Product · Intelligence
An agent drafts it. A named reviewer decides it. The run keeps both.
MaxIron AI Engine runs customer-trained role agents in the MaxIron Portal across delivery, support, operations, reporting and governance. Higher-impact work stops at a review gate, and each run keeps the draft, the reviewer edit, the approver and the time.
Where agentic AI stalls
Assurance signs the write boundary, not the quality of the output.
Six weeks after a change, the questions are which agent proposed it, who approved it, and what was edited before it ran. A chat log answers none of the three.
The register
Five gates on one run, and the refusal at each
Written with your security and assurance teams before the first workflow type goes live, so it can be signed rather than discovered.
-
Gate 1 Remit
Signed by The stream owner who defined the role
Passes
One purpose, one stream, one environment class, describable to an auditor in a sentence.
Held back
Widening its own remit, or taking work from a stream it was not scoped to.
-
Gate 2 Gathering
Signed by The scope, agreed at scoping
Passes
Prior occurrences, the responses taken, and environment state inside that scope.
Held back
Reading outside the stream and environment class its remit names.
-
Gate 3 Drafting
Signed by Nobody yet. The widest permission, still short of the estate
Passes
A diagnosis, remediation sequence, report or runbook, written in full and held inert.
Held back
Filling a gap with an assumption. An open question stops the plan and routes to a person.
-
Gate 4 The crossing
Signed by The reviewer, whose edit is kept beside the original draft
Passes
The draft, put to a named reviewer who approves, edits or rejects it.
Held back
Retrying, escalating itself, widening scope or acting on probability while it waits.
-
Gate 5 Learning
Signed by The stream owner, per change
Passes
A knowledge change proposed with the evidence behind it.
Held back
Changing its own behaviour. An unapproved lesson changes nothing.
How wide the boundary is set
Three widths, matched to a workflow type before it runs once
The width is a scoping decision taken with you. No agent can move it.
The agent informs, a person acts
The agent answers and a person decides and acts. Nothing is written by the agent, on any system.
- Who acts
- The person
- Recorded
- Question and answer
- Used for
- Advisory work
Written, held, released by a name
The output exists and stays inert until a named reviewer approves it. The default for anything with consequence.
- Who acts
- Reviewer, then agent
- Recorded
- Draft, edit, approver, time
- Used for
- Diagnosis, remediation, reporting
A narrow boundary, agreed in advance
Execution runs inside a boundary set before go-live and stays reviewable against a named owner. It widens only on recorded evidence.
- Who acts
- The agent, inside the boundary
- Recorded
- What ran, where, at what cost
- Used for
- Repeatable, understood work
The crossing, once
One run, 02:14 to 07:31
-
02:14
A scheduled check reports the same non-production condition for the third time this month. Nobody is woken.
Written A case, against the stream the support role holds.
-
02:31
The support role gathers two prior occurrences, drafts a diagnosis and remediation sequence, and classifies it as requiring approval.
Written The inputs used, and the draft, inert.
-
07:20
The named reviewer corrects the step order for this environment and approves.
Written The edit, beside the original draft, under the reviewer name.
-
07:31
The corrected sequence runs. The original does not. The stream owner later accepts the correction as a knowledge change.
Written What ran, where and at what cost. Two approvals: one to touch the estate, one to change future behaviour.
Between 02:31 and 07:20 the run holds its position and does nothing. Five hours at a gate is the behaviour assurance is buying. Condition and timings are illustrative.
The record
What a run retains, and where it runs
Retained on every run
- Stages
- Inputs and outputs per stage, including the resume point where a run continues an earlier one.
- The draft
- The original output, kept beside the reviewer edit that replaced it.
- The decision
- Approver, time, and whether the draft was approved, edited or rejected.
- The cost
- Run cost, against the workflow type and the stream that owns it.
Scope and platform
- Where it runs
- The MaxIron Portal, beside IBM Maximo and MAS estates it does not modify.
- Role agents
- Delivery, support, operations, governance and guided interaction, each holding one remit.
- Data handling
- Agreed with your security team before deployment: what is processed, where it stays, retention, access.
- Learning
- Human approval per knowledge change. Nothing is learned silently.
Used in the MaxIron Portal for trained role agents, controlled execution, review gates and governed operational evidence, workflow layer of MaxIron Intelligence; the same layer MaxIron delivery, support and operations teams work behind.
Scope and boundaries
Three limits that will shape the rollout
Better heard at scoping than after sign-off.
It stops, and then it waits for you
Where no reviewer is available, nothing progresses. AI capability arrives as a standing review commitment somebody must be named for and given time, so estates that cannot staff reviewing should start with fewer workflow types.
It needs an agreed procedure to work from
Where no procedure exists and nobody holds authority to approve the outcome, more role agents produce more drafts. Three workflow types run properly beats twenty configured badly.
It improves at the speed you approve
Every change to what an agent knows waits for a stream owner. The rate of improvement is set by how often somebody works that queue, which is a recurring cost. See enterprise AI governance.
MaxIron AI Engine, frequently asked questions
- Can it act on production without a person?
- Not on higher-impact work: it waits at the crossing gate. Controlled automation runs inside a boundary agreed before go-live, with a named owner and a reviewable outcome.
- What does a reviewer or an auditor receive?
- Stages with inputs and outputs, the original draft, the reviewer edit, the approver and time, the run cost, and the lineage where a run resumes from an earlier one.
- How do you stop it learning the wrong lesson?
- Corrections and patterns are proposed with their evidence, and the stream owner accepts or rejects each one.
- How does it relate to MaxIron AI Crew?
- AI Engine is the Portal workflow layer for role agents. MaxIron AI Crew is what a user meets inside IBM MAS Manage. Both sit in MaxIron Intelligence and complement IBM Maximo and MAS.
Bring one piece of work somebody did twice last month.
We draw the boundary against that one task: the remit, the refusal at each gate, the crossing point, who holds each key, and what the record holds when somebody asks in November.
Bring this to the first call
- One routine operational task done more than twice last month
- The person who would sign it off, and whether they have the time
- The last question audit asked you about an automated change