How we work

We build our own data pipelines, regular reports and statistical insight with teams of AI agents, under written rules, review and recorded checks. We hold our own agent teams to the same six controls we ask of any figure, as set out in Traced to Trust.

The ladder has five rungs: Tacit, Recorded, User-checked, System-enforced, Tested independently. The system-enforced rung is the step change, because the pipeline holds back work that lacks the control, whoever is on shift. Below we place each control, and say why. Counts come from our own release records, figure files and status notes for our data work, read on 21 November 2026. Times taken from version history are lower bounds, since the visible history may be incomplete.

Where each control sits on the ladder
ControlWhere we place itWhat holds it
TimelinessSystem-enforcedRe-check against current data at release
RecordSystem-enforced, for one named error classAn engine rule added after a miss
AttentionSystem-enforced for the spend gate; recorded for approvalsA usage gate; approval records
ComparabilityRecordedA written rule, applied in review
EvidenceSystem-enforced for stored values; recorded for source linksEngine refusal; figure records
DutiesSystem-enforcedRelease credentials held apart from preparation

No control has been tested independently. That is covered at the end of this page.

Timeliness

Timeliness is checked twice: when a figure is fetched, and again at release, against current source data. Of 76 release records, 5 were held at release because the re-check found the figure stale or superseded. The re-check sits in the release step, not in anyone's memory, so we place it at the system-enforced rung. It shows what the re-check caught, not that every provisional figure is stopped: the Record section describes one that was not.

Read more: Flash versus final.

Record

On 15 November 2026 a note recorded that published items overstated a fall, because the source series fills in late. Each had matched its source when prepared; the source was then revised. A change to the engine, dropping unfinished periods and the newest months of series marked as late-filling, followed 68 minutes 38 seconds after that note. On 16 November the engine refused a later draft in the same series.

For that class of error we place Record at the system-enforced rung. We do not claim every past correction has been turned into an engine rule; many lessons sit in written notes, which is the recorded rung.

Read more: If it isn't written, it wasn't learned and You cannot instruct agents into compliance.

Attention

We aim to spend human judgement where risk leaves the firm, and let checks handle the rest. Of the same 76 records, 50 carry a recorded human approval; we have not checked each for who gave it. Over those records, 11 items were held back at release or withdrawn afterwards because the data had moved on or failed a check, each with its reason recorded. Because approvals are recorded but not yet tested for whether they were read rather than ticked, we place approvals at the recorded rung.

Spend is the part held by the system. Our usage gate checks weekly use before scheduled work starts. On 18 November it returned stop at 98% weekly use, and no new scheduled work started until the reset on 19 November. That gate is system-enforced. It is not the whole answer: in the week to that reset, spend passed the budget set for the week, and a person reading the billing page found it, not the gate. That miss became a second gate that measures spend in euro against the budget; wiring it fully into live running was still open on 21 November, so it is not yet counted as system-enforced.

Read more: More approvals can make you less safe and Your AI budget needs a brake, not a report.

Comparability

Our rule is that two figures are not charted side by side until we can say whether they measure the same thing, and that when they do not, the useful answer is "cannot be compared, because...". The rule is written down and applied in review. We do not claim the system blocks every unruled pairing, so we place Comparability at the recorded rung.

Read more: Two numbers side by side are not a comparison.

Evidence

Our agents are not permitted to type a figure. Figures are fetched from source by code and stored with their evidence, and the engine is designed to refuse a figure that does not resolve to a stored value. That refusal is system-enforced.

The source trail is less complete. Of 9,435 figures logged, 9,377 carry both a source link and a fetch time; the other 58 carry a fetch time but no link. We have checked that the fields are present, not that each link resolves or each value still matches the source. Since a missing link does not yet stop a figure, the link sits at the recorded rung.

Read more: Accuracy is an architecture decision and Code for the exact, agents for judgement.

Duties

The agent that prepares an output does not hold the credentials to release it. A separate, restricted process holds them, re-checks the work against current source data, holds the release if the figures have moved, and writes a record. Of the 76 records, 8 were stopped before release by a check rather than by anyone's attention. We place Duties at the system-enforced rung.

The same separation applies to our engineering work. Changes land through pull requests that must pass automated checks and a review. As at 21 November 2026, 215 had been closed without being merged, against 3,157 merged. That count includes superseded and duplicate work, so it is not a rejection rate, but a control that has refused nothing cannot be told apart from one that is not running.

Read more: Segregation of duties applies to digital workers too and A control that never says no is not a control.

Where we are not yet at the top rung

No control here has been tested independently. No party outside the team has tested any of them on a sample and reported what it found. That is the next step, and it comes after the system-enforced rung for a reason: auditing a habit only confirms it is a habit.

The clearest gap is removal. Our correction policy sets a one-hour limit for a factual error to stay public. Three items withdrawn for the late-filled series were found at 11:40 Irish time on 15 November and deleted on the morning of 21 November, almost six days later. The policy had settled who could decide on a takedown, not who could carry it out, and our agents had no way to delete a live item themselves.

The miss became a written runbook the same evening. The clock starts when the error is found; within 15 minutes, a named role that already holds the access is asked to remove it; a reminder repeats every hour until removal is confirmed; anything queued from the same source is paused; and each found-to-removed time is recorded. The runbook has not yet met a real incident, so removal sits at the recorded rung until it is drilled. We do not claim it will meet the hour. The next takedown will be timed and on record, and measured against it.

Read more: Decide who pulls the plug before you need to.

What we share

  • Agentic team insight. What we learn from running AI agent teams under governance a finance function would recognise, written up for organisations moving off spreadsheets and email sign-offs.
  • Governed building. How we build our own data pipelines, regular reports and statistical insight with agent teams under written rules, review and recorded checks.
  • Controls for agent-built work. The checks we use, segregation of duties, evidence trails and human oversight placed where the risk leaves the firm, offered as a method any organisation can apply to its own work.