A service desk is usually described by its size. So many agents, so many tickets a month, so many euros a seat. None of those numbers tell you what the desk has to do, because none of them describe a path. A large European public-sector organisation preparing to re-tender its outsourced IT service desk started with a drawing of what already existed.
The drawing had three layers. Staff spread across several of the organisation's sites. Four ways to reach the desk: a call, an online chat, an email, or a local visit. One service desk behind all of them, running a single platform of record carrying five process areas: incident, problem, change, configuration and service level management. Several addresses, four doors, one tool, five disciplines. That is the entire environment, and it fits on one page.
The challenge
The count is the scope. Every site multiplied by every intake channel is the number of ways a request can arrive, and every one of those combinations is a path that has to be staffed, measured, priced and, when it fails, explained. Nothing in that arithmetic is exotic. It is the number nobody writes down, because a desk is bought as a service and drawn as a box.
Two features of the map do more work than the rest.
The first is that a local visit is a channel. It sits on the same line as call, chat and email, with the same status, so deskside support was not an informal courtesy but a route the organisation had committed to keeping open at every address its staff work from. A quarter of the named ways in is a person walking to a desk. Any claim about handling this demand without one has to survive that fact first.
The second is what sits inside the single platform. Four of the five process areas are the operational ones you would expect. The fifth is service level management, so the commitments made in the contract and the performance delivered against them were administered in the same system as the incidents themselves. That is unusual, and it is the quiet reason the study was possible on a short clock. Where service levels live in a signed document and achievement lives in a reporting tool, comparing the two is a reconciliation project. Here it was a query.
That distinction matters more now. An agentic system can only hold a service to its contract if the contract's numbers and the service's numbers are answerable from the same place. Most organisations still cannot do that, and no model quality fixes it: it is an architecture decision made years before anyone writes an agent loop.
The approach
The study was scoped as six objectives converging on one artefact, a detailed technical requirement. The six covered the existing contract and its agreements, requirements collected from business areas rather than from IT alone, contract provisions including terms, penalties and motivation strategies built on service levels, a supply market analysis, options for integrating additional support services, and a transition strategy. A requirement written that way is a synthesis product, the output of several evidence streams rather than the starting point of one.
The evidence was scheduled as work in its own right. Five classes had to be gathered before any current-state assessment could begin: contracts and associated agreements, financial data, performance reports, organisation structures and profiles, and processes. The phase that gathered them was the only one of seven to name no deliverable. Its key output column read as not applicable and its objective was to reduce project risk, an honesty most plans avoid, because a phase producing nothing a client can hold is the first one a client cuts.
The diagnosis method was two columns. Service levels and key performance indicators, agreed against achieved. That is the sharpest instrument in the engagement, and it carries no client data, which is why it can be described here. Most service reporting shows achievement without showing the commitment beside it, and the gap only becomes visible when a contract is up for renewal. Market work was scoped as secondary research, a speed decision inside seven phases across twenty-five working days, with a client review every fifth working day.
Then the sharpest thing in the material, which no single page states out loud. The final recommendations were to include options for integrating four adjacent services into the desk contract: mobile telephony, print infrastructure, senior management support, and optional local service desk assistance. That last one is the deskside visit. It appears on the as-is map as one of four live channels, and in the to-be scope as an option.
A re-tender is where a service people already rely on quietly becomes a line item that can be bought or dropped, and users find out afterwards. The test is simple enough to apply to any sourcing document: does the map of what exists and the list of what is optional use the same word for the same service? Here they did not.
The outcome
What this phase produced was a mobilisation: an agreed scope, a drawn as-is environment, an evidence checklist, a diagnosis method and a plan with a named output for six of its seven phases. It produced no measured result. No service level figures, no ticket volumes, no headcount and no cost appear in the source, and nothing here is a performance claim. The value is the method and the map.
Both travel. The map is what anyone deploying agentic systems into a service desk needs first and almost never has. The channel layer is where deflection lands, and it is not one number but a number per door, because a chat that resolves itself and an email that resolves itself are different systems with different evaluation sets. The process layer is where an agent's actions have to be recorded, using the same case types the humans use, or the record is not comparable to anything. The single platform of record is what makes both auditable at once.
Graded autonomy belongs on the paths, not on the desk. Some of those combinations can be closed end to end by an agent working from retrieval over the organisation's own knowledge base. Some should return a drafted answer for a human to release. Some, the ones that end with somebody walking to a floor, should never leave the copilot pattern. Deciding that per path is a design exercise. Deciding it per desk is a press release. Our Agentic OS work holds that distinction, because the alternative is one autonomy setting applied to every risk profile on the map.
The contract has to carry it too. Penalties and motivation strategies were named as separate levers on the same service levels, an admission that a purely punitive regime buys compliance rather than performance. That sharpens when a provider is deploying automation. Penalise per breached ticket and you have paid someone to suppress tickets. Measure resolved outcomes instead, per channel, and write down who owns the automation and the trained knowledge base when the contract ends, because that question is unanswerable at handover if it was not answered before signature.
The EU AI Act is in force, and the obligations that reach a service desk first are documentary rather than technical: knowing what the system does, telling a person when they are talking to one, and being able to produce the record. All three are easier when service levels and operations already share a system, which is what the map showed. The organisation was better positioned than it knew, for reasons that had nothing to do with foresight.
Draw the thing before you buy it. Four ways in, five disciplines, one system of record, repeated at every address staff work from. Every one of those is a promise to somebody, and a promise nobody has counted is a promise nobody has priced.