Most arguments about a support function happen after the evidence lands. Somebody presents a current-state pack, the numbers are worse than the room expected, and a decision follows. The decision that mattered most was usually taken months earlier, in a procurement document nobody in that room has read.
I keep the mandate slide of one such review because it is unusually honest about this. A large European public-sector organisation running an outsourced IT service desk across several sites commissioned a current-state analysis and a report of possible solutions. The work package that paid for it said on its face what the output was for: to support a possible retender of the IT service desk. The purpose was procurement, and understanding the desk that existed was a means to it.
What the work package actually bought
Set the four contracted items side by side and the centre of gravity is obvious. One of them examines the service that exists. A second reads the contract governing it, and was bought to produce recommendations for the next one. The remaining two look outward, at the supplier market, and forward, at sourcing and operating model options.
That is not a criticism of the brief. A retender needs a market study, a view on what the next contract should say, and a set of sourcing options, and none of those can be derived from measuring the incumbent. But it changes how the current-state numbers should be read by anyone who inherits them. Evidence commissioned to support a procurement gets collected in the shape a procurement can use. A metric that cannot be written into a tender document, scored against a bidder or defended at a challenge is a metric with no sponsor, and it tends not to get collected at all.
You can see the shape in the service level tables the same engagement produced. Every metric carries three columns: industry average, contractual target, measured performance. That is a procurement instrument rather than a diagnostic one. It answers what a bidder should be asked to commit to. On the incident acknowledgement table the contract had set no target at any priority level and nothing was measured. On the incident resolution table the targets existed and the measured column came back empty anyway. That emptiness is itself a finding, argued in The Desk Answered in 17 Seconds and Solved Fewer Than Half the Calls. The decision I want to draw out came earlier, and it is what made those tables the natural artefact.
The scope decision arrived as an instruction
The line on the mandate slide that does the most work is the one describing the future service. A single first point of contact for roughly 8,000 users, covering every site the organisation runs and the affiliated partner organisations alongside them.
Read that as a scoping statement for a study and it looks like background. Read it as what it is, a commitment about the shape of a service, and the assessment that follows has a fixed destination. Consolidation was settled. The reach across institutional boundaries was settled. What remained open was how to get there and what to pay for it.
I do not think that was wrong. Whether one front door should serve every site is a strategy question, and strategy questions are not resolved by counting calls. The failure mode is subtler: by the time the evidence is presented, months later, in a pack with tables and charts, the decision that was made up front reads like a conclusion the analysis produced. Nobody in the room remembers it was an input, so nobody asks what evidence would have overturned it. The honest version writes the decision down as a decision, with its reasoning attached, inside the brief. Then the evidence phase knows exactly which boundary it is working inside and which it is allowed to push on.
A brief that names the future user base is not scoping a study. It is scoping a service, and every number gathered afterwards is gathered inside a boundary somebody has already drawn.
The one thing the brief refused to prescribe
Against a fixed destination, the method was left genuinely open, and the choices made there are the reason the pack held up.
Three activity streams were run. A current-state assessment built on stakeholder interviews and an assessment framework designed for this organisation rather than reused from elsewhere. A market study drawing on three independent sources at once, peer client organisations, the supplier community and an internal research function. Then synthesis into recommendations and options.
Two details in that are worth copying. The first is that the framework was tailored. A generic instrument produces findings phrased in the instrument's vocabulary, which is precisely what a challenge from a losing bidder attacks. The second is that the current-state work deliberately reached past the desk in scope, into other service functions inside the same organisation that had nothing to do with IT. What that widening was worth is argued in How to Benchmark an Outsourced Service Desk Against Your Own Internal Desks, and I will not repeat the case here. The point for this piece is narrower: the brief did not ask for it. Method left open is method a team can use well.
- 4
- Contracted workstreams, one looking at the existing desk
- 3
- Independent sources behind the market evidence
- 8,000
- Users the future desk was to serve as first point of contact
The same document exists in every AI programme now
The 2015 work package has a direct descendant, and most data and IT managers reading this have one open in a tab. It is the business case for an AI-assisted or agentic service desk, and it behaves identically.
It names the population before anything is measured. All employees, all sites, all subsidiaries, usually all channels. It contracts four or five workstreams, of which one is a baseline assessment and the rest are tooling selection, integration design, operating model and change. It states an ambition in the same register as "works from the start". And it treats the reach of the future service as the premise rather than the output, which means no evidence gathered later can shrink it without the programme looking like a retreat.
Four things follow, and none of them requires a model.
Write the scope decision as a decision. One paragraph in the brief saying what has been settled, by whom and on what grounds, and what remains genuinely open. It costs nothing and it prevents an input from being read back as a finding.
Buy a method built for you. In 2015 that meant an assessment framework designed around this organisation. In 2026 it means an evaluation set drawn from your own tickets, your own applications and your own escalation boundaries. A vendor benchmark on a public dataset tells you a model is capable. It does not tell you the model can close your incidents, and that is the only question your business case is actually asking.
Triangulate the market evidence. The market study behind this review ran across three independent sources rather than a single supplier questionnaire. An AI business case resting on vendor-published figures has exactly the weakness that design avoids, and it will be found the first time the programme is asked to justify a renewal.
Protect the workstream that looks backwards. If one of four workstreams examines the service you already run, it is the one under pressure when the timeline slips, and it is the only one that can tell you afterwards whether anything improved.
None of this is hindsight about a failed programme. The record I am reading is the first deliverable of a staged engagement: a current-state picture, a market study, recommendations and sourcing options, produced so that a retender could follow. Nothing was built, and the target service described on that mandate slide was a target. What survives the decade is the structure of the document, which is the part everybody copies without noticing they have copied it.
Drawn from the mandate and method slide of a review of an outsourced IT service desk at a large European public-sector organisation, work I led. That engagement was the first deliverable of a staged piece of work, producing a current-state picture, a market study and sourcing options ahead of a possible retender; no new service was delivered, and the single point of contact serving roughly 8,000 users described above was a stated target, not an achieved one. The four contracted workstreams, the three activity streams, the triangulated market evidence and the tailored assessment framework are all on the source slide. Reading the commissioning document as the place the scope was actually decided, and reading an AI business case the same way, is ours.
“A brief that names the future user base is not scoping a study. It is scoping a service, and every number gathered afterwards is gathered inside a boundary somebody has already drawn.”
Get in touch
Put RealAI’s applied-AI team on your hardest data problem.
We help enterprises move from pilots to production: sovereign models, governed data, and agents you can audit. Start with a value-first assessment.