AI in M&E: what is real in 2026

Every M&E conference in 2026 has a session on AI, and most are either sales pitches or warnings. Both are unhelpful. The useful version is a list of what the technology does today, in this discipline, with real field data, and what it does not. Here is ours, from the inside of a platform that ships these features.

What works now

• Drafting a survey from a description. Twelve questions on household income for a rural livelihoods programme, phone-administered, produces a defensible first draft that a person who knows the programme edits in an hour. A real half-day saved.
• Transcription. Recorded interviews and focus groups transcribed in minutes, including accented English and the code-switching that defeated tools two years ago. The transcript is a base for coding, not the analysis itself.
• First-pass anomaly flags. A household of fourteen earning 8,000 a month, an enumerator completing forty visits in a day, a GPS point ninety kilometres from the assigned site. Flagged for a human. Never deleted by a machine.
• Narrative drafts. Given a month of clean data, a model writes a passable first draft of a report's factual sections: what happened, against which target, in which districts. It is mostly boilerplate, which is the point. The boilerplate was always the expensive part.
• Assisted coding of open-ended answers. A thousand free-text replies to 'what would improve the training' grouped into themes in minutes, each quote linked to its source. A person then reviews the groupings, merges two themes, splits one, and the code frame holds a line back to every response it claims to summarise.

What does not work

• Unsupervised analysis anyone would sign. Left alone with a raw dataset, a model produces plausible paragraphs with confident numbers, and some of the numbers are wrong in ways only an analyst who knows the programme would catch. Nobody should sign that.
• Numbers without a traceable source. If a model states that completion in one district was 71 percent, that has to be because 71 percent is computable from submissions, not because 71 percent sounds likely. Generated figures with no query behind them are worse than no figures. They are confidently wrong in your donor's report.
• Judgement. Whether a bad month needs a paragraph or a phone call, whether a partner's explanation holds, whether the revised target is honest. These are calls about context and consequence, and the model has neither.

The rule

One rule keeps this entire area safe: every figure traces back to a submission. When AI drafts a report, each number it places is a reference to real records, and clicking the number shows them. When AI flags an anomaly, the flag is attached to the specific submissions that triggered it. When AI transcribes, the audio sits behind the text. The moment a number can exist without its source, the tool has left monitoring and entered storytelling.

Two guardrails sit beside the rule. Your data should not leave your tenancy to be processed, so ask where inference runs and what is retained after a request. And field data is usually personal data, which means the access controls that apply to a submission apply to anything a model produces about it. AI features that cannot answer either question are features you cannot put in front of a donor.

Where this goes

The direction of travel is boring and good. Drafting gets faster, anomaly detection gets sharper, transcription becomes assumed infrastructure. The judgement layer, deciding what the numbers mean and what to do about them, stays with people who know the programme, because that layer is not a text problem. A tool that respects the line, automating the mechanical and citing everything, is worth piloting this year. A tool that offers conclusions without sources is not worth piloting ever.

The teams getting value today are modest about it. They pilot on one report section, check every citation for two months, and extend only what survives checking. That pace is not timidity. It is how a signature on a report stays meaningful.

If you are evaluating claims, ask one question of every AI feature: show me where this number came from. A good system answers with records. A bad one answers with a paragraph.