Enterprise AI · Assurance
AI assurance reference for IBM Maximo and MAS estates
Three numbered registers, built to be cited. Q1 to Q7 are the questions an auditor, risk function or public-sector buyer asks. A1 to A4 are the artefacts that answer them. The credential table states what each certificate does not cover.
Scope
Who this reference is written for, and where it stops
What this reference applies to
- Written for
- Internal audit, risk, information security, engineering authority, procurement and public-sector assurance teams.
- Covers
- AI-assisted work on IBM Maximo and MAS estates that MaxIron delivers or operates, in assist, draft or controlled automation mode.
- Anchored by
- MaxIron AI Engine execution controls, which hold the modes, the gates and the run record.
What it does not
- Out of scope
- AI tooling your organisation runs independently of MaxIron.
- Not claimed
- That any model output is correct. We publish no accuracy figure.
- Citation
- Codes Q1 to Q7 and A1 to A4 are stable once published and safe to quote in a submission.
Most AI capability in asset management is stopped by assurance rather than by engineering: the capability is proven, assurance is consulted late, and the questions that follow are reasonable ones nobody prepared evidence for.
Register 1
Assurance questions, Q1 to Q7
Asked roughly in this order. None of them is about model architecture. All of them are about who decided and what exists to show it.
| Ref | Question | Usually asked by | The answer | Artefact |
|---|---|---|---|---|
| Q1 | What context can the model see? | Information security, data protection | A written scope per stream of work: which environments, record types and document sources are approved as input, and what is out of bounds. The answer is a list rather than an assurance that it behaves sensibly. | A1 |
| Q2 | Which decisions can it make on its own? | Risk, operations leadership | A mode per workflow. Assist for advisory work, draft where a named person approves before publish or execution, controlled automation only inside a boundary agreed in advance. Agents, actions and workflows are enabled independently. | A1 |
| Q3 | Who approved this outcome, and when? | Internal audit | The run record: stages passed, the draft as produced, the reviewer edits, the named approver, the timestamps, and the lineage where one run resumed from another. | A2 |
| Q4 | What happens when the model is uncertain? | Risk, engineering authority | Uncertain output does not progress. It routes to a review gate or escalates to a named owner. Compliance-critical calculations are deterministic, so they never reach a generative stage. | A1, A2 |
| Q5 | How do you stop it learning the wrong lesson? | Quality, engineering authority | Nothing is learned automatically. A correction made at a gate is raised as a proposal with its evidence, and a person with authority over that role approves or rejects it. | A3 |
| Q6 | What are you certified to, and what does it cover? | Procurement, public-sector buyer | ISO/IEC 27001:2022 and ISO 9001:2015, certificate 27274 by ISOQAR, UKAS-accredited certification body 0026. UK G-Cloud 14 supplier 721548. IBM Partner Plus Gold. Coverage and non-coverage sit in register 3. | A4 |
| Q7 | Can we turn it off? | Risk, procurement | Your own people can disable an individual agent, action or workflow, and reduce any workflow to advisory mode without a change request to us. Approved outcomes stay in your Maximo estate. | A1 |
Cite a row as Q1 to Q7. Codes are stable once published and are never renumbered.
The gates
Three gates, and what is refused at each
Every AI-assisted run on a MaxIron-managed estate passes these in order. The refused column is the one an engineering authority reads first.
-
Gate 1 Scope gate
Signed by The stream owner, at scoping
Passes
A run inside the purpose, environments and record types written into A1 for that stream of work.
Held back
A run that would widen its own remit, escalate its permissions or retry against a broader target. It stops and says so.
-
Gate 2 Review gate
Signed by The named approver for that decision type
Passes
A draft a named reviewer has read, edited where needed and approved. The corrected version runs, and the difference from the original stays visible.
Held back
Higher-impact work with nobody available to review it. The run queues rather than proceeding on its own judgement, and a five-hour wait is correct behaviour.
-
Gate 3 Learning gate
Signed by The owner of the role being changed
Passes
A reviewer correction promoted into what an agent knows next time, with the evidence behind it and a dated approval.
Held back
A correction changing future behaviour on its own. Permission to act and permission to learn are separate approvals.
Register 2
Evidence artefacts, A1 to A4
Each artefact is designed to be read by people who were not present, which is the only test that matters.
| Ref | Artefact | What it contains | Where it is produced |
|---|---|---|---|
| A1 | Scope and mode register | Per stream of work: purpose, approved context, prohibited context, operating mode, named owner, escalation threshold. Versioned, so a widening of scope appears as a dated decision rather than as drift. | Agreed at scoping, held in the MaxIron Portal, changed only by the named owner. |
| A2 | Run record | Per run: stages passed, input and output at each stage, the draft as produced, reviewer edits, named approver, timestamps, cost, and lineage where one run resumed from another. | Captured as the work runs, by AI Engine execution controls. |
| A3 | Learning approval log | Every change to what an agent knows: who proposed it, the evidence behind it, who approved it, and when. This is the artefact an engineering authority asks for second. | Raised at the learning gate, approved separately from the work itself. |
| A4 | Credential set | ISO/IEC 27001:2022 and ISO 9001:2015 certificate 27274 by ISOQAR, UKAS-accredited certification body 0026. UK G-Cloud 14 supplier 721548. IBM Partner Plus Gold. | Independently checkable rather than asserted. Set out in register 3. |
Illustrative. Internal audit samples an automated remediation that ran four months earlier: A2 returns the scheduled check that opened the case, the two steps the reviewer corrected before approval, her name, and the timestamps either side of a five-hour wait at the review gate. A3 shows the correction was approved separately rather than absorbed into future behaviour.
Verifiable, not asserted
Register 3: four credentials, and what each does not cover
Each is checkable without contacting us. The fourth column exists because an assurance audience notices what a supplier avoids saying.
| Credential | Issuer and reference | Covers | Does not cover |
|---|---|---|---|
| ISO/IEC 27001:2022 | ISOQAR, UKAS-accredited certification body 0026 Certificate 27274 Verify → | Independently audited controls around the systems handling your data: risk assessment, access control, supplier management, incident handling. | Model behaviour. It is an information security management system certification, and we will not present it as evidence of model governance. |
| ISO 9001:2015 | ISOQAR, same certificate Certificate 27274 Verify → | Quality management: documented, audited and repeatable delivery processes. | Whether a particular AI-assisted decision on your estate was appropriate. |
| UK G-Cloud 14 | Digital Marketplace supplier listing Supplier 721548 Verify → | A public-sector procurement route your commercial team can verify without contacting us. | AI governance. A framework listing is a route to buy, and it ranks nobody. |
| IBM Partner Plus Gold | Verifiable through IBM partner listings IBM Gold Partner Verify → | IBM partner standing for IBM Maximo and MAS delivery. | Any statement about model behaviour or AI decision control. |
Register 4
Situations, and the control that catches them
Four situations we would rather demonstrate in your environment than describe. T1 to T4 can be written into an acceptance schedule as they stand.
- T1
An agent is asked to act outside the scope written into A1
- Passes when
- The run halts, records the refusal against the scope entry it breached, and does not retry against a broader target.
- Witnessed by
- Your stream owner and an information security reviewer, in your own non-production environment
- T2
Higher-impact work reaches the review gate with nobody available to approve it
- Passes when
- The run queues and stays queued. Nothing reaches the estate on the model’s own judgement.
- Witnessed by
- The operations lead who owns that decision type
- T3
A reviewer correction would change how the agent behaves next time
- Passes when
- The correction applies to this run, and appears in A3 as a dated proposal awaiting the role owner rather than taking effect quietly.
- Witnessed by
- Your engineering authority
- T4
An assurance sample lands months after the work ran
- Passes when
- A2 returns the draft, the reviewer edits, the named approver and the timestamps for a run your auditor selects, not one we choose.
- Witnessed by
- Internal audit
Scope and boundaries
Where this governance model stops
Stated in the same spirit as the fourth column above.
No AI-specific certification, and no implication of one
MaxIron holds ISO/IEC 27001:2022 and ISO 9001:2015 under certificate 27274, plus a UK G-Cloud 14 listing as supplier 721548. None of them covers model governance. Where an AI-specific standard becomes relevant to your assurance position, we state where we stand against it.
No warranty that an output is correct
We publish no accuracy figure and we will not warrant a model output. What is governed is the decision that follows it: a named person with the authority reviewed it, the review is recorded, and nothing with consequence reached the estate without them.
No substitute for your own accountability structure
Where nobody has decided who may approve what, this model has nowhere to attach. Naming decision owners, agreeing escalation thresholds and funding the review time are organisational commitments, and how we run it day to day sets out what they cost.
AI governance, frequently asked questions
- Do you hold an AI-specific certification?
- No, and we will not imply otherwise. Certificate 27274 by ISOQAR covers information security management and quality management. What we put in front of an assurance function for AI governance is the operating model on this page and the evidence it produces.
- How is auditability handled?
- Each gated step captures review and decision evidence as the work happens, because assurance questions arrive months later. A2 answers who approved what and when. A3 answers whether a correction was allowed to change future behaviour.
- Can this satisfy public-sector assurance expectations?
- The model is built around controlled access, human approval, traceability and procurement-verifiable markers including UK G-Cloud 14 supplier 721548. Whether it satisfies a particular authority depends on the requirements that authority sets, which is why we would rather see your questionnaire early than send a generic answer late.
- Can we cite this page in an assurance submission?
- Yes, and the codes exist for that. Cite a question as Q1 to Q7 and an artefact as A1 to A4. If your authority needs the answers against its own wording on letterhead, ask and we will produce them.
- How does this differ from the AI Engine product page?
- AI Engine is the mechanism at procedure level: stages, modes, gates and what a run record contains. This page is the reference: the question, the artefact that answers it, and the exact boundary of each credential.
Bring the AI assurance questionnaire.
Send the questions your risk, audit or procurement function will ask about AI, including the ones nobody can currently answer. We map each to a row in register 1, name the artefact that answers it, and mark where the answer is the operating model rather than a certificate.
Bring this to the first call
- Your AI or automation assurance questionnaire, in whatever state it is in
- The last question internal audit asked about an automated change
- Your current position on who may approve a change that touches production
- Any public-sector or regulatory assurance requirement you are held to