Work & case studies · Professional Services
Multi-agent research and reporting system for a consulting firm
A team of specialised AI agents that gathers sources, cross-checks facts and drafts cited briefings, cutting first-draft research time from days to hours.
Consultants at a strategy firm were spending the first days of every engagement building market briefings by hand. We designed a multi-agent system, part of our ai-agents practice, in which planner, researcher, fact-checker and writer agents collaborate to produce cited briefings from public sources and the firm's own archive. A reviewed first draft now takes around four hours instead of two to three days.
Challenge
Every kick-off needed a market-sizing note, a competitor landscape and a regulatory summary. Analysts assembled these from web research, past decks and paid databases, and partners routinely found unsourced numbers that had to be traced back at the last minute. A single-prompt "write me a report" experiment had produced fluent text with invented figures, which set the bar: anything we built had to show its sources or say nothing.
What we built
- A planner agent that decomposes a brief into research questions and assigns them to researcher agents with different tools: web browsing (Playwright), the firm's document archive via retrieval (
rag-systems), and a licensed data API. - A fact-checker agent that re-verifies every quantitative claim against its source passage and downgrades or removes anything it cannot confirm.
- A writer agent that assembles the briefing in the firm's house structure, with inline citations that resolve to the exact page or URL, and an explicit "gaps and uncertainties" section.
- Human checkpoints after planning and before final assembly, plus a run log showing which agent produced which paragraph and why.
- Cost and time budgets per run, so a briefing cannot silently balloon into hundreds of model calls.
Results
First-draft briefings that took two to three days now arrive in roughly four hours including review. Every claim in the output carries a resolvable citation, and analysts report about 35% less time on engagement kick-off work. The gaps section proved unexpectedly valuable: partners use it to decide which primary interviews to prioritise. Related pillar: generative-ai-llm for the drafting layer.
What made the difference
We treated the fact-checker as the most important agent, not the writer. Splitting "find" from "verify" from "write" made each step testable on its own, and gave the firm something it could trust in front of a client.
It does the legwork the way a diligent junior would and, unlike a junior, it never forgets to cite.
Next case study
Computer-vision defect detection on a packaging line
Ready to chart your AI course?
A free 30-minute consultation: we listen, ask hard questions and tell you honestly where AI will and won't help.