Mention artificial intelligence in a conversation about social impact and the discussion tends to leap toward the speculative — algorithms deciding who receives a bursary, predictive models forecasting which communities need intervention before anyone asks them. That version of AI in development work is mostly still theoretical, and arguably should stay that way. The version already changing how well-run programmes operate is far less dramatic, and far more useful: AI as an administrative tool, quietly absorbing the hours that used to disappear into paperwork.
That distinction matters, because it changes what a programme manager should actually be evaluating when a vendor pitches an "AI-powered" solution. The question worth asking is not whether the tool is impressive. It is whether it gives staff back time they were previously spending on tasks a machine can do adequately, so that staff can spend more of their time on the judgement calls only a person should make.
Where the hours actually go
Anyone who has run a bursary, learnership or grant programme at scale knows where the administrative burden actually sits. It is in screening hundreds of applications against eligibility criteria that are simple in principle and tedious in volume. It is in chasing supporting documents — certified IDs, academic transcripts, proof of income — from applicants who submit them incomplete, late, or not at all. It is in compiling the same beneficiary data into a different report format for every funder, every quarter, by hand. None of this is intellectually demanding work. All of it consumes a disproportionate share of programme staff time that could otherwise go toward the parts of the job a person is actually needed for: engaging beneficiaries, solving problems on the ground, exercising judgement in ambiguous cases.
This is the terrain where AI tools currently do genuine, unglamorous work — not because the technology is exotic, but because the tasks it is replacing were never good uses of a skilled person's time to begin with.
Practical AI, today
Applied sensibly, AI can pre-screen applications against defined eligibility criteria and flag the clear passes and clear fails, leaving genuinely borderline cases for a human reviewer to decide. It can check submitted documents for completeness and basic consistency before they reach a case officer, catching the missing-page-and-wrong-date errors that otherwise surface only after a delay. It can flag beneficiaries showing early signs of disengagement — attendance dropping, milestones slipping — so a programme coordinator follows up while there is still time to intervene, rather than discovering the dropout after the fact. And it can produce a first draft of a quarterly report from structured programme data, turning a task that used to consume days into one that consumes an editing pass.
The common thread across every one of these applications is that a human being remains the decision-maker. The tool narrows what a person has to look at and speeds up how quickly they can look at it. It does not, in any of these uses, decide on its own who receives funding, who is removed from a programme, or what a report ultimately says.
Dashboards beat documents
A quieter but equally consequential shift is happening alongside AI specifically: the move from quarterly PDF reports to live programme dashboards. A quarterly report is a snapshot, already out of date by the time a funder reads it, and expensive to produce in staff time each cycle. A well-built dashboard, drawing on the same underlying data, shows a funder or a board what is happening now — enrolment against target, spend against budget, early outcome indicators — without anyone needing to compile a document first. For programme managers, the same dashboard becomes an early-warning system rather than a retrospective scorecard, surfacing problems while they are still small enough to fix.
This shift does not depend on AI at all; it depends on programme data being structured and centralised well enough to be queried live. But it is usually the same infrastructure investment that makes AI-assisted screening and flagging possible in the first place, which is why the two tend to arrive together.
The guardrails
None of this is a case for deploying AI without discipline. Beneficiary data in South African social programmes is personal information under POPIA, and any tool processing it needs a lawful basis, appropriate consent, and real data-security controls — not an afterthought bolted on once the system is live. Screening tools trained or configured carelessly can encode bias as readily as a careless human reviewer can, quietly disadvantaging applicants who do not fit a pattern the tool was tuned on, and that risk needs active monitoring, not a one-off check at launch. And no beneficiary decision — who is admitted, who is removed, who receives funding — should ever be made by a model without a human accountable for the outcome.
That is a modest claim compared to the more dramatic promises circulating about AI in development work, and it is deliberately modest. Social impact management does not need artificial intelligence to reinvent judgement. It needs the administrative weight lifted off the people exercising that judgement, so they can exercise it more often, and sooner.