An open-source framework and implementation assistance for the programs that support micro and small businesses and the self-employed: a coach in your program's name that works with each participant every week.
Training teaches 20 to 30 good practices, and participants typically adopt one or two. Support works better when it is continuous and personal. But that is the costliest part of a program to deliver in person.
Illustrative; programs differ, from a course with no follow-up to weekly graduation visits.
5 to 10%average gain in profits and sales from classroom business training
About 3×that effect when the training is personal and action-focused (Personal Initiative, Togo)
US$10–30kper firm for one-on-one consulting, which works but rarely scales
03
The coaching method
A coach that reaches out first
Every coach we build runs on the method we developed over three years across multiple countries. It knows the participant's business and goal, agrees one step at a time, comes back to hear what happened, and adjusts. On WhatsApp, in text, voice notes and photos, free to participants.
1 · CUE
Proposes one small step
Toward the owner's goal, and does the first part itself: the text, the list, the calculation.
2 · COMMIT
The owner takes it on
With a tap, a yes, or their own plan in their own words. If they don't say when, the coach suggests a day.
3 · REFLECT
Follows up on the day
Asks how it went, and for the number when there is one: orders, sales, costs.
4 · ADAPT
Adjusts the plan
A stalled step is a question about the step: make it smaller, change the tactic, or revisit the goal.
A Q&A chatbot isn't enough: in a trial with 640 Kenyan entrepreneurs, a GPT-4 assistant on WhatsApp had no average effect, and struggling owners did worse. MAIA starts 45% of its conversations and follows each step through.
‹MAIAonline
For your bodega: what matters more right now, selling more or spending less?9:12
Selling more. Sales are slow on weekdays.9:14 ✓✓
Let's make WhatsApp a weekly buying routine: a bundle your regular customers come back for. Here's this week's post with your logo. Send it today, and I'll check in tomorrow.9:15
23 hours later
How did the bundle go? How many orders came in?8:40
3 orders the first day8:52 ✓✓
Good start. Next step: send it to the customers who haven't answered yet.8:53
Illustrative conversation, composed from real threads in Peru and translated from Spanish.
04
How it works
Your program's own coach
You bring the program. We bring the AI coaching method, the technology and a team that builds, runs and improves your coach with you.
You bring
Your program, content and participants
Your method and materials, the participants who trust you, and your own name and voice.
Curriculum and methodMaterials and FAQsParticipantsYour name and toneWho takes sensitive cases
MAIA brings
Everything a coach needs to work, tested on real users
The AI coaching method, WhatsApp and AI infrastructure, safety, testing and daily monitoring, and a team that keeps improving your coach. Fixes, features and test results from every program reach yours.
WhatsApp and AI modelsProactive follow-upMemorySafety and escalationEvals and daily reviewSurveys and dashboardsUpdates from every program
Yours to keepYour prompt, knowledge base, configuration and participant data belong to you.
Open sourceThe base code and prompts are released CC0, free for anyone to use.
No lock-inTake it in-house or to another builder at any point, and we'll help you move.
05
In practice
What entrepreneurs use it for
Besides the weekly follow-up, entrepreneurs ask questions, get work done and find programs that fit. Sector knowledge is developed with local institutions and checked against source manuals.
Ask anything about the business
Pricing, costs, customers, permits, crops and animals. Text, voice notes or a photo of a price list, a receipt or a sick plant. Answers draw on your materials, sector manuals and current web sources.
Does the work with them
Flyers, logos, product photos, price lists, customer messages, simple forms and calculations. The owner leaves with something ready to use.
Connects to opportunities
An opportunity radar checks public programs, funding calls and fairs against each owner's sector, region, tax registration, sales, years in business and gender, asks about what it doesn't know yet, and shows only the ones that fit, with deadlines. Live in Peru, and set up for each new country during the build.
Tracks their progress
Their goal and every step they take on, closed as done, stalled or dropped, with the result in their own words. Owners can also log sales and costs in the chat, and the coach builds on what worked.
Live today in seven countries, in Spanish, English, Portuguese, French and Arabic.
06
Results so far
Three years in production
MAIA launched in Peru in November 2023 and has run every day since, through governments, the ILO, NGOs and private partners.
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
What owners report · early versions of the coach
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
In-chat surveys, self-reported, n=357 to 528; net promoter on a 3-point scale. September 2026.
Independent evaluation
2% → 17%
of Colombian coffee farmers formally registered after using the coach; 55% directly implemented its recommendations, and record-keeping rose from 78% to 89%.
Before-and-after study, ILO Colombia, n=102. A randomized trial in Peru is under review by study funders.
Who we work with
10×
AMPYME, Panama's small-business authority, scaled its ILO-funded pilot to 6,000+ enterprises. Also national programs with PRODUCE in Peru, and ILO programs in Colombia, Brazil, Kenya and Jordan.
Funded pilots that became ongoing contracts
07
How we work together
Three stages
You see a working coach in weeks, before committing to a launch, and can leave at any stage with everything built for you. Scale and timing stay your decisions.
Stage 1 · about 6 weeks
Design and first build
Fixed fee
Your coach's prompt and knowledge base: your materials and method, built into MAIA's coaching method
A working test coach on WhatsApp for your team, in your language
A plan for linking to your systems, if needed: how participants are matched and what data flows each way
A readout of what testing showed
Stage 2 · launch, then 12 months
Launch and operate
Launch fee, then monthly
Your own WhatsApp number and name
Links to your systems, a testing round with staff or coaches, then launch
We run it: daily review, testing before every change, outcome surveys, dashboards, reports
Stage 3 · at your pace
Scale
Scoped with you
More programs, countries and languages
Your staff update content and settings
New capabilities as they ship
Off-ramp · after Stage 1
Take the prompt, knowledge base, integration plan and readout to another builder, or build in-house.
Off-ramp · any time we run it
Take the open-source code, your prompt, knowledge base, data and number, and run it yourselves. We'll help you transition.
08
Stage 1, step by step
The first six weeks
When
Step
Your team
Our team
Week 0
Kick-off
Shares materials and names a program focal point
Sets up the workspace and reviews the materials
Weeks 1–2
Discovery and first drafts
Working sessions on the method, scope, tone and topics to hand to a person
Builds the knowledge base and drafts the prompt; first answers shown side by side for review
Weeks 2–3
Test coach live
Names testers, as many as you like
Opens the test coach on WhatsApp; tests it on difficult and unusual questions
Weeks 3–5
Test and improve
Testers use it and log feedback in one shared log
Reviews every conversation daily and revises
Week 6
Readout and handover
Decides whether to go to Stage 2
Hands over the prompt, knowledge base, integration plan and readout
Nothing to set up on your side: no servers, no WhatsApp account, no software licences. The test coach runs on MAIA's line, and you decide how many people test and how many rounds.
09
Understanding the cost
Third-party delivery cost
Third-party costs for WhatsApp and the AI providers: about US$10 a year per active user, against hundreds to thousands of dollars per person for traditional approaches.
Per active user, per month · Peru, list prices
US$0.82 about US$10 a year
WhatsApp messages (Meta)US$0.48
Mostly the messages the coach starts: the weekly message and follow-ups. Replies within 24 hours of an owner's message are free today. Rates vary by country.
AI modelsUS$0.26
Replies, deeper analysis for harder questions, the strongest model for money and legal questions, memory, and images (about US$0.12).
Servers and databaseUS$0.08
Google Cloud.
Cost per participant as each source reports it: classroom training per trainee (US$20 to 740 across the trials McKenzie and Woodruff reviewed; US$756 in Togo), a full graduation program per household (BRAC), consulting per firm in a Colombia trial. AI coach: per active user, Peru.
Not everyone is active every month, so in Peru the first year costs about US$2 per person who starts a conversation.
10
Trust
Safety, privacy and quality
Your participants' data
No sensitive personal data is requested: no ID, bank or health details
Encrypted in transit and at rest; access limited to named staff
Data you supply is used only for your coach; you own it, we act as processor, and it is returned or deleted at the end on request
AI providers do not train on it; hosted on Google Cloud in the United States
Safety in every conversation
Topics you choose, such as violence disclosures, are handed to a person you name
Money and legal questions go to the strongest model, with sources checked
Contacts, fees, addresses and rules the sources don't support are caught and flagged before a reply goes out
Misuse is flagged and confirmed by a second model, then warned or suspended
Quality before it ships
Every prompt, model or pipeline change is replayed against 1,500 cases built from real conversations
Every past incident becomes a permanent test case
Facts such as doses, fees and requirements are checked against your manuals and official sources
Watched while it runs
Daily human review of live conversations, including every new participant's first chat
An hourly automated health check and a watchdog for lost messages
Randomized tests decide what changes stay, with results shared
11
The benefits
What changes
Participants
A coach every week, for as long as they want it
On the app they already use, in their language, at any hour
Advice on their own business, with their own numbers
Help doing the work itself
Free to them
Your staff and human coaches
Each visit goes further
Routine questions answered between visits
Practice continues after the session ends
Where linked to your systems: what each participant is working on, and who needs a visit
Sensitive topics handed to the person you name
Your program
Results you can see, during and after
Steps taken and completed, by participant and cohort
Sales, costs and usefulness surveys at months 2 to 10
Dashboards and regular reports
Reach after the program ends, at a cost of cents per participant per month
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.
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. Both courses cost about US$756 per invited participant.
McKenzie and Woodruff (2014), "What are we learning from business training and entrepreneurship evaluations around the developing world?", World Bank Research Observer 29(1). Training cost per trainee ranged from about US$20 (Dominican Republic) to US$740 (Ghana) across the trials reviewed.
Iacovone, Maloney and McKenzie (World Bank WPS 8854). Colombia: individual consulting cost about US$30k per firm and group consulting about US$10k; both improved management practices by 8 to 10 percentage points.
Bruhn, Karlan and Schoar (2018), Journal of Political Economy 126(2). Mexico: consulting raised employment and wage bill by about 50% five years later.
Banerjee, Duflo et al. (2015), Science. Poverty graduation in six countries, with frequent coaching visits as a core component: lasting gains in five of six.
IPA Best Bets: soft-skills training for business owners and consulting for small businesses are 2 of IPA's 14 best bets for poverty reduction.
Otis, Clarke, Delecourt, Holtz and Koning, "The Uneven Impact of Generative AI on Entrepreneurial Performance" (Management Science). A pre-registered randomized trial of a GPT-4 business assistant on WhatsApp with 640 Kenyan entrepreneurs: no significant average effect on revenues or profits. High performers gained and low performers did worse, a gap the authors trace to which tasks owners took on and which advice they acted on.
BRAC puts the cost of a full graduation program at US$300 to 2,000 per household.
The "20 to 30 practices taught, one or two adopted" pattern is what MAIA is designed around: the hard part is putting practices to work.
How the coaching works
Goals, tactics and next steps
Layer
Horizon
What it is
Goal
Months
One lever on the business's profit, named for what it is about ("more lunch orders a day"), ideally with a number and a date. Confirmed in the owner's own words.
Tactic
Weeks
How the goal will move: a new price list, winning back past customers, a bigger pen. The owner's own plan, or a practice that works for similar businesses, checked once with numbers.
Next step
Days
A small dated task with a result to report. Followed up when due, and closed as done (with the number), stalled, dropped, or expired.
A good reply
Answers what the owner actually asked, first.
When money or a decision is at stake, says whether it is right for them, with their own numbers.
Gives one step they can take now.
Asks at most one question.
Simple messages get a fast answer. Harder ones go to an expert pass with the sector's manuals and web search attached, and high-stakes questions (money, legal risk) to the strongest model available.
Proactive follow-up
Channel
When
What it looks like
Day-after touch
About 23 hours after the owner last wrote
Short, with the first part of the next step already done, and one-tap answers. If it asks about a step: "Done / Not yet / I had a problem".
Weekly message
For owners who haven't written recently
A personalized message toward their goal, a check on a due step, or a matching public program or funding call. Owners who go quiet always get a one-tap way to stop.
Some rules are hard limits in code: no message while one is unread, contact limits and opt-outs always respected, less frequent contact for owners who stop answering, and no weekly messages after six months of silence.
The diagnostic
From the conversation, the coach scores each business on eight dimensions, with no questionnaire: customers and market, offer and quality, sales channels, operations and supply, financial records, costing and pricing, people and productivity, and planning and risk. It reads those scores whenever it writes a reply or a proactive message. Owners get what they ask for, and also what they wouldn't think to ask: the diagnostic weighs the topics they choose against the change most likely to raise this business's profit.
Where each owner stands
The coach keeps track of where each owner is in the loop: new, working on a step, a step due, stalled, gone quiet, or paused. That state decides the next move: welcome a new owner, ask how a due step went, adapt a stalled one, pick up where a returning owner left off, re-engage someone who has gone quiet, or step back when they stay quiet.
Memory
The coach remembers the business, the owner's settled facts and decisions, what they tried and how it went, and their open goals and steps. Facts are stored only when the owner's own words back them, and code checks the quote against the conversation. Participants never have to repeat themselves, and the coach doesn't push ideas they already turned down.
Co-developed with local institutions
Sector knowledge bases are built with local institutions (PRODUCE, INIA, ILO curricula) and validated against source manuals. The coaching loop, memory and follow-up cadence come from three years of testing in Peru.
Outcomes, as reported
Sales and costs. In-chat surveys of owners with real businesses who sent 3+ messages: 41% report higher sales at month 2 (n=528) and 43% lower costs at month 3 (n=467). At months 9 and 10 the shares hold or rise (41% sales, n=105; 52% costs, n=68).
Satisfaction. Would recommend at month 5: 83% yes, 4% maybe, 13% no, a net score of +71 on a 3-point scale (n=357). Useful at month 7: 64% very, 19% somewhat (n=377).
Return. From the ranges 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. Halving every self-report still gives about 9 to 1.
Independent evaluation. Colombian coffee farmers (n=102), before and after: formal registration 2% to 17%; 55% directly implemented recommendations from the past three months and 25% partly; accounting records 78% to 89%.
Engagement. Owners active every other week report higher sales about twice as often as those who stop after the first week, so how engaging the coach is matters as much as what it says. About half of new users drop off during onboarding, which is where much of our current testing is focused.
Self-reports carry survivor and responder bias. A randomized trial of the Peru deployment is under review by study funders, and these results come from early versions of the coach.
Who we work with
Every program came to us, each with its own method and population. The pattern repeats: a funded pilot or a short build, then a contract to keep it running.
Partner
Who it serves
How it started, and where it is now
Ministry of Production (PRODUCE), Peru
Micro and small businesses and farmers in the national MSME program
Live since November 2023, with agronomy content developed with INIA. Registered in Peru's national AI catalogue. Where every new feature is built and tested first.
AMPYME, Panama
Small businesses served by the national small-business authority
An ILO-funded pilot after AMPYME's own bot saw little use. Scaled more than 10× to 6,000+ enterprises; now moving onto AMPYME's own budget.
ITCILO (SABY), Kenya and Jordan
Graduates of the ILO's Start and Improve Your Business training
A scoping exercise and a short build of a white-label coach, then 12-month hosting contracts. In English, French and Arabic.
ILO, Colombia and Brazil
Coffee producers in Huila and Minas Gerais
A funded pilot in Colombia with an independent evaluation (July 2024), then a contract in Brazil to August 2027, in Portuguese.
MOVE-ETI, Germany
Participants in Personal Initiative training
A fixed-fee build of a coach that carries the Personal Initiative method, ready for their next training projects.
Also
A lender to small businesses in Uganda, with a data feed back to the lender; a coffee exporter's producer program in Colombia; and a poverty graduation NGO in Latin America, building a coach that works alongside its human coaches.
Two cases
AMPYME, Panama. Panama's small-business authority first built its own bot. Limited functionality and the work of maintaining it meant very little use. It then ran a pilot of MAIA funded by the ILO, scaled use more than tenfold to over 6,000 enterprises, and is moving to funding MAIA from its own budget as its main AI chat service for small businesses.
SABY, with the ILO's training centre (ITCILO). The ILO's Start and Improve Your Business program trains entrepreneurs in classrooms around the world, with little contact afterwards. After a scoping exercise and a short build, SABY went live as a white-label coach for SIYB graduates, carrying the SIYB curriculum and the ILO's own name. It now runs in Kenya and Jordan (in Arabic) on 12-month hosting contracts, with SIYB-specific content owned by the ILO and the code available to them as open source.
The same path, a short engagement on the shared code and then hosting, is what Stage 1 and Stage 2 describe.
Other kinds of programs
Most of MAIA's work is with programs for existing small businesses: national MSME programs and business training. The same coach also fits programs whose participants run small businesses of their own, and each brings methods worth learning from.
Farm businesses
A farm is a small business. Coffee growers use MAIA with the ILO in Colombia and Brazil, and in Peru agronomy content comes from INIA, with IFAD funding the agriculture work in 2026 and 2027. In Peru, 726 users are in agricultural value chains, 411 of them producers (classified from their conversations and checked against blind hand labels). They ask about selling and prices more than about crops, so they use MAIA first as a business coach. The independent study of Colombian coffee growers is on the results slide.
Economic inclusion and graduation
Graduation programs already coach participants in person, with strong views on how. A coach on WhatsApp can work between visits: follow up on the plan agreed in the visit, answer questions, and tell the field coach who is stalled. A graduation NGO is building one now for its programs in Latin America.
Youth self-employment
Young people starting out on their own need the same follow-up after a course or a grant. There is no dedicated youth deployment yet.
Lenders and value chains
A lender or buyer can offer the coach to its clients or suppliers, with agreed data sent back. A small-business lender in Uganda and a coffee exporter's producer program in Colombia work this way.
What is shared, and what stays yours
Every program runs on one shared base: the WhatsApp and AI infrastructure, the coaching method (proactive follow-up, goals and steps, memory), safety guards, and the testing and evaluation machinery. Each program's coach sits on top with its own prompt, knowledge base, number, name and data.
What reaches every program
Fixes, features, test results and lessons about what works. When one program hits a problem, the fix becomes a permanent test case for all of them. Our own testing in Peru ships to every coach.
What stays yours
Your prompts, content, knowledge base and participant data. We never reuse them in another program.
Pooled evidence
Results are pooled across programs only on common instruments and with each program's agreement. Programs can opt in to randomized tests run across partners, which is how the evidence on what works for whom builds up: for example, AI alone against AI alongside human coaches.
What your team brings
People
A program focal point: knows the method, signs off the prompt and tone.
A technical focal point, only if the coach will link to your systems: reviews the integration plan.
Testers: as many as you like. Program staff first, then a few coaches or trainers.
Materials
Curriculum, training guides and any participant handouts
FAQs, local program and referral information
A data dictionary, if linking to your systems; sample or synthetic records are enough, so no real personal data is needed in Stage 1
Decisions
Language and countries for the first version
Whether the coach works on its own or alongside your human coaches
Topics the coach hands to a person, and who receives them
How many testers and rounds, within the six weeks
Building alone vs. with MAIA
Alone
With MAIA
Time to a live pilot
9 to 18 months
6 to 10 weeks
Cost to build
US$80–200k for an in-house team
A fixed fee for Stage 1, scoped with you
Cost to run
US$80–200k a year for an engineer and a prompt specialist, or it degrades
A monthly fee for running and improving it, plus messaging and AI
Team you need
ML engineer, DevOps, conversation designer, analyst
The program specialists you already have
Testing and evidence
One pilot, measured its own way
A 1,500-case test suite, daily review, outcome surveys and randomized tests
Exit
Locked to your own code and the people who wrote it
Open source: take the code and run it yourself at any time
Technical and security detail
Platform
Google Cloud serverless services (Cloud Functions, Cloud Run, Firestore, BigQuery) in the United States (us-central1 and the US multi-region), with Turn.io on Meta's WhatsApp Business Platform.
One codebase; each WhatsApp number maps to its own program configuration, prompt and knowledge base. Partner deployments are separated by configuration and data scoping.
Named accounts for staff. Internal services run under least-privilege technical identities, and each external integration can reach only the data it uses. Secrets sit in Secret Manager, and the WhatsApp webhook verifies a signature on every inbound message.
Encryption in transit and at rest. Administrative changes and data access are logged, and data-access logs are kept for two years. The database is protected against accidental deletion.
Autoscaling with queues to smooth spikes and enforce rate limits, daily usage limits per user, an hourly health check with alerts, and a watchdog for lost messages.
AI models
Current-generation OpenAI models on the main path through MAIA's own routing layer, with Google Gemini on Vertex AI as a second provider and failover. Model choice follows offline evaluations and can change.
What is sent: the conversation turn, the recent history, the business profile and relevant knowledge-base passages, over TLS. Providers do not train on data sent through their APIs, and OpenAI calls are made with response storage turned off. MAIA does not train or fine-tune models on usage data.
Misuse and instruction-injection attempts are part of the test suite. A misuse detector, confirmed by a second model, warns or suspends the user.
Data
Stored: the WhatsApp number and profile name, what the person sends (text, voice notes, photos), survey answers and the coaching record derived from the conversation (goal, steps, reported results). Bank details are removed from messages before storage.
Kept while the person uses the service and until they ask for deletion. On deletion the content is erased and only anonymized counts remain. Backups are kept 7 to 30 days and technical logs up to 30 days.
Staff review conversations to improve the coach, and test cases are built from them with the data that identifies the person removed.
Data a partner supplies about its participants is scoped to that partner's coach and used only there. Outbound data to partner systems is limited to what the partner specifies.
Documentation
An AI system card and a data protection note, prepared for Brazil's government IT and data protection review, are available in English and Portuguese, along with a full technical and security overview. We are happy to complete your own security questionnaire.
Testing and evaluation
Three layers of checking, each answering a different question. Does a change break anything? (the test suite). Is the live service behaving? (monitoring). Does a change help owners engage and act? (randomized tests).
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.
695 of the 1,500 cases are pinned from real incidents and never removed, so a fixed problem stays fixed.
Named coaching failures are screened with exact quotes (answers beside the question, offers instead of delivering, no clear next step, a price accepted without checking the market), and facts are checked against the sector's manuals. A person reads every newly flagged case.
While it runs
A check after every deploy that the new version is actually serving.
An hourly health check: error replies, unanswered messages, safety and unsubscribe keywords, volumes against the week before.
A daily review: coaching quality, every phone number, address, regulation and link checked against sources, safety in five languages, every new owner's first conversation and a random sample, plus checks specific to each partner.
Randomized tests
Experiments run on Evidential, the open-source experimentation engine from IDinsight and The Agency Fund. Each is pre-registered with its hypothesis, measure and decision rule before launch. Seven have been read since May 2026, about 15,000 owner assignments, and the null results are reported alongside the wins. Programs can opt in to tests across partners, which is how the evidence on what works for whom builds up.
9:15