Case study: AI agents that research organizations

Five days of manual research, done in eight minutes

Ten AI agents working in two phases. And one rule that is not negotiable: no finding exists unless it carries its source and its date.

5 days → 8 min

One full research run

end to end

Verifiable findings

each with source and date

Agents across two phases

with a dependency barrier

Findings without a source

the schema does not allow it

The starting point

A strategic alliances consultancy in the social sector needs to know, before bringing two organizations to the same table, who each one really is: how it is incorporated, who runs it, where its funding comes from and what it has done so far. That research was done by hand, by an expert, one source at a time.

What hurt

Each research run took around five days of work from someone with expert judgment.

The bottleneck was not the analysis but the collection: searching, opening, reading, verifying and noting.

The information is scattered across public registries, own websites, press coverage and financial statements locked in PDFs.

With similarly named organizations, a single misattributed fact contaminates the entire assessment.

What was built

Ten agents in two phases, not ten loose agents

Six collect in parallel: five search the web and one reads public PDFs to interpret financial statements. The four synthesis analysts only start once all of them have finished. The barrier is deliberate — an analyst synthesizing over an incomplete corpus will confidently produce false conclusions.

Instrumented unit economics

Every agent reports tokens, searches, latency and cost, visible in the interface. A research run has a known, predictable cost before it is launched, not a surprise at the end of the month.

Failure handling and a finished deliverable

Paused turns are resumed and good results are protected when a later run fails: in a ten-agent system, one agent falling over cannot cost the work of the other nine. The final output is a three-page document exportable to PDF, not a data dump.

Six controls that keep the system from inventing

Hallucination is not something you patch: it is a property of the model. What you design is a system where a wrong answer never reaches the deliverable. None of these controls depends on asking the model to behave.

A fixed JSON schema per agent

Every agent answers inside a closed structure. It cannot add a field that does not exist, or return prose where a value is expected.

Source and date are mandatory

Every finding carries them. It is not a prompt instruction: without a source the finding does not fit the schema, so it does not exist.

An explicit channel for what was not found

"I did not find it" is a valid, recorded outcome. When a model is not allowed to abstain, it invents.

Identity anchors

Before a fact is accepted, the system verifies it belongs to the organization under research and not to a similarly named one. A single misattributed namesake contaminates the whole assessment.

A dependency barrier between phases

The synthesis analysts do not start until all six collectors have finished. Synthesizing over an incomplete corpus produces false conclusions with complete confidence.

Expert approval as the gate

No deliverable ships without human review. The system removes the collection, not the judgment.