Clear, source-led guidance for accountable business AI.
Build a compact AI-use inventory, rate exposure across four explainable dimensions, and route every use to proportionate review as its context changes.
ModelFold Editorial Team9 min read
An AI inventory should be a routing layer for governance work, not a spreadsheet that attempts to hold every assessment. Record each AI-enabled use in its real workflow, give it accountable owners, rate its inherent exposure and link deeper evidence only when the route requires it. A customer-support drafting aid, for example, becomes a materially different use when it gains access to identity documents, permission to issue refunds and authority to send messages. The supplier may remain unchanged, but the record and review cannot.
What teams should retain
Create one record for one AI-enabled use in a defined workflow, purpose, user group and output-to-action path. This unit captures the context that determines exposure: who operates the system, who is affected, which information enters it, what it produces and whether that output informs or executes an action. Official frameworks examine AI through people, context, data, models, tasks and outputs; a model name or supplier contract alone cannot describe those relationships.
Split the record when purpose, affected parties, input sensitivity, action authority, deployment exposure or accountable ownership changes materially. The same hosted model might therefore support separate records for internal drafting, customer-service recommendations and automated account updates. Link the shared supplier, model, application, data source, contract, assessment, control, incident and approval records instead of copying them. This keeps common evidence consistent while preserving a searchable history for each use.
Use a compact discovery record that identifies ownership, context, exposure and the next governance route. Detailed impact assessments, validation reports, supplier reviews, test results, control evidence, incidents, approvals and monitoring plans should remain conditional linked artefacts. The base record needs enough information to find a use, understand its boundaries, assign responsibility and decide what further work is required; it should not invite an owner to write an assurance report during initial intake.
A listed control is not evidence that the control works. Summarise the design in the row, then link the test, operating record or reviewer finding that supports it. Use controlled lifecycle values and dates so teams can find stale or retired records. Review timing should remain risk-based: a rapidly changing, higher-exposure use may need closer attention than a stable, bounded use, but one universal interval will rarely fit the whole portfolio.
Rate consequence, autonomy, scale and sensitivity separately, using three concrete levels and one use-specific evidence sentence for each rating. This four-dimension rubric is an internal editorial proposal synthesised from broader characteristics in NIST, OECD, Canadian and UK materials; none of those authorities prescribes or endorses this exact method. Mark an unknown for resolution instead of guessing, especially where missing information could change the route.
Write evidence such as “the pilot serves one trained support team and one message class”, not “scale is low”. Rate the actual proposed or operating use rather than the supplier’s generic product. The dimensions are related, but they answer different questions: a small deployment can still handle highly sensitive information, while a low-sensitivity public assistant can create broad exposure through scale or autonomous action.
Set the provisional tier from the highest material dimension: all Level 1 ratings produce Tier 1, any Level 2 with no Level 3 produces Tier 2, and any Level 3 produces Tier 3. Do not average a severe exposure away. This rule and the three routes are editorial design choices, not externally validated scoring formulae; each organisation must calibrate them against its risk tolerance, policies and applicable obligations.
Escalate where an override identifies out-of-tolerance activity, severe affected-party impact, sensitive data with broad access, ineffective human control, difficult reversibility, cascading dependencies, a material incident, applicable obligations or unresolved facts. Rate inherent exposure before giving credit for controls. Reviewers should then examine control design and evidence, record the residual-risk decision and specify conditions. Merely listing human review, access control or monitoring must not lower the inherent rating.
| Internal tier and route | Minimum review | Decision and evidence | Monitoring and reopening |
|---|---|---|---|
| Tier 1 — Registered | Owner confirms the record, approved boundaries, standard controls and operating instructions. | Approval follows the organisation’s standard policy path, supported by rationale and control attestation. | Risk-based owner attestation and reopening on material change. |
| Tier 2 — Assessed | Cross-functional review covers affected parties, data, oversight, suppliers, dependencies, tests, fallback and correction. | A named owner records conditions, control evidence, residual risk and the monitoring plan after required specialist review. | Defined measures, issue routes, evidence refresh and event-driven reassessment. |
| Tier 3 — Enhanced Review | Independent challenge and domain expertise test necessity, alternatives, authority, reversibility, controls, incidents and exit. | The authority named in policy explicitly approves, constrains or stops the use, with findings and residual-risk evidence. | Closer exposure-based monitoring and immediate reopening after incidents, control failure or material scope change. |
Keep applicable legal, regulatory, contractual, privacy, security, procurement, labour, records and sector classifications in a parallel track owned by qualified teams. Tier 1, Tier 2 and Tier 3 are internal routing labels only. They cannot establish that a use is lawful, exempt, prohibited or legally high-risk, and they should never be mapped automatically to a jurisdiction’s categories.
The tier changes when the use’s data, permissions, autonomy and reach change, even if the model and supplier remain the same. Consider a fictional pilot that retrieves approved help content and drafts a response for a trained support employee, who edits and sends it. Account decisions, policy exceptions, unsupervised sending, payment data, identity documents, credits and account changes remain outside scope. This illustration applies the proposed method; it reports no measured outcome and does not validate the rubric.
The highest rating makes the pilot provisionally Tier 2. Required pre-send review, source links, blocked write actions, role-based access, sampling, a complaint path and manual fallback belong in the control record; their presence does not erase inherent exposure. If the proposal later adds identity documents, payment-dispute details, refunds, account updates, automatic sending and cross-region deployment, every dimension becomes Level 3. Reopen the record before expansion, route it to Tier 3 and run the separate obligations review.
The inventory earns trust when a change in data, authority or scale changes the route—not merely the row.
Maintain the inventory through event-driven reopening, supported by owner attestation at a risk-based interval. Feed discovery from product and workflow intake, procurement and renewals, application and architecture review, access and integration administration, approved-tool and subscription processes, model and data governance, employee disclosure, support cases, incidents and complaints. These feeds widen visibility; they do not prove that every AI-enabled use has been found.
Use a small dashboard to show records and affected parties by tier and lifecycle, missing fields, overdue decisions, open exceptions, incidents, control gaps, routing time and retirement closure. Coverage measures are discovery indicators, not certificates of completeness. Higher exposure, rapid change, recent incidents or weak evidence can justify closer review, while a stable use may support a longer interval. Do not impose one quarterly or annual cadence across every record.
Treat retirement as a controlled state rather than deleting the row. Preserve the accountable owner, dependency and migration decisions, access-removal confirmation and whatever evidence applicable policy requires to show that the use has stopped. The method remains an internal routing proposal, not a compliance determination. Qualified legal, compliance, privacy, security, procurement, records, labour and sector owners should assess obligations, while appropriate domain expertise and accountable human review should precede consequential or high-stakes uses.
Include a stable use ID, business and technical owners, purpose and boundaries, users, affected parties, inputs, outputs, action authority, suppliers, dependencies, controls, lifecycle, dates, ratings and obligations status. Keep assessments, tests, approvals and control evidence as linked artefacts when the route requires them.
Define explainable dimensions, anchor each with three levels and require an evidence sentence for every rating. This proposal uses consequence, autonomy, scale and sensitivity, assigns the provisional tier from the highest material rating, and adds explicit escalation overrides. Calibrate it to organisational risk tolerance and obligations.
Make the primary governance record one AI-enabled use in its workflow and action context. Link shared model, supplier, application, data and assurance records so common evidence is maintained once without hiding materially different uses.
Reopen a record whenever a material change affects its purpose, people, data, authority, technology, scale, controls, incidents or obligations. Add owner attestation at a risk-based interval; there is no universal quarterly or annual cadence suitable for every use.
No. An internal tier routes organisational review and cannot determine a legal or regulatory classification. Qualified legal, compliance, privacy, security, labour, records, contractual and sector owners must assess the requirements that apply to the specific use.
This article was researched using the following sources:
We report on how AI actually lands inside a business. Our work starts from named sources, separates what we found from what we think, and uses AI assistance for research and drafting under documented editorial controls. We are not a substitute for individual expert review.
A practical guide to assigning AI decision rights, choosing an operating structure, and turning pilot evidence into durable portfolio learning.
A practical guide to binding prompts, models, tools and workflow logic into one testable AI release, with staged promotion and complete rollback.
A practical method to bound an AI assistant’s data, tools, permissions, actions, approvals, refusals, evidence and operating controls.
