Imagine a service director responsible for reducing a backlog. The AI project belongs to IT. Finance has already included savings in next year's plan. Operations still owns the old approval process, and nobody has agreed who can change it.
Every function can provide an update. The backlog remains.
An enterprise AI operating model defines how the organisation makes and executes decisions about AI. It covers authority, funding, shared capabilities and the responsibilities that remain with each business unit. Its value becomes visible when a decision crosses departmental boundaries.
The first question I would put to that programme is simple. Who has permission to change the work?
Give the business owner a usable mandate
A named sponsor is helpful. A sponsor who cannot alter a service process has a limited range of useful actions.
For the hypothetical service operation, define a mandate that includes the target service outcome, the scope of process change and the resources available. Make any excluded decisions explicit. Access to customer information, contractual commitments and workforce changes may require additional authorities.
Then agree which decisions stay with the workflow owner and which need escalation.
| Decision | Proposed accountable authority | Required contribution |
|---|---|---|
| Change queue routing inside approved policy | Service director | Operations and technical delivery |
| Recognise a financial benefit | Finance controller | Evidence from the business owner |
| Expand access to restricted data | Designated data and security authorities | Purpose, access design and risk review |
| Increase production scope or funding | Executive sponsor within investment limits | Operating results, cost and control evidence |
| Suspend unsafe operation | Pre-authorised duty role | Incident record and escalation to relevant owners |
Build a usable mandate
Select a field to see an example.
Set the service result and its quality condition
Reduce the service backlog without increasing incorrect resolutions.
Agree how the team will measure both backlog and quality.
Specify these six fields, then test the mandate against a real exception.
This is an example to adapt to the organisation's actual delegation rules. It is not a universal organisation chart.
Make deputies and escalation routes explicit. A role without available people becomes a waiting room.
Put the mandate on one page before the next delivery meeting. Include the following fields.
- Outcome. State the service measure and the quality condition that must hold.
- Process boundary. Name the queues, systems and decisions the owner may change.
- Resources. Record the approved budget, delivery capacity and access constraints.
- Reserved decisions. Specify which data, contractual and workforce changes need another authority.
- Containment. Name the duty role, its deputy and the actions they can suspend.
- Review. Set the next decision date and the evidence needed to continue.
Walk a recent exception through that page. If nobody can authorise the required action, revise the mandate while the gap is still cheap to resolve.
Decide what the centre should own
An enterprise needs shared controls. It also needs people close enough to the work to recognise a bad result.
A central team can maintain approved integration patterns, security requirements, evaluation infrastructure and procurement guidance. Business units can own process outcomes, operational acceptance and the benefit case. Finance should apply a consistent basis for recognising returns across the portfolio.
Specify the boundary when a shared service causes a local problem. If the central platform changes a model and a business unit's exception rate rises, the unit needs a route to restrict use or roll back within agreed controls. The central team needs evidence that it can investigate.
A 2026 Google Cloud and National Research Group survey of more than 2,400 executives associated stronger reported returns with clear ownership, workflow integration and ongoing capability development. This is vendor-commissioned, self-reported research. It supports examining those practices, not assuming that a new reporting line causes higher returns. Google Cloud research.
Keep the test local. Can the service director get a harmful change contained before the next steering committee?
Review a real exception before reviewing the presentation
Choose a recent case that the existing process handled badly. Follow it from arrival to resolution with the people who did the work.
Ask where it waited, which information was missing and who had to make a judgement. Check the source systems. A process map assembled entirely from management interviews can omit the workaround that keeps the service running.
Our article on management use of ChatGPT discusses this process-level investigation. The ownership question comes immediately after it. Who can remove the wait, change the rule or fund the missing integration?
Training should prepare managers to make those decisions. They need enough technical understanding to challenge claims about data access, tool permissions and evaluation. They also need to understand what the system cannot establish.
An external adviser can run that examination with the internal team, bring specialist knowledge and help develop the internal capability. Keep the business decision and its accountability inside the organisation.
Make portfolio reviews capable of ending a project
An AI portfolio can accumulate pilots because continuing each one is easier than explaining why it should stop.
Give every funded initiative a next decision, a date and the evidence required. That decision might approve a limited rollout, request more evidence or end the work. Establish these possibilities before the team becomes attached to the demonstration.
For a project with promising speed gains and rising rework, the relevant decision may be to restrict the scope to simpler cases. For a project that reduces a genuine operational risk, finance may accept a different justification from direct cost reduction. Make that rationale visible and reviewable.
Make the next decision
Select an outcome to follow its route.
Approve a defined production scope
Record the scope, accountable owner and operating limits. Keep service quality, cost and risk evidence available for the next review.
A review has three legitimate outcomes: roll out within approved limits, request further evidence or end the work.
Use a consistent enterprise AI ROI account for financial claims. Keep service quality and risk evidence alongside it, with their own units. A portfolio score assembled from incomparable numbers can conceal the decision instead of informing it.
Do not infer workforce savings from a model demonstration. Establish sustained capacity under real demand and exception conditions, then make any staffing decision through the appropriate management and HR process. Include training, redeployment and resilience in that discussion.
Write the authority down before expanding it
The operating model gives people authority. The AI governance framework translates the approved scope into controls and evidence. Both need to survive changes in staff, suppliers and technology.
Before the next budget meeting, take one active initiative and write down the person who can change the process, the authority that validates the benefit and the role that can suspend operation. If those people disagree about the mandate, resolve it before increasing spend.
Amalgama is an enterprise AI consultancy led by Kirill Kiryushin, whose experience spans consulting, enterprise technology and business ownership. That combination matters when the difficult part is getting a technical capability to work inside an organisation with existing responsibilities and commercial constraints.
Bring the initiative and its unresolved decision to a workflow assessment. The useful next step is to establish who can act and what evidence they need.