A supplier that misses its targets is a manageable problem. A supplier that hits them and leaves you where you started is not, because at that point there is nothing left to escalate.
A large European public-sector organisation asked us to review the outsourced service desk carrying first-line IT support for its staff, ahead of a decision about whether to renegotiate, bring the function in house, or take it back to market. The performance evidence was not ambiguous. First call resolution was running at 47%. The contract required more than 65%.
The finding that writes itself is to hold the supplier to the number. The finding that mattered sat under it. That floor of more than 65% had been agreed once, at signature, and left alone for the life of the agreement. By the time we read it, the recommendation we ended up writing was not to enforce the target but to replace it, alongside new measures the contract had never carried at all. Which is an odd thing to write about an agreement still in force, and is the reason this piece exists. Full compliance would have delivered a desk nobody in the organisation was asking for.
The challenge
Three of the root causes recorded in the assessment were not failures of delivery. They were absences in the document. No minimum skill set was defined in the contract for an agent to be placed on the production floor. Reporting requirements were not clearly defined in the contract. Meetings and reviews were not clearly defined in the contract. Each one is a sentence somebody chose not to write, and each one removed a lever the buyer needed years later.
The reporting silence is the expensive one. The service management tool in use could not generate service-level reports. Text inside tickets was not searchable, and the automatic path for updating the knowledge base did not work. So the number that governed the agreement could not be produced from the system that recorded the work it described. A floor you cannot compute from your own data is not a floor. It is a sentence both parties are free to interpret in the review meeting.
The skills silence produced most of the issue log. Incidents were categorised wrongly and routed wrongly. Tickets moved back and forth between the desk and the second-line groups. Escalations arrived carrying information too thin for anyone to act on. Duplicate incidents accumulated, including the case where a user ringing to ask about an open ticket had the call logged as a new incident rather than as a diary entry against the existing one. Critical incidents were detected late and escalated slowly. Second and third line reported spending their time on routine desktop and application issues the desk could have closed. The organisation was paying for that work twice, once under the contract and once inside its own retained teams, and the second payment appeared in no service report anywhere.
The desk and the retained groups also had no shared knowledge management system, and turnover ran higher than expected on both sides, so every departure took knowledge out of the building. None of that is visible in a resolution percentage, which is why a contract built on a small set of easily reported numbers survives so long without anyone noticing what it never asked for.
The approach
The organisation had written design principles for the desk it wanted next: one desk as the single point of contact for all issues, higher first call resolution, accountability set at the outset rather than argued later, service that works from the start, value for money. Those became the evaluation criteria the sourcing options were later scored against.
The remedy was then written as contract language rather than as an improvement plan. Revise the service level and performance targets to current industry standards instead of defending the ones already signed. Add measures the agreement had never carried: acknowledgement and resolution times split by incident severity, customer satisfaction and call quality, agent utilisation and agents as a share of total desk headcount, new-recruit training hours, annual training hours, agent job satisfaction. Tier the service levels by user group so that price, reward and penalty follow business impact. Avoid exclusions and thresholds to service level achievement, since every exclusion is a way to report green while a user waits. Put minimum qualifications and business-context training into the contract, with training outcomes measured by the buyer through formal assessment and remedial work at the supplier's cost. Move to outcome-based pricing, rewarded above target where the buyer actually gains and penalised below it. Name the desk as owner of the ticket to closure, and write operational level agreements between the desk and the second and third line so the handover has terms.
Half of that list is about measurement rather than about service, and that split is deliberate. A service level nobody can compute independently is a negotiating position, not a commitment.
All of this transfers directly to the service agreements being signed this year for AI-assisted support. The shape is the same. A copilot over an internal knowledge base, retrieval-augmented generation across a vector store built from the documentation that desks like this one never maintained, a deflection target, an accuracy floor, a term measured in years. Those floors are being set once, in a single negotiation, with the same energy that produced a resolution target above 65% and then outlived everyone who agreed to it.
Three things carry across. First, ask how long the floor in front of you is meant to hold, because a number agreed once at signature gets quoted back at you for the rest of the term, and this desk is what a floor nobody reopened looks like from the inside. Second, insist the number be producible from your own data: a held-out evaluation set that you own, refreshed on your own traffic, run on a cadence written into the agreement, with enough lineage recorded that a disputed answer can be traced back to the document it was drawn from. Third, put the skill floor in the contract, because the failure that hollowed out this desk, staff put on the floor with no defined minimum, has a counterpart on the supplier side in data readiness and MLOps discipline. Ask what gets retrained, on what, how often, and who checks. The European AI Act makes documentation and logging far easier to argue for in a first contract than to retrofit into an amendment later.
The same review recommended automating password resets through the phone system, and the most frequent request types outright. That instinct still holds, and now extends to early, carefully scoped autonomous handling of a narrow slice of standard requests. The clause naming who measures whether it worked is the part that gets skipped.
The outcome
This engagement produced a review, not a result. What was delivered was a current-state assessment across people, process, technology and governance, the organisation's design principles turned into evaluation criteria, four sourcing options scored against those criteria, a recommended option, and a recommendation to stand up a project to write the new service description, service levels, transition terms and financials before anything went to market. An exit and transition plan was recommended in parallel, useful whoever ended up holding the contract. No service level moved while we were there. The 47% is a number we found, not a number we changed.
The one sentence worth carrying out of it is that the organisation should design what it wants from the start, because the moment before signature is the last one in which anyone can still change the terms cheaply. Performance against a contract is reviewed monthly for years. The contract itself is reviewed once, by people under time pressure, and then treated as physics. Benchmark the floor, not just the distance you sit below it. Our Consult engagements now open by reading the agreement before the dashboard, on service desks and AI services alike, because a dashboard can only tell you how you are doing against a number somebody already agreed to accept.