A kick-off document is the least glamorous artefact an engagement produces, and it settles more of what follows than the analysis that comes after. It is written before any evidence has been read. Everything in it is a commitment made in advance: what will be gathered, from whom, and what the client must hand over for the schedule to hold.
This one belonged to a sourcing study for a large European public-sector organisation preparing to re-tender the outsourced IT service desk its staff depended on. The plan set out seven phases across twenty-five working-day columns, about five working weeks, with a client checkpoint every seven days. Nothing in it reports a finding. It is a plan, and what makes it worth rebuilding is the arithmetic it accepted.
The challenge
Six objectives had to converge on one document: an analysis of the existing contract and its accompanying agreements, service levels and key performance indicators; input from the business areas on what they would need next; provisions grounded in the supplier market, covering terms and conditions, penalties and, in the brief's own phrase, motivation strategies using SLAs; a supply market analysis; options for folding additional support services into the contract; and a strategy for moving the service to a new provider with minimal risk.
Those six were to feed one artefact, the detailed technical requirement that would become the tender: the requirement is an output of the study, not its input. Most re-tenders begin with the previous specification and edit it, which is how an organisation buys the same service twice.
Twenty-five working days is a hard boundary, and harder than it looks, because much of the window was already spoken for. Four progress reviews sat at seven-day intervals, and the final presentation occupied the last working-day column on the chart, with no buffer behind it. The assumptions added one more: one draft per deliverable, with one week for the client to comment. A fifth of the window was reserved in advance for the client's reading, not the team's writing.
The approach
| Phase | What it was for | What it was to produce |
|---|---|---|
| 1 Kick-off | A common understanding of the deliverable and its content | The kick-off document |
| 2 Planning and data gathering | Plan later phases, gather the inputs that reduce risk | Nothing. The plan names no output |
| 3 Current state assessment | Understand where performance was failing | Recorded service levels, indicators, contractual terms |
| 4 Future state requirements | Understand user needs for a future desk | Business user requirements |
| 5 Market evaluation | Understand the supply market well enough to weigh the options | A supply market report |
| 6 Reporting | Present findings, recommend a future desk strategy | First full version of the final report |
| 7 Review and sign-off | Test the deliverable against the objectives | The final report, taken to sign-off |
Three rows repay a stop.
Phase 2 carries no output, and that is deliberate. Gathering was scheduled as work in its own right rather than buried inside the analysis that consumes it. One workstream gathers five classes of evidence: contracts and associated agreements, financial data, performance reports, org structures and profiles, and processes. The other plans the workshops: participants, dates and logistics, and the questionnaires people would answer. A phase with no deliverable is the first thing a compressed schedule deletes, and deleting it is how a study reaches its fourth week without the contract in hand.
Phase 3 carries two tasks, and the second is the whole diagnosis: service levels and indicators, agreed against achieved. Agreed in one column, achieved in the next. Most reporting on an outsourced desk shows achievement without the commitment printed beside it, which makes every month read like weather rather than a contract. No performance figures from this engagement appear here; the method is the publishable part.
Phase 5 made the most honest trade in the plan. Market evaluation was scoped explicitly as secondary research on published outsourcing models, contract terms, penalty regimes and motivation strategies, rather than a primary supplier survey, which will not fit twenty-five days at a defensible cost. The logic is plain: buy the market's published structure, spend the scarce days on the questions only this client's own evidence can settle.
Look at the shape of that. Phases 2, 3 and 5 are document work: evidence in, structured summary out, with the workshop logistics the one exception inside phase 2. Phase 4, the workshop, and phases 6 and 7, the recommendation and sign-off, are judgement and negotiation, running at the speed of people's calendars.
That split is where a plan document stops being history and becomes something we sell. The Platform work we do starts at the dull end: a client's contract set, performance packs and process documents in one indexed corpus, with lineage back to the page a claim came from. A copilot pointed at documents nobody has gathered is a demonstration, not a capability.
The outcome
What this document set out to reach was agreement. It put six things on the table at kick-off: objectives, the client's expectation of the end deliverable, the inputs required to produce it, risks and assumptions, governance and roles, and the review schedule. It carried fifteen written assumptions and no risk register, so the assumptions were the risk treatment: where the team would sit, what the client would supply, who led which deliverable, and what converted into a change request. It specified a weekly report on a seven-day period carrying progress by aspect, every risk and issue paired with an agreed action, effort per person by location, and the budget remaining after that week, countersigned by both sides.
That is the whole of it. No finding, no recommendation, no result. Those were still five working weeks away when this was written, and this piece does not claim them.
The reading phases are the ones current tooling compresses. A contract set, financial reports, performance packs and process documents are the workload retrieval-augmented generation was built for: index the corpus into a vector store, ask what the agreed service levels say, and get passages back with the document and page attached. The secondary market scan has the same shape, run over public material. None of that decides anything. It removes the days between having the documents and having them summarised, which here was most of three phases.
Two conditions decide whether that is real. The first is data readiness, which the assumptions had already named without the vocabulary: the plan required existing service reports, their owners and their documents to be accessible, preferably by self-service. An organisation that cannot hand over its own contract set in week one cannot index it either. The second is evaluation. Extraction over a contract is worth something only once an evaluation set says how often it reads a service level correctly, and lineage back to the clause is what makes a wrong answer findable. That is ordinary MLOps discipline pointed at documents.
So the arithmetic. Three of seven phases are compressible in principle and four are not. The workshop still needs its users. The recommendation still needs someone accountable for it. The client still needs their week with the draft, a week written into the assumptions before anyone knew what the draft would say. Sign-off is still a meeting. A study planned at five working weeks is bounded from below by the calendar rather than by the reading, which points closer to two weeks than five. We have not run that end to end on a study of this size, so treat it as arithmetic, not as a result.
Public procurement adds one line. With political agreement reached on the European AI Act, a public body using automated assistance inside a tender will want to say what that assistance did, on what evidence, and who signed the conclusion. The kick-off document already had the instinct: it recorded who produced each output and who countersigned it. Our Consult engagements start there, keeping the early and carefully scoped autonomous experiments inside the phases where a wrong answer is caught by someone who was going to read the passage anyway.
Cutting a study from five weeks to two does not make it a better study. It makes the judgement phases a larger share of it, and the judgement phases were always the ones the client was paying for.