Results | MAIA, MYPE AsesorIA

Independent evaluation: Colombia coffee pilot

The ILO's FORLAC programme commissioned an external monitoring and evaluation consultancy to run baseline and endline surveys of coffee farmers in Colombia's Huila region who used MAIA, in coordination with the Comité Regional de la Cadena de Café del Huila.

Metric	Result
Businesses with dual formal registration (RUT + Cámara de Comercio) Formalization is notoriously hard to move.	2% → 17%
Implemented MAIA's recommendations directly in their farm or business (past 3 months) A further 25% implemented them partially.	55%
Keeping accounting records	78% → 89%
Minimizing waste and residues in harvest	72% → 87%
Setting clear, measurable goals	89% → 94%
Harvesting only mature beans	80% → 85%
Fertilizer application per technical requirements	88% → 91%
Using AI tools (sometimes or constantly) Expected, since MAIA itself is an AI tool; reported as adoption, not an outcome.	18% → 71%

Methodology: Baseline and endline surveys with a 58-question instrument covering business management, formalization, decent work, and technical practices. All baseline-to-endline comparisons use the panel of n=102 MAIA users surveyed at both rounds (144 total surveyed at endline, including 27 users without baseline data and 15 non-users). 60% of surveyed users were women. This is a pre-post design without a comparison group: it measures change among participants, not causal impact. With n=102, small movements (under ~10 percentage points) should be read as stability, not change. Evaluation reports (Producto 3, March 2025, qualitative; Producto 4, June 2025, endline quantitative) available on request: [email protected].

User-reported outcomes and experience

MAIA asks users about results in-chat at fixed tenure milestones. Latest responses, November 2025 to June 2026:

Metric	Result
Average self-reported sales gain per respondent (n=239) The median respondent reported no increase; the average is driven by the 44% who did. Full distribution: 56% no increase, 31% up by S/150 (~$40)/month or less, 13% up by more.	~S/43 (~$11)/mo
Average self-reported cost reduction per respondent (n=236) Full distribution: 58% no reduction, 28% down by up to S/150 (~$40)/month, 14% down by more.	~S/42 (~$11)/mo
Net satisfaction, month 5 (n=94) 83 would recommend MAIA, 7 would not, 4 unsure. Simplified 3-point scale, not a 0-10 NPS.	+81
Usefulness, month 7 (n=136) A further 24% “somewhat useful”.	63% “very useful”

Methodology: In-chat button surveys sent at fixed tenure milestones (months 2-10). Latest response per user; MAIA test numbers excluded. Averages use bin midpoints with the open top bin counted at its floor (S/150), zeros included, so the true means are higher. These are survey responses, not verified sales records, and reported sales gains are revenue, not profit; do not add the sales and cost figures together (different respondents). Respondents are active users replying in-chat, so results are subject to survivorship, courtesy, and non-response bias. Directional evidence, not causal.

Reach and engagement

Live operational metrics from MAIA's production database.

Metric	Result
Entrepreneurs reached	13,640
Engaged beyond onboarding (3+ messages)	10,930
Monthly active users (May 2026)	3,547
Messages exchanged (all time)	1,025,377
Where Peru 6,965 · Panama 4,505 · Colombia 1,093 · ILO SIYB pilots (incl. Jordan, Kenya) 818 · other 259.	5 countries
Still active 18 months after starting n=208, 95% CI ~15-26%. 18 months of data, excluding two 2024 third-party door-to-door recruitment campaigns. Firms up as larger 2025 cohorts pass the 18-month mark.	~20%

Methodology: Internal database query (BigQuery), June 12, 2026; MAIA test numbers excluded. Reached = distinct WhatsApp numbers with at least one inbound message, all time. Engaged = 3 or more inbound messages (i.e., interacted beyond the onboarding minimum). Active = distinct users sending at least one message in the calendar month.

The measurement pipeline

MAIA is run as a learning system, not a finished product. The numbers above are a snapshot of a deliberately staged measurement program, following the four-level AI evaluation framework from The Agency Fund and the Center for Global Development: model behavior, then product use, then user-level change, then independent impact evaluation.

Continuous randomized experiments. Coaching designs are tested head to head in pre-registered A/B experiments with primary metrics defined in advance, from outreach framing to check-in design. Winners ship; results are documented either way.
Behavioral execution measurement (REX). We are building and validating a measure that detects, in the conversations themselves, whether users actually executed what their coach proposed. It is validated against survey outcomes and designed to complement self-reports with observed behavior at full scale.
Longitudinal outcome tracking. Since April 2026, in-chat outcome measures repeat at months 9 and 10 of tenure, so the same users' sales and cost reports can be compared over time.
Independent and experimental evaluation. The Colombia evaluation above was pre-post; we are open to working with research partners on experimental designs. If you are a researcher interested in studying MAIA, write to us.