Adelbury Labs · Prototype
Claims Ledger records what organisations say they will do, when they said it and when it can be checked. AI agents do the filing and people do the judging.
Ledger as generated on 9 October 2026.
In March, a chief executive says a new plant will open next year with 350 jobs. It makes the trade press and then everyone moves on. By next year, almost nobody remembers the date or the number. We wanted to know whether AI agents could keep an accurate record of promises like this, good enough that a team could come back months later and test them.
Comms teams write these promises. Reputations depend on whether they’re kept. Journalists, investors and campaigners are getting better at going back to check, and the first sign is usually a question you didn’t see coming.
Most organisations keep this knowledge in people’s heads, old press folders and annual reports nobody rereads. A dated record of what you, your competitors and your stakeholders have committed to shows you the question months before anyone asks it.
Agents read published material such as annual reports, results statements, council papers and company announcements.
They pick out forward commitments that carry a number or a date, and skip strategy language and past results.
Each claim goes into the record of the organisation that made it, with who said it, when and a link to the source.
Each claim is given the month when it can first be tested.
A script gathers every claim into one list, sorted into testable now, due within six months and later.
When a claim falls due or new material touches it, the system flags it and a person decides what it means.
Every claim is filed with who made it, where it was said and the month it can first be tested. Notes record the context a person will need when the date comes round.
The company, people and claim are fictional.
Early batches padded the claims with vague statements and cited a data-resale site, because nobody had written down what counted as a claim or a source. Those rules now live in the system itself, not in a one-off prompt.
A claim needs a number or a date, and checking it later has to tell someone something. If an organisation has made no promise worth recording, its record says so.
In one batch we included a role with no named holder, to see whether the agent would make up a name. It didn’t. It found the correct name on the organisation’s own website.
Every change to the rules is logged with what was rejected and when to revisit it, so the next agent or the next person doesn’t reopen settled questions.
Claims Ledger is a prototype that we’ve used in our own publishing work since September 2026. It isn’t a product you can log in to. Next, we want to test the same approach on one client’s set of competitors and stakeholders.
We build systems like this for clients. Agents monitor, archive and categorise new signals, and your team keeps the judgement.
See our AI intelligence systems service