A research contract is two documents wearing one cover. The first is the scope: the phases, the deliverables, the science that will be attempted. The second is the schedule of milestone gates, and it is the one almost nobody reads and the one that actually decides whether the programme happens. Scope tells you what a vendor promises to try. The gate schedule tells you when each commitment is made, what has to be true before the next stage is released, and who is carrying the risk in the gap between a commitment and the result it was meant to buy.
We are writing as the people who ran the technical side of a multi-phase applied-AI engagement with a mid-sized Middle East carbonate operator and sat close to how it was structured. Other whitepapers in this series cover how a single research phase was planned and staffed, and the year-two services agreement that handed the capability over. This paper sits between and beneath both. It is about the founding contract for the whole twenty months, and specifically about its architecture: how the programme was cut into phases and milestone gates over time, how sign-off at each gate was bounded, and what kinds of work the programme was made of. For a programme director, a procurement lead, or an investor deciding whether an applied-AI programme is a fundable shape, this is the anatomy that matters. We do not publish the contract's commercial values, and none of the argument depends on them.
The headline total is the most inert number in the contract
Every contract has a headline total, and stated alone it tells you almost nothing useful. It is a size, not a structure. A buyer cannot tell from the total whether the programme is front-loaded or back-loaded, whether the vendor is exposed or the client is, or whether a bad first phase can be stopped before it drains the programme. All of that lives in how the programme is cut, and it is cut two ways at once: along time, into phases and milestone gates, and along category, into the kinds of work it resources.
The temptation with any headline contract value is to reason about it as a lump. That instinct is exactly backwards. The total is the one quantity in the contract that is deliberately inert. It does not move, it does not gate anything, and it carries no information about who is at risk when. Everything that makes the contract a risk-management instrument rather than a wager is in the two decompositions, and the rest of this paper is about reading them.
Along time: five gates, and the one that leads the work
The programme was scheduled as five milestone gates across three phases. Phase 1 carried two: a mobilisation milestone at kickoff, and a review gate around month 5. Phase 2 carried one, a gate around month 10. Phase 3 carried the last two, a gate around month 15 and a close-out gate around month 20. Five gates, the first of them at kickoff, each tied to a milestone rather than to a calendar tick.
The word that does the work in that schedule is mobilisation. The first milestone was not an advance against future science. It was earmarked, explicitly, to clean and prepare the data infrastructure and the cloud before any model work began. That distinction is the whole argument of the structure. In an applied-AI programme on real subsurface data, the work that decides whether the science is even possible is not the modelling. It is the unglamorous plumbing: pulling image logs out of the operator's systems, reconciling tool outputs whose value ranges do not agree, standing up the compute, and building the ingestion path that turns raw logs into training data. That work has to be resourced before it can be done, and it has to be done before anyone can honestly promise a model metric. A mobilisation milestone resources the data plumbing that everything else stands on.
Read the staircase above and the shape of the programme is legible in one glance. The first step lands at kickoff, before any model has run, and every later step is a gate at which the client signs off before the next stretch of work is released. This is the opposite of a milestone-only contract where the vendor carries the entire first phase on its own balance sheet and is settled only on delivery. Here the buyer accepts more early exposure in exchange for something concrete: a vendor that is resourced to do the data preparation properly rather than one cutting corners on plumbing to protect its own position. The mobilisation milestone is a choice about who carries the setup risk, and the operator chose to carry it, deliberately, because the alternative is a first phase built on a rushed foundation.
There is a second, quieter piece of structure in the schedule: a phase-0 runway. Before phase 1 formally starts, the contract allows a setup and governance window, the period in which the mobilisation work is actually done. It does not appear as a gate of its own, but it appears as time, and it matters because it is the interval where the programme's fate is largely decided. If the data does not arrive, if the imaging tools defeat normalisation, if the compute is not standing, the whole twenty months is compromised regardless of how good the eventual model is. The runway is the contract acknowledging that the first risk to retire is not a modelling risk at all.
Why five gates turn one bet into five cancellable ones
Splitting a programme into five stages is not, by itself, de-risking. Anyone can chop a plan into five. What makes the structure a risk-management instrument is that each stage is gated: the next one is released against a milestone, and the client signs off before the work moves on. That converts one large, open-ended research bet into a sequence of smaller, cancellable ones.
The mechanism deserves to be stated plainly, because it is easy to wave at and hard to design. At each gate, the vendor submits a deliverable, the client reviews it, and the next stage is contingent on acceptance. If phase 1 fails to convince, phase 2 does not automatically proceed. The buyer's downside on a programme that goes wrong is not the whole contract; it is the stages already released plus the one in flight. For a research engagement, where the honest expected outcome is a distribution and not a point, this is the difference between a fundable structure and an unfundable one. A board can approve a bet it can stop. It struggles to approve a bet it cannot.
Milestone gates are common enough that the interesting question is not whether to have them but what they cost in time. A gate is a stop point, and a stop point where the next stage is contingent on sign-off puts the client's review latency directly onto the critical path. The vendor can finish a deliverable on a Friday, but if approval takes three weeks, three weeks of schedule are gone, and on a fixed twenty-month clock that latency is expensive. Left uncapped, the client's own slowness silently lengthens the programme the vendor is being judged on.
Along the SLA: capping the latency the buyer controls
The contract's answer to that is a 96-hour client approval SLA. Every milestone sign-off is contracted to happen within 96 hours. Read as a legal clause, it is unremarkable boilerplate. Read as a schedule instrument, it is doing real work: it caps the schedule cost of the one input the vendor cannot control, the buyer's own review speed, and it does so at a known, plannable four days per gate.
The comparison above is the point. Under the 96-hour cap, review across all five gates consumes a fixed twenty days, a number a project plan can absorb without flinching. Without the cap, that same review is an open-ended hole, and the bar fans out as latency drifts. The gap between the two bars is not a nicety. It is schedule risk that the SLA retires, and it is retired symmetrically: the SLA protects the vendor from a client who sits on deliverables, and it protects the client by forcing the vendor to package each milestone into something reviewable in four days rather than a sprawling artefact that cannot honestly be judged that fast. A 96-hour SLA is a two-sided discipline. It says the vendor will produce gate-shaped deliverables and the client will act on them promptly, and it puts a bound on the cost of the handshake between them.
It is worth being precise about what the SLA does not do. It does not guarantee the deliverable is good, and it does not guarantee acceptance. A client can review inside 96 hours and reject. What the SLA guarantees is that the decision is fast, which is the part that sits on the critical path. The quality of the deliverable is governed elsewhere, by the scope and the acceptance criteria; the SLA governs only the latency of the verdict. Conflating the two is a common misreading of approval clauses, and it leads buyers to over-value an SLA as a quality lever when it is really a schedule lever.
Along category: reading the split of work for where the risk really sits
The gate schedule tells you when commitments are made. The split of the contract by category tells you what it resources, and it is where the character of the engagement is most exposed. We group the proposal's cost lines into three buckets by what they de-risk: people, meaning the research staff, scholarships, overhead and the research partnership through which the staff were engaged; compute, meaning cloud services, additional big-data cloud and capital equipment; and enabling work, meaning travel and the remainder.
Read by what they de-risk, the shape is unambiguous. People come to more than two thirds of the contract. All compute together, cloud and capital equipment, comes to under a third. Enabling work is a small remainder. This is a people contract, not a compute contract, and the split says so in the plainest possible terms.
The instrument above lays the three buckets on a single bar, and the counter-intuitive fact is the one worth sitting with. The category most buyers assume dominates an applied-AI programme, compute, is the minority here. The scarce, risk-bearing input is human: the research staff who build and validate the models, and the partnership through which they are engaged. Compute is real and it is not trivial, but it is the smaller share.
Two consequences follow from that split, and both are decision-relevant for a buyer. First, the people side of the programme is doing double duty. It is not only resourcing the work; it is funding local capability, the training of local researchers and students through a university partnership, an in-country capability dividend that sits inside the people work rather than in a separate programme. The investment that de-risks the science is the same investment that builds durable local skill, which is a large part of why an operator structures an engagement this way at all.
Second, the compute share tells the buyer something about scalability that the headline total cannot. If most of a contract is people and the model, once trained, serves every subsequent log at a low marginal compute cost, then the heavy part of the programme is the one-time build, not the ongoing run. Other pieces in this series make that operating-model argument in detail, and this split is its upstream evidence: the resources went where the one-time risk was, into people, and only a minority went into the compute that the eventual served model amortises across every scan. We treat that downstream story as a pointer rather than re-derive it; the point here is only that the split predicts it.
The formula the whole structure is protecting against
There is a way to make the logic of the schedule exact, because a research programme has a governing inequality and the gate structure exists to keep the buyer on the right side of it. What a buyer should be willing to commit at any gate is bounded by the expected value of what the programme delivers, weighted by the probability that it delivers at all, and set against what is still to be committed. Written compactly, the decision at each gate is whether the forward expectation still clears the forward commitment:
The milestone structure is what makes this inequality evaluable rather than notional. Every gate updates the left-hand side: passing phase 1 raises the estimated probability of success and sharpens the estimate of value, and the buyer re-checks whether the remaining stages on the right are still worth committing to. A lump-sum contract has no such update; the buyer commits the whole right-hand side once, at the start, when the probability estimate is at its most uncertain. The gated schedule lets the buyer commit in step with what has been learned, which is the entire point of de-risking a research bet by phase. The mobilisation milestone is the one deliberate exception, a commitment made before the first update, and it is defensible precisely because the thing it resources, the data plumbing, is a prerequisite for generating any update at all.
What this structure is not, and where it can fail
It would be dishonest to present this as a template that de-risks by itself. The structure has real failure modes, and naming them is part of reading it properly.
The mobilisation milestone front-loads the buyer's exposure. A meaningful share of the contract is committed before a single review gate is passed. If the vendor is the wrong vendor, that commitment is the least protected in the whole schedule, because it is spent before the first gate can catch anything. The mobilisation milestone trades the vendor's setup risk for the buyer's early-exposure risk, and that is only a good trade if the buyer has done the diligence to be confident in the vendor before signing. It is a structure that rewards a well-chosen partner and punishes a poorly chosen one more than a milestone-only contract would.
The gates are only as good as the acceptance criteria behind them. A milestone gate with vague criteria is theatre: the client cannot honestly reject, so the next stage is released regardless, and the cancellability that justified the structure evaporates. The gate de-risks only if failing it is a real possibility, which means the deliverable for each phase has to be specified sharply enough that a 96-hour review can reach a genuine verdict. A soft gate is worse than no gate, because it looks like protection and is not.
The SLA cuts both ways on quality. Forcing a 96-hour decision is a schedule discipline, but it can pressure a client into rubber-stamping a deliverable they have not truly absorbed, especially a technical one. The cap protects the schedule at some risk to the depth of review, and a buyer who leans on the SLA to move fast can find they have approved something they did not understand. The mitigation is not to weaken the SLA but to invest in making deliverables reviewable, which is effort that lands on the vendor and belongs in the scope, not the schedule.
And the people-heavy split concentrates key-person risk. When most of a contract is people, the programme's fate rides on a small number of researchers. That is efficient and it builds local capability, but it means the loss of a key person is a larger shock to a contract like this than to a compute-heavy one. The split that makes the engagement a good capability-building instrument also makes it fragile to attrition, which is one more reason the year-two agreement that follows is written to transfer the capability out of individuals and into the client's own team.
Reading a research contract as a buyer
Pull the pieces together and the engagement's architecture is a coherent argument, not a collection of clauses. The mobilisation milestone resources the data plumbing that decides whether the science is possible. The milestone gates turn one open-ended bet into five cancellable ones and let the buyer commit in step with what each phase reveals. The 96-hour SLA caps the schedule cost of the buyer's own review latency at a plannable four days a gate. The phase-0 runway names the setup risk as the first to retire. And the split of work, more than two thirds people and under a third compute, says where the value and the risk actually sit and, not incidentally, funds the local capability the operator wanted to build.
For a buyer looking at a proposal for applied AI, the decision criteria fall out of that architecture directly:
- Is there a mobilisation milestone, and is it earmarked for the data and infrastructure work that gates whether the science is even possible, rather than being a generic advance?
- Is each stage milestone-gated, with acceptance criteria sharp enough that a gate can genuinely be failed, so the structure is cancellable and not theatre?
- Is there a cap on the buyer's own approval latency, so the client's review speed does not silently lengthen the schedule the vendor is judged on?
- Does the split between people and compute match the risk profile of the work: people-heavy where the risk is the one-time build, compute-light where a served model amortises the run?
- Does the people work do double duty, resourcing delivery and durable local capability at once, or is capability-building bolted on separately?
This engagement was structured to answer yes to all five, and the structure is transferable even though the specifics are not. The point is not that every applied-AI contract should copy this programme's gate months or its mobilisation milestone. It is that the gate schedule and the split of work are the real risk-management instrument of a research engagement, and reading them as such, rather than reading only the scope and trusting the total, is how a buyer tells a fundable programme from a hopeful one.
Limitations
This is a reconstruction from the founding engagement proposal and the surrounding archive, anonymised, and it carries the limits of that provenance. The three-phase structure, the five milestone gates and their planned months, the mobilisation milestone and what it was earmarked for, the 96-hour approval SLA, the phase-0 runway and the people, compute and enabling split are all sourced from the proposal. The grouping of the proposal's cost lines into those three buckets is an editorial lens we laid over the sourced lines, and we report the split only as rounded shares; the contract value, the payment values and the individual cost lines are deliberately not published. The month positions of the gates are the proposal's planned milestones, not a record of when each was actually signed off. The proposal also predates the programme's real start, which slipped later than planned and ran a different technical core than the one first scoped, so the schedule described is the intended architecture rather than a retrospective audit of what happened. The uncapped-latency baseline in the SLA instrument is an illustrative comparison, not a contracted number. None of the arguments here should be read as disclosing the client, the field, or the commercial terms of any specific named counterparty; they are offered as a transferable reading of how applied-AI research can be phased, gated and de-risked, grounded in one real contract.
References
Proposal V3.0, founding engagement proposal, March 2021. The source document for the phase structure, the five milestone gates, the mobilisation milestone, the 96-hour approval SLA, the phase-0 runway, and the cost lines behind the people, compute and enabling split. Held in the project archive; anonymised here.




