What 11 weeks and 8 AI agents taught us about building for real teams

We spent a quarter building eight AI agents for Scout Talent, running the whole programme as one focused sprint. Some agents are live and doing real work. One got cut entirely once we realised it wasn't needed. Here's what that process actually looked like, and what we'd tell anyone starting a similar build.

Start with the process, not the model

The first agent we shipped wasn't the most technically interesting one. It was a candidate summary agent that reads interview transcripts and recordings and produces a structured summary automatically, removing a manual download-and-upload step recruiters were doing by hand.

It's live and being used daily. The unexpected win: it also gave the business clean visibility into how many interviews each team member runs each month, data they didn't have an easy way to pull before. That's the pattern we look for. The best first agent is rarely the flashiest one. It's the one sitting on top of a process people already do, badly, by hand.

Ship the boring agent first

The second agent automates something genuinely unglamorous: logging into a job board and posting ads on a consultant's behalf. It's live in two parts of the business. The first version used a browser-based flow. Phase two replaces that with a direct API integration, because browser automation is fragile and an API is not.

Boring problems are usually the right place to start an AI programme. They're well understood, the value is obvious, and nobody is emotionally attached to the old way of doing it.

The best agent is sometimes no agent

Partway through the sprint, we descoped one of the eight agents entirely. It had been scoped as a landing page generator. Once we got into the workflow, it became clear the actual problem didn't need an agent at all. So we removed it from the plan.

We think this is worth saying out loud because it's the part most vendors don't put in a case study. Give away the recipe: if a client is paying for outcomes rather than agents, then not building something is sometimes the most useful thing you can do. It also builds the kind of trust that gets you invited back for the next project.

Governance can't be an afterthought

The most sensitive agent in the programme lets recruiters query candidate application data in plain English, things like "show me applicants with five or more years in sales." Before this went anywhere near production, we built and tested PII tokenisation and confirmed how personal information would be handled. That work happened alongside the build, not after it.

If your business handles personal data (most do), governance isn't a phase you bolt on at the end. It's a design constraint from day one.

What this means if you're thinking about your own agent programme

Eight agents in eleven weeks only worked because we treated it as a portfolio, not eight separate projects. Some things shipped fast. One got cut. One is still in exploration because the governance questions needed more time than the build did. That's a healthy ratio, not a failure rate.

If you're weighing up where to start with AI in your own business, we're happy to talk through it honestly, including telling you if you don't need an agent at all.