We use AI for parts of our work: research, synthesis, drafting, and analysis. Before a client, or anyone else, relies on that work, it goes through the checks below. They are written down, and they apply to everything a client or third party will read.
We look before we report
We don't say a step worked, or that a result came out a certain way, until we have seen it ourselves. When an AI agent reports that it finished a task, we check the output, not the agent's account of it. Facts come from the primary source: the signed document, the live page, the sent email. A summary of that source doesn't count.
A claim that something is missing names the places we searched, and rests on a second search framed a different way. If we can't check a statement against a source, the work says so and goes out marked unverified.
Every figure keeps its origin
Each number a result depends on is restated with where it came from: the file it sits in, or who supplied it and when. Figures someone gives us in a meeting or an email count as sources and get the same treatment.
When we derive a count or a comparison, we trace one case through by hand. Until the figure has been re-derived, it stays marked provisional in any draft.
Quantities get checked against what the contract grants
When a client or counterparty gives us a quantity to model, such as capacity, area, or headcount, we check it against what the governing document actually covers: the contract, its schedules, its annexures. A number can be correct and still sit outside the scope that applies. Any review finding that touches scope, quantity, or entitlement is checked against that primary document first, however minor the reviewer rated it.
Three independent reviews before anything goes out
Work a client or third party will read, including financial and governance documents, gets three separate reviews before we present it. Separate AI reviewers, on more than one model, each read the whole piece cold with no sight of the others' notes. Each looks for anything wrong, not only the items on a checklist. We take every issue any of them raises, then verify each flagged item against its source.
We use three because a single reviewer proved unstable: its findings varied from run to run.
A check that stops failing gets harder
When a check passes everything for several runs in a row, we tighten it: harder criteria, a fresh line of attack, a new reviewer with a different brief. A clean streak is a prompt to harden the check. It isn't evidence the work has no defects left, and the checks have to stay ahead of the work they test.
