The AI Asset Register: Keeping It True - Part four of four
Most AI registers are accurate on the day they ship and quietly wrong six months later. The gate most organisations are missing, the triggers that should force a reassessment, and who gets to challenge a classification once it is set.
Boundary, schema, classification, discovery. If you have read the previous three parts, you now know how to build a register that reflects the estate rather than a comfortable subset of it.
This part is about something the build phase does not naturally provide: an answer to what happens next. Most registers are accurate on the day they are signed off, and quietly wrong six months later. Nothing in the previous three articles prevents that on its own, because building a register and running one are different disciplines, and the second one needs its own design rather than a maintenance note at the bottom of a project plan.
The line worth carrying through this whole article: a classification with no reassessment trigger is a point in time assertion being presented as a standing one. Everything below is about closing that gap.
Two gates, and the second one is usually missing
Most organisations have an intake gate at project initiation. Someone proposes to build or adopt an AI system, a project gets opened, and somewhere in that process the system gets registered. This catches the first entry path from part three: built internally.
What is usually missing is a gate at procurement. This is the one that actually matters for the two paths that produce nothing on their own, the vendor feature that ships into a product you already own and the tool an employee finds and starts using.
A procurement gate cannot catch adoption by individual staff, that is what discovery in part three is for, but it can catch the vendor feature path if it asks the right question at the right moment: not just "does this purchase include AI" but "will this vendor's product develop AI capability during the life of the contract, and how will we find out." ISO 42001 A.2.3 requires AI specific requirements to be incorporated into existing organisational policies, procurement among them, which is the clause that makes this a compliance question rather than a nice to have. A.10.3 covers the supplier due diligence that should sit behind it, and A.6.2.5 covers the deployment authorisation and impact assessment sign off that should follow once something is approved.
A gate is only worth having if it produces three things: a register entry, a provisional classification, and a named owner. A gate that produces an approval and nothing else has moved the problem rather than solved it, because six months later nobody can say what was approved or why.
One constraint carries over directly from part three. If the gate is slower than the workaround, the gate manufactures shadow AI rather than preventing it. A procurement process that takes eight weeks to approve a browser extension will simply be routed around, and every time that happens the register gets a little less true. This is why part three argued for a triage tier rather than one process for everything: the gate has to be fast enough to win the comparison with not asking at all.
Telling a field update from a reassessment
Most of what happens to a register entry over its life is small. An owner leaves and gets replaced. A vendor changes its support contact. None of that touches classification.
Some changes do, and the register needs to force the distinction rather than leave it to whoever happens to be editing the record that day. Four triggers are worth naming explicitly, because each one changes something the classification in part two actually scored.
Scope expansion. The system starts serving a second business process without a new project being opened. The purpose dimension from part two was scored against the original use, and it no longer describes what the system does.
Repurposing. The output starts being used for a decision it was never built to support, often because it happened to be good enough for the adjacent case and nobody stopped to ask whether good enough for one decision means good enough for another. ISO 42001 A.9.4 speaks to this directly: use beyond the documented intended purpose has to be identified, assessed, and controlled, not simply noticed after the fact.
Population change. The people affected grow or change character. A pilot that ran against staff moves to customers. The impact dimension from part two, the one scoring population and severity and reversibility, was never assessed against this new group.
Vendor model update. The component underneath the entry gets replaced while the entry itself sits still. This is the deployment pattern problem from part one showing up again at maintenance time rather than at build time.
The EU AI Act gives one of these an external anchor, and it is worth using precisely rather than loosely. Article 3(23) defines substantial modification as a change to a system after it is placed on the market that was not foreseen in the original conformity assessment and that affects compliance or changes the intended purpose the system was assessed against. That is close to scope expansion and repurposing as defined above, though not identical, and the fit is worth checking against your own drafting rather than assumed.
One thing to flag honestly rather than gloss over. Article 111 sets out when a substantial modification pulls a system that predates the Act's obligations into full scope, and Article 111 was itself amended by the Digital Omnibus discussed in part one. At the time of writing, most secondary commentary on Article 111 has not caught up with that amendment, and the practical trigger date needs checking against the consolidated enacted text before you rely on it in anything client facing. The definition in Article 3(23) is stable. The transitional mechanics around it are not, right now.
ISO 42001 A.5.3 gives the cleaner and more durable anchor for this section: impact assessment results have to be updated when significant changes occur to the system or its operating context. Write the four triggers into the record as a named field with a date, so the absence of a trigger being fired is something your reporting can see, rather than something everyone assumes.
Reconciliation as a standing process, not a project
Part three's reconciliation loop, pulling from detection sources and comparing against the register, was framed there as a discovery exercise. Run once, it finds what you missed at build time. Run on a cadence, it is what keeps the register honest afterwards, and that second use is arguably the more important one.
Cadence should track the rate of change in your estate rather than the audit calendar. An organisation adding AI capability every month needs a different rhythm than one that touches its estate twice a year.
The metric worth watching is the delta rate: newly discovered systems as a share of the registered estate, per cycle. Read it carefully, because the intuitive reading is backwards. Zero deltas does not mean your estate is complete. In an organisation of any size, it almost always means detection has quietly stopped working. What you want to see is a delta rate that is stable or gradually falling, evidence that discovery is still functioning and that the gap it is finding is shrinking rather than a rate pinned at zero that could mean either total success or a broken pipeline.
A short set of numbers is worth reporting alongside it: the delta rate itself, the proportion of entries with a confirmed rather than merely provisional classification from part two, the proportion with a reassessment date already in the past, and the time from detection to registration. Framed for a board, these are not a count of how many AI systems the organisation has. They are a measure of whether the register can be relied on, which is a different and more useful question.
ISO 42001 Clause 9.1 requires the organisation to determine what needs monitoring and by what method, and Clauses 9.2 and 9.3 are where these numbers should land, in internal audit and in management review respectively, rather than sitting in a dashboard nobody with authority looks at.
Who can challenge a classification
The governance question the first three parts deferred. Short, but it is the one an audit reader has been waiting for since part two.
Name the default failure mode plainly: self classification by the team deploying the system, reviewed by nobody, with the incentive running the wrong way, because a lower tier means fewer controls and a faster launch. Nobody sets out to game their own classification. The incentive gradient does the work regardless of intent.
Independent challenge should sit with a second line governance function rather than with internal audit, and it is worth saying why rather than just asserting it. Internal audit reviewing an estate it will later be asked to audit is a conflict baked into the org chart, even where every individual involved acts in good faith. ISO 42001 A.3.2 requires roles and responsibilities for AI governance to be defined and documented, including accountability for individual systems separately from AIMS oversight, and A.3.3 requires a mechanism for raising concerns, which is what gives a challenge somewhere to go beyond an informal conversation.
What should a downgrade have to show? A change in the underlying facts, not a change in who is doing the assessing. This is exactly what the rationale field from part one exists to make testable: if the reasoning was written down properly at the time, a downgrade has to explain what changed since, not simply restate the same facts with a lower number attached. Article 6(4)'s documentation duty for the EU AI Act's own derogation, discussed in part two, is the clearest external example of a downgrade that has to stand on its own written reasoning rather than on authority.
One distinction resolves most of the ownership arguments that otherwise recur. Who owns an entry is the business owner of the use case, a role part one's schema already names. Who owns the register is a governance function, a different job entirely. Conflating the two is a common cause of registers going stale, because the person accountable for shipping the system is rarely the right person to be accountable for whether the register about it stays current.
What is realistic, in stages
A fortnight is enough to write the boundary rule and exclusion list from part one, fix the schema, and mine the passive sources from part three into a first draft register. This is decision work, not tooling work, and it does not require a platform.
A quarter is enough to get a provisional classification across the estate through inheritance as described in part two, confirm the top tier by hand, and get the procurement gate live. This depends on access to your existing data classification and asset registers, and on someone being named as the register's owner, not on new software.
A year is enough to have both intake gates operating, the four change triggers embedded in the schema, reconciliation running on a cadence, and the metrics from this article actually reaching management review. This depends on sustained ownership rather than a project budget, which is the harder thing to secure and the reason most registers that make it to the quarter mark still fail here.
A spreadsheet with a disciplined schema beats a platform with a bad one. Tooling is a scaling decision, not a starting one, and buying a platform before the schema is right just gives you a more expensive way to store the wrong fields.
That comes with one condition attached, and it closes the loop back to where this series started. Start small only if the schema is right, because the schema is the one thing in this entire series that cannot be retrofitted cheaply. Everything else described across these four articles is recoverable with effort. That one is not.
Closing
Every downstream obligation, whatever framework it comes from, attaches to the register. Get the unit of registration and the classification dimensions right, and mapping to the next framework that arrives becomes a lookup exercise rather than a project.
Five questions, one drawn from each part of this series, worth taking back to your own estate:
Can you produce your written exclusion list, with reasoning, right now?
Does one record in your register represent a use case or a model, and do you know which?
Could you reperform any given classification from its rationale field alone, without asking the person who wrote it?
When did you last reconcile the register against a source outside itself?
Who can overturn a classification, and what would they have to show to do it?
If the honest answer to more than one of those is "I'm not sure," that is not a failure. It is the place to start.