AI playbook for Programmes and Monitoring: five to delegate, three never

This is the first of seven department playbooks we are publishing properly. Six of the seven departments in most organisations we work with are named in adoption plans and never actually illustrated: everyone agrees AI belongs somewhere in Programmes, in Finance, in HR, and nobody writes down what that actually means task by task. Programmes and Monitoring goes first, because it produces more raw data than almost any other department, and more of that data is genuinely safe to hand to a tool than most teams currently assume.

What follows sorts the department's real work into two piles: five tasks worth delegating, and three that should never be handed over, whatever the pressure to move faster.

What the department actually spends its time on

Programmes and Monitoring lives on indicator tracking against logframes, donor report drafting, data collection oversight, beneficiary feedback review, and monitoring visit write-ups. Most of it is mechanical aggregation of numbers that already exist. A meaningful slice of it is judgement about what a specific person's experience actually means, and those two categories need to be treated completely differently.

Five tasks worth delegating

• First-draft summaries of quantitative survey data. Turning clean indicator numbers into a narrative first draft is aggregation, not interpretation, and a tool can do the mechanical part correctly.
• Cross-referencing indicator definitions across donor logframes. Flagging where two donors define "beneficiary reached" differently is a matching task, not a judgement call, and it catches errors a tired analyst misses at month end.
• Drafting routine donor report sections that restate already-approved data. Once the numbers are approved, turning them into report prose is exactly the kind of task that should never eat a programme officer's afternoon.
• Organising field notes into a searchable format. Transcribing and tagging raw field notes so they can be searched later is clerical work with no judgement attached to it.
• First-pass charts of indicator trends over time. Producing a draft visualisation from clean numeric data is mechanical, and a human still decides which chart actually goes in the report.

Three tasks that never get delegated

• Interpreting what a beneficiary's qualitative feedback actually means. A complaint or a comment read out of local context can be misclassified in ways that matter, including missing a genuine safeguarding concern hiding inside an ordinary-sounding sentence.
• Signing off on whether a programme met its stated outcome. This is a judgement call with organisational accountability attached to it, and that accountability cannot be delegated to a tool that will never answer for the decision.
• Handling any case-level data that identifies a specific vulnerable individual. We have written before about the four-class data table: casework and safeguarding records sit in the class that never touches a general-purpose tool, whatever efficiency gain is on offer.

The five tasks save time. The three tasks protect judgement. Confusing which is which is how a monitoring department loses the trust it exists to build.

The trade-off in drawing the line this precisely

Naming exactly five delegable tasks and exactly three that are not invites an obvious pushback: the line will move as the tools improve, so why fix it in writing now. The honest answer is that the three "never" tasks are not there because today's tools are not capable enough. They are there because they carry a kind of accountability that does not transfer to a tool regardless of how good it gets, and pretending the line is only about current capability sets the organisation up to quietly erode it the moment a more capable model arrives.

The practical takeaway

Delegate aggregation, drafting, and mechanical cross-checking in Programmes and Monitoring; never delegate the interpretation of a beneficiary's experience, the sign-off on programme outcomes, or anything touching case-level safeguarding data.

No external statistic cited; this article presents an internal department playbook rather than third-party evidence.