Who they are

They win contracts to design and run donor-funded development programmes around the world. Winning a programme is the start, not the finish: each one then has to be staffed with specialists whose experience matches what that programme demands.

So the reading never stops. Every programme they win brings a fresh set of required specialisms and a stack of candidate CVs to judge against them. The more they win, the more people there are to assess.

The work, as it actually happened

Each programme they won had to be resourced, and a new one came round roughly every six weeks. That meant hundreds of CVs to work through, each judged against the specialisms that particular programme called for: public financial management, inclusive economic growth, media freedom, democratic institutions, and so on, and how much real experience the person had in each.

This was never a keyword match. It was expert judgement: a subject-matter expert reading a CV and deciding whether the candidate had real, direct experience in the areas that mattered, or just happened to mention the right phrases. Hundreds of times, for every programme.

And it fell to the people least able to spare the time. The screening that stood between winning a programme and getting it staffed was done by the same senior specialists the delivery itself relied on.

Why it hadn't been fixed already

Because it looked like a judgement job rather than a clerical one, and judgement was the part nobody expected a tool to handle. Ordinary CV filters match keywords, which is worse than useless here: they reward a candidate for writing “public financial management” and can't tell whether the experience behind the phrase is real, direct and relevant. So the work stayed with the experts, because only an expert seemed able to make the call. Automating it would mean automating the judgement, and that wasn't on the table.

What we did

We learned how the experts actually judged a CV. Before writing anything, we sat with how a reviewer really reached a verdict — which specialisms mattered for a given bid, what counted as genuine direct experience, how depth was weighed — so we were automating their judgement, not our guess at it.

We built the judgement into the workflow, not just the plumbing. CVs land in their document library, and the workflow reads each one and scores the candidate against the specialisms and experience levels that programme requires, writing the results into a spreadsheet the team works from. Moving the files and the results around was the straightforward part, that's APIs. The real work was the analysis: engineering the prompts so the model assessed the right capabilities the way a subject-matter expert would, crediting real direct experience in a theme and refusing to be swayed by a buzzword that appears on the page.

We proved it matched their experts. This is the part that mattered most, and the reason it could be trusted. We didn't ask anyone to take an AI's word for it. We ran the workflow's scoring against the experts' own past reviews of the same CVs and tuned it until it reached the same judgements they had. Once it agreed with the experts, it could take over that screening pass.

We built it to be re-pointed, not rebuilt. Every programme asks for a different mix of specialisms, so the workflow was made to take a new set of criteria rather than be rebuilt from scratch each time. A new programme, in a new country, was a change of criteria, not a new project.

We handled the data carefully. CVs are sensitive personal information, spread across people in many countries, so how the data was accessed, where it was processed and how long anyone retained it were designed in from the start, including that the AI providers in the workflow hold none of it afterwards. On a project like this, that isn't a footnote; it's a condition of starting.

None of this was data being shuffled from one place to another. It was an expert judgement, made hundreds of times for every programme, rebuilt as an automated workflow with the analysis at its centre, and held to the standard of the experts whose judgement it took on.