Count the doors before you size the desk.
A large European public-sector organisation ran a single outsourced IT service desk for staff working across several of its sites. The kick-off document for its re-tender opens with that arrangement drawn in three layers: the people, the ways in, and the desk itself. Staff in more than one place. Four intake channels. One system of record carrying five process areas. None of it is a finding alone. The counts are the finding, because each multiplies the others, and every combination is a path somebody must staff, script and answer for.
The challenge
The four ways in were a telephone call, an online chat, an email, and a local visit. Three arrive down a wire. The fourth is a person walking over, and its presence is what makes the map honest: a desk that counts walk-up support as a first-class channel admits that part of its demand never entered a queue, and so cannot be routed, held or deflected the way the other three can.
Behind the channels sat one system of record running five process areas: incident, problem, change, configuration, and service level management. The last reads like housekeeping and is not. The commitments the contract made, and the measurements against them, lived in the same system as the operational record. Most estates split them: service levels in a signed document, achievement in a reporting tool, and putting the two side by side is a project rather than a query.
Now hold the counts against each other. Every site has to be reachable through each of the four channels, or the four is really three in some places and nobody has said so out loud. Every route lands in the same five process areas, so the same categories, priorities and escalation rules have to hold whether a request came by chat from one office or from a corridor in another. A channel is not a feature. It is a commitment repeated at every site for the length of the contract, and the tender document is where that repetition gets priced or gets discovered afterwards.
The re-tender then did something the as-is map gives no warning about. Four adjacent services were to be assessed for folding into the new contract: mobile telephony, print infrastructure, senior management support, and optional local service desk assistance. That last one is the walk-up channel. It appears in the current-state drawing as one of four live ways in, and in the future-state scope as an option that could be bought or dropped. Neither page says so. Read them together and you watch a service people already relied on becoming a line item. Re-tenders are where existing services quietly turn into options, and the defence is a map drawn first.
The approach
Six objectives fed one artefact: analyse the existing contract with its agreements, service levels and performance indicators; collect requirements from the business areas rather than from IT alone; design provisions drawn from what the supplier market offers; run a supply market analysis; find options for integrating further support services; recommend a way of moving the service to a new provider with least risk. All six converged on a single detailed technical requirement, which makes the specification an output of the evidence rather than its starting point.
That evidence came from a phase which named no deliverable, and the plan says so by leaving the output column empty. Five classes of material had to be in hand before current-state work began: contracts and their associated agreements, financial data, performance reports, organisation structures and profiles, and processes. Four of the five are unstructured document sets.
The diagnosis was then reduced to one comparison across two axes: service levels and performance indicators, agreed against achieved. Three words carry the method: you put the commitment and the delivery in adjacent columns and let the difference nominate the problems. That fitted twenty-five working days for a reason that is architectural rather than methodological: the commitments and the achievement data were already in one system. Where service levels sit in a contract and achievement sits somewhere else, the same task is a reconciliation exercise that does not fit a five-week plan.
The same property decides something larger now. Put an agent loop on the three wired channels and it will classify, retrieve, draft and resolve faster than any shift roster; give it graded autonomy and it works under a written rule about what it may close alone and what it hands back. Neither is worth much if the agent cannot see, as it acts, whether the action breaches a commitment somebody signed. An agentic system holds a service to its contract only when the contract's numbers and the service's numbers can be read in one query.
The five process areas are the other half of that requirement. An agent that resolves an incident, raises a change or touches a configuration record has to leave its work where a person would have left it, same fields, same trail. With the EU AI Act in force, that trail is the evidence that a system acting on staff requests did what it was specified to do and stopped where it was told. Our Platform work begins at that join, because a copilot wired to a channel and writing nowhere is a demonstration rather than a service.
The outcome
What this record contains is an agreement and a plan. Six items were put up for agreement at the kick-off before any analysis started, including the inputs the client itself had to supply. The plan ran to seven phases inside a five-week window, held to account by a progress review every seventh day, and fifteen written assumptions carried the whole risk treatment on an agenda that had promised assumptions and risks. Market work was deliberately scoped as secondary research across four topics: outsourcing models, contract terms, penalties and motivation strategies. Requirements were to come from at most twenty-five business users over video, and the plan assumed no more than three vendors would respond.
None of that is a result. What the plan committed to produce was a tender-ready requirement plus recommendations across three areas: contract terms with penalties and motivation strategies built on the same service levels, the scope-integration options, and the transition to a new provider. No service level figure appears here, because the record we hold carries the method and not the values.
Two things would be scoped differently today. The evidence sweep is the first. Four of the five classes are document sets, and retrieval-augmented reading over a contract, its annexes and a year of performance reports returns a cited answer in an afternoon. That does not shorten the study by the same margin: the phase with no deliverable was never about reading speed, but about reaching the current-state conversation with nothing left to look up. The sample is the same story: a cap of twenty-five users over video, standing in for staff spread across every site, was a defensible calendar decision and is no longer forced.
The second is the tender document, which now has to state what the provider may automate, which evaluation sets that automation is measured against, and who owns the agents, the prompts and the accumulated knowledge when the contract ends. The brief already asked for penalties and for motivation strategies built on the same service levels, an admission that a punitive regime buys compliance rather than performance. That matters more once a supplier is deploying automation. Penalise a provider per breached ticket and you have paid them to suppress tickets. Measure them on resolved outcomes across all four channels and the incentive points where you want it.
The count at the front of that kick-off document is not an inventory. It is the scope. Four ways in, every site they have to reach, one system of record, five process areas, and every combination is a path somebody answers for, whether by a person, by an agent, or by nobody. Sizing the desk before counting the paths is how an organisation buys a service that works properly in one office.