AI can draft M&E narrative and charts from settled data, but outlier calls, gap-filling, causal claims and summary categories must stay human or the dataset is quietly corrupted.
The last piece in this series was about keeping AI-assisted research traceable back to a source. Monitoring and evaluation raises the same risk one level up: M&E is a data-integrity discipline first and a reporting discipline second, and the moment AI assistance touches the data layer rather than the reporting layer, a corrupted number can enter a dataset looking exactly as legitimate as a correct one. Nobody catches a fabricated citation by reading it twice. Nobody catches a silently smoothed data point either, and this dataset gets used to decide whether a programme continues, scales, or closes.
This piece draws the line between where AI genuinely helps in M&E work and which specific steps have to stay human regardless of how capable the tool gets.
AI assistance is genuinely useful on the reporting layer sitting on top of already-verified data: drafting the narrative sections of an M&E report from an approved data table, generating a first-pass visualisation from clean figures, writing the standard methodology boilerplate that appears in every quarterly report, or reformatting a dataset for a different funder's template. None of this changes a single underlying number, which is exactly why it is safe.
The risk sits specifically in tasks that look like tidying but are actually judgement calls: interpolating a missing survey response as though it were real data, summarising a spread of answers into a single confident sentence that quietly drops the distribution, or inferring a causal story from a correlation because a causal story reads better in a report. Every one of these produces output that looks more polished than the raw data did, which is precisely what makes the corruption invisible until someone tries to reconcile this year's figures against next year's and finds they no longer mean the same thing.
A summary that hides the distribution is not a simplification. It is a different dataset.
Keeping these four steps human costs real M&E lead time, and it costs the most during the busiest reporting periods, which is exactly when the temptation to let a fast, capable tool touch the raw data is strongest. The honest answer is that this is precisely the wrong moment to relax the boundary: an M&E dataset compounds silently across reporting cycles, so a corrupted baseline this year becomes a corrupted comparison next year, and by the time anyone notices, the affected decisions about programme continuation or scaling have already been made on bad data.
Let AI draft the narrative and visualise the numbers once the data is settled, but keep outlier judgement, gap-filling, causal claims, and the translation of raw responses into summary categories entirely human, because these are the specific points where a dataset can be quietly corrupted in a way that looks like tidying.
No external statistic cited; this article presents an internal trade-off analysis rather than third-party evidence.
If this is the question on your desk, a thirty-minute call tells you whether the service fits, or that you do not need us yet.