports and impact evaluation, so changes are judged on whether owners act on them and whether their sales and costs improve.
Illustrative conversation, composed from real threads in Peru and translated from Spanish.
06
We partner with governments, NGOs and researchers
Programs bring what makes them unique, we bring everything else
The program brings
Its method and content
Its participants and their trust
Its name and voice
+
We bring
Infrastructure layer: WhatsApp and AI, plus the database, surveys and results behind the evaluation loop
Behavioral layer: best practices for how a coach should remember, follow up and re-engage
Build expertise: a consulting engagement to turn a program's method into how the AI actually behaves, from having done it ourselves and with others
Monitoring, evaluation and continuous improvement, with what we learn across every program
=
The program gets
Its own coach, with its prompts, content and data its own
Continuous support for every participant
Evidence on common instruments
No lock-in: open source, take it in-house any time
Why they work with us
6 to 10 weeksto go live, against 9 to 18 months alone
A fractionof the cost of building and running it alone
Every programgets the fixes, features and test results from all the others
07
Traction
Three years at the forefront of AI support for small businesses in emerging markets
Since launching in Peru in November 2023, MAIA has supported small businesses in emerging markets through the programs that serve them, coaching first and adding new tools every year.
18,100small business owners reached in seven countries
1.47Mmessages between owners and their coach
3,000owners active each month; 800 to 1,100 each week
Early results · this is just the beginning
41%report higher sales by month 2
43%report lower costs by month 3
+71net promoter score at month 5; 83% find it useful at month 7
~20×return: about US$20 a month in owner profit for US$1 of cost
2% → 17%formally registered, Colombian coffee farmers, independent study
55%directly implemented the coach's recommendations, same study
In-chat surveys, self-reported, n=357 to 528; net promoter on a 3-point scale; return about 9× if self-reports are halved. Before-and-after study, n=102. First RCT in preparation.
Proven sustainability
Pilot → budget
Funded pilots turn into ongoing contracts. The ITCILO signed 12-month hosting contracts for Kenya and Jordan after a short build; AMPYME* scaled its pilot more than 10× and is taking it onto its own budget.
Contracts; *vendor registration in progress
Proven demand
Zero outbound
and still a steady flow of new programs, each with its own approach: ILO programs, MOVE-ETI's Personal Initiative training, a leading NGO's graduation program, IFAD with Profonanpe. All came to us, and demand already exceeds what the current team can serve.
Coaching is the entry point: once an owner talks to the coach regularly, it can carry much more. We have reached what a very small team can serve and are only scratching the surface of what this tool can do. The next stage builds the team to go further.
What the next stage builds
Shipped · 2023 to 2026
Scale the core, build the team
Push the frontier
Owner
Coaching on WhatsApp · Nov 2023
Photos and voice notes · 2024
Image generation · 2025
Owners log their own sales · 2026
Goals, steps and memory · Oct 2026
Stickier onboarding
Improved retention through optimized follow-up and re-activation
Further improvements to memory, state tracking and diagnostic
Easier incorporation of non-digital context
Continuously improving coaching
Owners can take their data with them
Links to finance; a sales and cost ledger
Sales help: doing more of the work
Voice out; surfaces beyond WhatsApp
Integrations with digital services
Help moving on to other tools when ready
What the team finds next
Program
One platform running several programs
A partner's own method built into the coach
Surveys and partner reports
Hardened to run many partners
A new program live in 6 to 10 weeks
Partner dashboards and reporting
Alternative configurations at lower cost per user
A hybrid human-and-AI design for graduation
A menu of validated coaching tools
An open standard programs own, with no lock-in
What one program learns reaches the others
What the team finds next
Government
Opportunity radar: connecting owners with public programs · 2026
Pulso, in beta with Peru's government · 2026
Pulso rolled out
Government dashboards
Small-business representation in policymaking, such as Peru's mesas ejecutivas
Links to public programs
A standard part of livelihoods projects
An open standard they own
Direct channel for programs and disasters
What the team finds next
Evidence
Eval suite of 1,500 cases; daily review
7 pre-registered A/B tests read
Impact surveys at months 2 to 10
Independent before-and-after study
RCT proposals under review
Common instruments published
First test across 3+ partners
Annual evidence report, nulls too
Potential Peru RCT
How big is the effect, and on what?
For whom? Weaker businesses too? Does it last?
Substitute or complement to human coaches?
Cost per result against traditional delivery
Open evaluations AI companies can build on
Questions partners bring
Today: demand beyond our capacity
09
Team
A small high-impact team focused on innovation
These are the functional areas the team covers. They can be combined or split depending on the people we select. Together they harden and scale the core service and push the frontier.
Bailey KlingerDirector & Chief ScientistHarvard PhD. Founded EFL and scaled it over ten years to a team of 63 across four continents, mobilizing about US$1 billion in credit for small businesses without credit histories. Built MAIA from the ground up and runs it today.
In place today · Hacer Perú, Lima
Piero Ghezzi Founder and director of Hacer Perú; former Minister of Production and head of emerging-markets macro research at Barclays; PhD, Berkeley. Part of MAIA since inception, including Pulso.
Mónica Moreno Government relations and communications; led communications for Peru's Presidency and Council of Ministers.
Finance and admin Accounting, treasury and contracts for MAIA's ILO, IFAD and government work.
Engineering
Designs and builds new capabilities with product; architecture, security and reliability for many partners.
Engineer
Runs and extends the platform.
Partnerships
Signs and launches partners; makes each build a repeatable product.
Implementation specialists
Launch and run each partner's coach, as partners sign.
Product
What gets built and why: the coaching loop, onboarding, the test pipeline.
Research
Evals, tests across partners, surveys, RCTs.
Operations
Runs the organization's operations, including contracting, pricing, financial oversight & reporting.
Finance & admin operations
Contractors.
The future we see
The individual
A trusted place to start
Starting out with AI in a business of your own is confusing. You don't know what to use or who to trust.
Your coach is the one place you start from. It works in your interest, connects you to what you need and helps you move on when you're ready.
Hit the ground running
You wake up knowing what to work on today and how it moves you toward your goal.
When people can see a clear step now that really moves them toward their goal, they are highly motivated.
Your numbers without friction
Your coach knows your business intimately, with minimum friction and manual input.
You always know how the business is doing and where it is heading.
Opportunities find you
You never have to scan for programs, services or opportunities.
Your coach watches all of them and brings you the ones that fit your business.
Done with you
Your coach works alongside you. It doesn't just tell you what to do. It takes some things off your plate.
That lets you focus on what makes your business uniquely you: your customers, your products and your goals.
Free by design
There is no upgrade to push and nothing to sell you.
Your data is yours and goes with you when you move on.
The world
AI as a tool for catch-up
Today's AI tools are made for the digital world and help the highest performers. In a trial with 640 Kenyan entrepreneurs, owners already doing well gained about 15% from a Q&A chatbot and those struggling did about 8% worse (Otis et al.).
We invert this, to make AI a tool for catch-up rather than divergence. We fill a hole the private sector is missing and show AI companies how to help the businesses who need it, with open evaluations and evidence of what works for these owners.
More shots on goal
Livelihoods programs no longer work in silos, duplicating effort without knowing what works. A wide variety of approaches each get a best-in-class attempt.
What works best is shared globally and quickly scaled.
The livelihoods standard
A range of validated coaching tools is a standard part of every livelihoods project, at governments and at multilaterals.
Project teams know exactly what to do. They choose from a menu with comparable results, all on one reliable platform.
The macroeconomic payoff
The biggest productivity gap between emerging markets and rich countries is in small businesses. In the UK they reach 84% of a large firm's productivity. In Kenya they reach 6% (McKinsey).
Counting the self-employed, they hold 80 to 90% of jobs in low- and middle-income countries (ILO). So even modest gains have real macroeconomic payoffs.
A standard they own
NGOs, the World Bank or a government don't have to choose among what big vendors pitch them.
They own the standard. It is open source and transparent, with no lock-in, and a menu of third-party providers for components like LLMs, with evaluations specific to livelihoods coaching.
A seat at the table
Governments can rapidly reach targeted firms for specific programs and at-risk firms for disaster guidance. And the channel runs both ways: governments can learn directly from small businesses and the self-employed what is holding them back. They can then focus improvements on those constraints and see how they change over time. That gives these businesses a seat at the table normally reserved for larger firms.
In five years: 50+ programs · 20+ countries · 1M people reached · 2 of 14 of IPA's best bets at scale
These programs already exist, with budgets and participants. Together they come to hundreds of programs and thousands of partner organizations, reaching tens of millions of entrepreneurs and households around the world, with at least US$1 billion a year spent on business training alone.
Segment
Size
Source
Business training
At least US$1 billion a year spent training at least 4 to 5 million potential and existing entrepreneurs in developing countries; the numbers are fragmentary, so this is a floor. ILO SIYB alone has 10,000+ partner organizations and about 13,000 active trainers. Personal Initiative runs in World Bank projects in about 8 countries
McKenzie 2021, in McKenzie, Woodruff et al., VoxDevLit 2025; ILO; Leuphana
Economic inclusion and graduation
405 programs in 88 countries, reaching about 15 million households and benefiting over 70 million people directly or indirectly. Governments cover nearly three-fourths of participants. 38 government graduation programs are running or in design
PEI 2024; World Bank; BRAC 2026
Youth entrepreneurship
Youth Business International: 75 members, and 277,879 young people building entrepreneurship skills in 2025. This is one network; we found no global count of youth programs
YBI 2025
National and regional small-business agencies
One in nearly every country; Brazil's SEBRAE alone has 27 state units and is building its own AI advisers
OECD 2024; SEBRAE
Lenders to small businesses
Thousands: 5,332 microfinance institutions in the West African monetary union alone. Lenders also want the data back
BCEAO 2024
Value chains and cooperatives
Coffee, cocoa, dairy buyers and cooperatives; one Brazilian coffee cooperative has 21,000 members
Partner data
Sources
Business training
McKenzie (2021), "Small business training to improve management practices in developing countries: re-assessing the evidence for 'training doesn't work'", Oxford Review of Economic Policy 37(2). Pooled estimates: profits and sales up 5 to 10% on average. Early trials were mostly too small to detect effects of that size.
Campos, Frese, Goldstein et al. (2017), Science. Togo, 1,500 firms: Personal Initiative training raised profits about 30% over two years, against a statistically insignificant 11% for a standard course (IFC Business Edge). A seven-year follow-up found the gains still there.
Bruhn, Karlan and Schoar (2018), Journal of Political Economy 126(2). 432 firms in Puebla, Mexico: consulting raised productivity and return on assets in the first year (no effect yet on sales or profits), and employment and wage bill by about 50% five years later, from social security records.
Iacovone, Maloney and McKenzie (World Bank WPS 8854). 159 auto-parts firms in Colombia: individual consulting (about US$30k per firm) and group consulting (about US$10k) improved management practices by a similar 8 to 10 percentage points.
IPA Best Bets: soft-skills training to boost business profits and sales (#7) and consulting services for small and medium-sized businesses (#8).
Poverty graduation
Banerjee, Duflo et al. (2015), Science. 21,000 people in Ethiopia, Ghana, Honduras, India, Pakistan and Peru: gains in assets, savings, work and food security a year after the program ended; benefits of US$1.33 (Ghana) to US$4.33 (India) per dollar spent, positive in five of six countries.
Banerjee, Duflo and Sharma (2021), long-term follow-up in West Bengal: consumption about 20% higher than controls after ten years, with effects growing for the first seven.
Costs: BRAC puts implementation at US$300 to 2,000 per household, with assets and staffing about 30% each. Trials in Uganda, the Philippines and Bangladesh found group coaching as effective as individual coaching (summarized on VoxDev).
Youth livelihoods
Blattman, Fiala and Martinez (2020), AER: Insights 2(3). Uganda's Youth Opportunities Program: grants of about US$400 raised work 17% and earnings 38% after four years; by nine years the control group had caught up.
How they fit together
The three literatures cover different people (existing businesses, the extreme poor, young people starting out), but they keep finding the same thing: results improve when support is personal and continuous, and that is costly.
Partners and outcomes
Governments, the ILO, NGOs and research groups keep coming to us, each with a different method and population, and the pattern repeats: a funded pilot or a short build, then a contract to keep it running. On October 26 a leading graduation NGO started the first stage of its own coach, for its programs in Mexico and Colombia: the collective's model, running today.
Partner
Program
Since
Basis, and what it showed
PRODUCE, Peru
National MSME program
Nov 2023
Reference instance, operated directly, and the lab where every behavioral change is A/B tested first. 7,500+ MSMEs; registered in Peru's national AI catalogue; agronomic coaching developed with INIA
AMPYME*, Panama
National small-business authority
2025
ILO-funded pilot scaled more than 10× to 6,000+ enterprises; moving to a service contract. AMPYME had built and abandoned its own bot first
ITCILO (SABY), Kenya and Jordan
Graduates of the ILO's SIYB training
2025
Two 12-month hosting contracts after a scoping exercise and short MVP build
ILO, Colombia (FORLAC, CaribeCafé)
Coffee producers in Huila
Jul 2024
Funded pilot with an independent before-and-after study: 8.5× formalization, 55% direct implementation, record-keeping 78% to 89%; 41% uptake, highest seen
ILO, Brazil
Coffee growers in Minas Gerais
Aug 2026
Contract to August 2027; national deployment now in government IT due diligence
MOVE-ETI, Germany
Personal Initiative training
2026
Fixed-fee build of a Personal Initiative coach, ready for their next training projects. A partner's method can be fused into the behavioral layer
A lender, Uganda
Loans to small businesses
Feb 2026
12-month contract for its clients, with a data feed back to the lender
A coffee exporter, Colombia
Coffee exporter's producer program
2026
One-off build of a learning platform and coach
A graduation NGO
Graduation, women in extreme poverty
Oct 2026
First stage, staged to implementation and hosting
In conversation: a university research program in Uganda on a coach for its post-training follow-up, and IFAD with Profonanpe on a Peru Amazon pilot. IFAD funded the Peru lab's next year through its M4D award and is connecting us to pilots in its programs, as the ILO did. An RCT of the Peru deployment is under review by study funders. *AMPYME: vendor registration in progress.
Outcomes: the detail behind the slide
The survey results come from in-chat surveys (Sept 2026 dashboard; owners with real businesses and 3+ messages). The gains hold later: at months 9 and 10, 41% report higher sales (n=105) and 52% lower costs (n=68).
Would recommend at month 5: 83% yes, 4% maybe, 13% no. Useful at month 7: 64% very, 19% somewhat (n=377).
Return: from the size bands owners report, about US$14 a month in cost savings plus about US$6 in margin on extra sales, against about US$1 a month of variable cost.
In the Colombian coffee study (n=102), a further 25% partly implemented recommendations, on top of the 55% who did so directly.
All self-reports carry survivor and responder bias, and the RCT in preparation is what will settle impact. These results come from early versions of the coach, before most of the features built since.
Engagement, as it stands
Onboarding loses about half of new users; sustained use at months 3 to 6 is around 20%, varying by intake channel.
Engagement is a dose: users who never return after onboarding week report higher sales 24% of the time; users active every other week report it about 50% of the time (self-reported, correlational). Impact rises with engagement up to every-other-week use, then plateaus.
Return rate rises steeply with how much someone gets from the first session, so onboarding is the first lever. The opportunity radar is lifting the month-3+ tail.
45% of two-way sessions are opened by the coach (last 90 days, verified in BigQuery).
Farm businesses
A farm is a microenterprise, and the stack already coaches them: INIA agronomy in Peru, coffee growers in Huila and Minas Gerais, the IFAD-supported work. Sector belongs to the partner, so the collective can partner with agricultural advisory organizations. Farm owners use the coach mainly as a business coach: among 973 agri-food users in Peru (farmers, traders and retailers), conversations about selling and market access (about 70%) and pricing (27%) outnumber those about agronomy (23%).
Source: MAIA conversation data, July 2026 (topics overlap, so shares add to more than 100%).
Why a “collective” and not a company
Programs rightly fear vendor lock-in, and want the security of an option to bring a tool in-house in the future. Open source, with usage passed through near cost, is what makes the yes easy for governments and organizations. No equity value by design; what is created is scale, impact, and public goods. We directly reduce costs in existing budget lines of partner organizations, which makes the service easier to sustain than innovations that need new budget commitments.
How the team works
Senior builders, not managers. A small peer team where each person owns an outcome and does the work, with AI tools doing much of what used to take a team.
Product and engineering design together. Product and engineering shape new capabilities side by side. Product has the final say on what gets built and why, engineering on how and when it ships.
Core and frontier in parallel. Hardening the platform and building new capabilities run side by side. New work runs as bets, each with a written hypothesis, a lead indicator and a rule for stopping it, moving from exploration to the Peru lab to partners.
Short loops. Few layers between the people building and the beneficiaries using it. Decisions that can be reversed are made fast and tested.
Headcount last. A role is added when the work can't be done with today's tools and the people in place. Implementation specialists follow signed partners; engineers follow the roadmap. Fewer, stronger people.
Operating design. A virtual team hired for overlap with Latin American and African time zones, async by default. Fractional-to-full-time is an acceptable trial path for senior roles.
Why strong people join. Impact on the livelihoods of millions of small business owners and self-employed people, ownership, a real AI product rather than a feature bolted on, a short loop from idea to users, and a product already used by thousands of owners across seven countries.
Further reading
Dell'Acqua, Ayoubi, Lifshitz, Sadun, Mollick et al., The Cybernetic Teammate (NBER Working Paper 33641; Organization Science). A pre-registered field experiment with 776 professionals at Procter & Gamble: individuals working with AI matched the performance of two-person teams without it, and produced more balanced technical and commercial work. Ethan Mollick's summary.
Shopify CEO Tobi Lütke's memo, "Reflexive AI usage is now a baseline expectation" (April 2025): before asking for headcount, teams show why the work can't be done with AI. The same rule as "headcount last" above. Coverage.
Illustrative profiles
These are general functional areas. They can be combined or split up depending on the candidates we select, and the profiles below are illustrations of the experience that fits each area.
Area
Covers
Illustrative profile
Partnerships
The consultative sale, build engagements, partner success; works with the implementation specialists
Has sold to governments and development programs and seen implementations through. Spanish and English
Engineering
Designs and builds new capabilities with product; architecture, hardening for many partners, security and data protection, reliability and on-call
A senior engineer who has shaped products, not only built them, and still ships, now with coding agents; has run AI systems in production and answered government IT and security reviews
Product
What gets built and why: the coaching loop, onboarding and the test pipeline, and finding, testing and shipping the next things owners need
Zero-to-one product for mass-market consumers who aren't tech-savvy; builds with AI themselves; instruments what they build and runs experiments
Operations
Runs the organization's operations and owns the P&L: finance and budgets, pricing with partnerships, contracts, hiring and people, compliance and data protection, funder reporting
A senior leader who makes executive-level calls as a peer of the rest of the senior team; has run operations, finance and people for an international organization or program
Research
Evals and regression suites, pre-registered tests across partners, survey instruments, dashboards, liaison with RCT partners
Applied researcher from the IPA / J-PAL world, at home in code and data
Engineer
Runs and extends the platform; product work end to end
Strong generalist; owns a problem from start to production
Implementation specialists
Configure, launch and run each partner's coach and knowledge base; train partner staff; recurring reports; first-line support
Customer success or implementation with a government or NGO counterpart; native Spanish, professional English; based in Latin America
Bailey Klinger, founder
Director & Chief Scientist who founded and will lead the collective. An economist who turned research on small-business finance into a successful global company, and who since 2023 has built MAIA himself.
US$1Bcredit mobilized by EFL for small businesses without credit histories
63people on the EFL team, built over ten years across four continents
18,100owners coached by MAIA, in seven countries
PhDPublic Policy, Harvard
MAIA, 2023 to now
Founded MAIA and built it hands-on with AI coding tools: the coaching engine, the evaluation suite and A/B program, and the partner programs with the ILO, IFAD, AMPYME and others. Three years in production, 1.47 million messages, zero outbound sales.
Entrepreneurial Finance Lab (EFL)
Created the EFL technology at Harvard's Center for International Development and led the company as founder, CEO and executive chairman. EFL's psychometric credit scoring let lenders serve small business owners with no credit history or collateral, at scale and profitably. Over ten years it grew to a team of 63 across four continents and mobilized about US$1 billion in credit. Successful exit in 2017.
Research and policy
Senior Research Fellow at Harvard's Center for International Development. Books and papers on entrepreneurship, growth diagnostics and structural transformation, in journals including Science, and published impact evaluations of entrepreneurship training. Senior advisor to governments and multilateral institutions, including Peru's Ministry of Production from 2014 to 2016, where he first worked with Piero Ghezzi and Mónica Moreno.
Education
PhD in Public Policy, Harvard University. MPA in International Development, Harvard Kennedy School. BBA, Simon Fraser University.
His role in the collective
Leads the team and sets the vision. Owns the product thesis and the research agenda, the anchor relationships (ILO, World Bank, flagship governments) and fundraising, and keeps protected time in the Peru lab reading conversations and running experiments.
Partner journey
The journey
Three stages. A program sees a working coach before committing to a launch, and can leave at any point with everything built for it.
Design and first buildAbout 6 weeks
We build the coach from the program's materials and method. Its staff test it on WhatsApp and we revise it daily. Then it decides whether to launch.
Launch and run12 months, renewable
Its own WhatsApp number and name. We run it: daily review, testing before every change, surveys, dashboards and reports.
ScaleAt its own pace
More programs, countries and languages, and new capabilities as they ship.
Under the hood: the engine
The engine is the system already running MAIA and SABY, offered as a shared base: multi-tenant infrastructure, the behavioral layer (proactive cycles, follow-up, re-engagement), coaching-specific state tracking and memory compaction, and the safety testing and evaluation machinery. Partners bring their program, content, participants and trust. We directly help them build, deploy, monitor, test and scale.
What stays the program's
Partners control content, knowledge base and prompt customizations, through the build engagement, and own them: we don't reuse one partner's prompts, content or participant data in another program. What flows back to the shared engine is general: fixes, features, test results and lessons about what works. Our own R&D and testing ship to every instance. The behavioral layer, state and memory, safety and testing machinery are shared. Each coach carries the partner's name and voice. Pooled results use common instruments, with each partner's agreement.
Valuable data for policymakers
A coach in thousands of small businesses learns something policymakers can rarely see: what is actually getting in the way. Policy conversations with the private sector reach the larger firms, because there aren't enough seats at the table for micro firms.
Pulso
Pulso is a tool built on de-identified MAIA data that lets officials ask what small businesses are living through, in their own words. It shows, for example, which part of tax formalization owners get stuck on, and which part of formalizing employees actually stops them. Because the data is continuous, it can show how those change after a policy changes. Peru's government is starting to use it.
Why it matters beyond one tool
Most workers and much of GDP in these economies sit in a very large base of small, mostly service businesses, so a small productivity gain across that base is macroeconomically relevant. The coach is one way to reach that base, and the same relationship can show whether it is working.
Testing and evaluation
Three layers, each answering a different question. Does a change break anything? (the eval suite). Is production behaving? (monitoring). Does a change make owners engage and act? (randomized A/B tests).
1,500test cases, 695 of them regression pins from real incidents, never trimmed
10production surfaces, each with its own gate
205daily reviews of live conversations since March 2026
7randomized experiments read since May 2026, about 15,000 owner assignments
Before a change ships
Every prompt, model or pipeline change is replayed against cases built mostly from real conversations, old version against new, several times each.
Every push
27 deterministic test files, no model calls
Surface gate
Every pin and typical case, old vs new, several reps
Defect screens
Coaching defects with quotes; domain facts checked against manuals
Human read
Every newly flagged quote, for the judgement-heavy defects
Deploy and watch
Clean snapshot, traffic check, a named watch in the daily review
Instrument
Used for
Deterministic checks
Routing, repetition, language leaks, rendered messages, decision rules: everything they can cover, with no judge needed
Per-case assertions
Regression pins: the pass criterion is the incident's lesson
Defect screens
Ten named coaching failures, each binary with an exact quote: answers next to the question, re-summarizes, offers instead of delivering, no clear next step, a price accepted without a market check, interrogation, and others
Fact check
Doses, fees, requirements, deadlines, prices and contacts, checked against the sector's manuals, cited links and official sources (on 48 pinned replies: 4 flagged, all real, no false alarms)
Blind pairwise preference
Supporting evidence only, never a gate
In production
Layer
Runs
Catches
Deploy check
After every deploy
A new version that is built but not serving
Platform alerts
Continuous
Error bursts, an AI provider failing over, a batch job killed by a timeout
Lost-message reconciler
Every 30 minutes, on separate infrastructure
Messages delivered to the WhatsApp provider but never received
Hourly health check
Hourly, read-only
Error replies, unanswered messages, untranscribed voice notes, safety and unsubscribe keywords, incomplete jobs, volume against a week earlier
Daily review
Daily
Coaching quality, factual errors (every phone number, address, regulation and link checked against sources), safety in five languages, every new owner's first conversation, a random sample, partner-specific checks
Dashboards
For partner reporting
Acquisition, retention by cohort, survey impact, execution, each against its history
Automated checks only report: none sends a message to an owner, changes a user's status or deploys. Every finding cites the user, the time and the quoted message. A production failure worth fixing becomes a regression pin, so every later change is tested against it.
Randomized A/B tests
Experiments run on Evidential, the open-source experimentation engine from IDinsight and The Agency Fund. Owners are randomized one at a time and stay in their arm for the life of the experiment. Outcomes are computed in our own warehouse, so raw conversations never leave it, and any error falls back to the control experience.
Pre-registered. Each experiment's hypothesis, primary measure with its query, size, decision rule and read dates are committed before launch, and read the way they were written.
Three tiers, each with its own claim. Engagement (replies, returns, active weeks), execution (steps done, by an arm-blind scorer), and impact (sales and cost surveys, and the RCT).
One yardstick for every test. Among owners who answered the day-60 sales survey, those active in 4 or more weeks of their first 60 days report a gain 50% of the time, against 22% for one active week, and the share goes flat beyond that (correlational, and it holds within each country). So every experiment reports the share of owners "at dose": active in at least 2 of the 4 weeks after enrolment.
Experiment
Result
What we did
Outreach to quiet owners
Interactive offers lifted positive engagement from 8.3% to 10.5 to 11.1%, with about a third of the opt-outs
Shipped
Deliver first in the core reply
Opt-outs fell from 8.35% to 6.26%; returns flat
Adopted into the baseline
Check-in wording; profile-quoting outreach; nudge in a first-session silence
Clean nulls
Retired
Targeted practice and peer norm
Null, but reached only 15 to 21% of its arm
Idea untested; led to an early delivery check
MAIA 3.0b (in the field since 28 October 2026)
A coach that picks the change with the biggest profit stake and follows it up weekly, against one that follows the owner's lead; 350 owners per arm, with a weekly sales panel in both arms
If it ships, it is the frozen treatment for the RCT
What changed in MAIA, week by week. Sources: the MAIA Eval Guide (20 October 2026) and the MAIA A/B Testing Guide (28 October 2026), both internal and available on request.