A configuration tells you what an AI workflow does today, not why it was built that way. The Rationale Ledger keeps the reasons, especially the boundaries, alive through staff turnover and edits.
Consider an illustrative scenario. A programme team rebuilds its monthly donor report workflow. In March, an AI assistant is configured to leave out any beneficiary story that names a village, because a case review raised a protection concern. In September, the colleague who set it up moves to another post. In October, a new team member adjusts the workflow because the reports "feel thin", and the exclusion quietly disappears. Nothing was hidden and nobody was careless. The reason the rule existed lived in one head and one email thread, and both are gone.
This is the quiet cost of AI adoption in mission-driven organisations. Tools, prompts and agents change every few weeks. The decisions behind them, meaning what was delegated, what was ruled out and what evidence justified it, change more slowly and are almost never written down where the next person will look. Our view is that the durable asset of an AI programme is not the configuration. It is the reasoning attached to it.
A system you can rebuild but cannot explain is a system you will eventually break.
Organisations have always lost knowledge when people leave. Three features of AI work make the loss faster and harder to see.
First, the artefacts are soft. A prompt, a project instruction or a skill file is plain text that anyone can edit in seconds, with no review step and no version history unless someone builds one. Second, the rules embedded in them are usually reactions to something: a near miss, a funder request, a data protection concern. The reaction is encoded. The event that caused it is not. Third, platforms change constantly, so people revise these artefacts for reasons unrelated to their original purpose, such as a new model behaving differently, and remove protections as collateral damage.
The result is drift. The working system slowly diverges from the intent that justified it, and nobody can say when.
We use a simple device we call the Rationale Ledger. It is one page per AI-supported workflow, held by a named owner, with five entries that together answer the question a successor will ask: why is it built this way, and what happens if I change it?
The ledger is deliberately short. A ledger that takes an hour to update will not be updated. Aim for something a colleague can read in three minutes and act on.
The configuration tells you what the system does today. Only the ledger tells you what it must never do tomorrow.
Most organisations already hold parts of this record, scattered across places. Decision logs, board minutes, data protection impact assessments and policy annexes all contain fragments of the reasoning. The Rationale Ledger does not replace them. It points to them, and it puts the one-page summary next to the workflow itself, which is where the next editor will be standing when the question arises.
It also connects to capability work. Under the 4D AI Fluency framework, Diligence is the competency of owning the outcome. Institutional memory is Diligence extended over time: owning the outcome not only today, but under the next owner. And the record of review, one of the artefacts in a well-built evidence file, is exactly what an auditor or funder will ask to see when someone wants to know why a control changed.
Leaders set the tone by asking for the reasoning, not just the result. In a review of an AI-supported workflow, the useful first question is not "does it work?" but "what would a new colleague need to know to change this safely?" If the answer is "ask the person who built it", the organisation is carrying a single point of failure, and that risk should be logged like any other.
Turnover is a fact of life in project-funded organisations, where contracts end when grants do. That makes the problem more acute in our sector than in most. Build the habit while the workflows are few. Retrofitting ledgers across forty undocumented workflows is a project. Writing the first five is a routine.
Treat the reasoning behind every AI workflow as an asset with an owner: one page, five entries, reviewed whenever the workflow changes, so that the system outlives the people who built it.
Pick the three AI-supported workflows that would cause harm if their owner left tomorrow, and write a one-page Rationale Ledger for each this month, starting with the Boundaries entry and its reasons.
No external statistics cited. Frameworks referenced: the 4D AI Fluency framework (Anthropic) and the M.A.I. Consulting evidence-file model. The Rationale Ledger is a framework proposed by M.A.I. Consulting; the opening scenario is illustrative.
An agent only its builder understands is a personal project running in production. A named owner, a written operating guide and a periodic test set are the minimum bar before calling it finished.
Custom Agents & Tools · 4 minAnalysis · 9 October 2026When AI helped shape a decision and a funder asks how it was reached, "we have the chat history" is not an answer. The Decision Ledger: four lines, five minutes, one named owner.
AI Use Policy · 6 minGuide · 27 September 2026The five documents that together let an organisation prove its AI governance to an auditor, board or funder: policy, tool inventory, training record, competency assessment and review record.
AI Use Policy · 4 minIf this is the question on your desk, a thirty-minute call tells you whether the service fits, or that you do not need us yet.