A research contract is a bet on an outcome that might not arrive. The model might not converge; the data might not carry the signal; the well count might be too small to generalise. A software licence or a build-to-spec statement of work assumes the opposite, that the deliverable is knowable in advance and the vendor is on the hook to produce it. Sign an AI research engagement on that template and you have quietly promised a result you cannot guarantee, and the operator has quietly bought a lawsuit for the day the result does not come.
We spent roughly twenty months building borehole-image interpretation models with a major operator in Oman, on carbonate reservoirs, alongside an academic partner. The technical story is elsewhere on this site. This piece is about the document underneath it, the draft service agreement and the proposal, because the way it allocated risk is what let the engagement run a failed experiment or two without either party reaching for a way to punish the other. Standard software procurement mishandles almost every clause below.
The obligation is effort, and the word matters
The first sentence of the risk architecture is that the provider carries an effort obligation, not a result obligation. The agreement says in plain terms that the provider is not responsible for achieving a specific result, because it is a research project. That single distinction reorders everything that follows. Under a result obligation, a model that fails to hit an accuracy target is a breach, and breach is the doorway to damages. Under an effort obligation, the provider owes competent, documented, good-faith work; a negative result honestly reported is delivery, not default.
This is not a way to escape accountability. Effort obligations still require you to show the work, and the phase-gated structure below makes that showing concrete. It is a way to price the engagement honestly. A vendor who has to insure against a result they cannot control will either refuse the work or load the fee with a risk premium the operator pays for nothing.
Payments are gated, and each gate is an exit
The fee did not arrive as a lump sum against a final deliverable. It was phase-gated, released stage by stage across the three phases of the programme, with an amount due at signing and further tranches tied to each phase. A gate is a checkpoint for the operator: if a phase does not convince, the next tranche is a decision, not an obligation already paid.
The same gates cut the other way for the provider. The agreement gave the provider a right to exit after Phase 1 or Phase 2 on 14 business days notice. Read the two together and you have a bilateral off-ramp. Either party can decide, at a phase boundary, that the bet is not worth continuing, and leave without litigating the whole contract. Money already earned for completed phases stays earned; money for phases not yet started is not owed. The failure gets contained to the phase it happened in.
The cap is the spine
The clause that lets both sides genuinely walk away is the liability cap. Total liability under the agreement is capped at USD 10,000, and all indirect and consequential damages are excluded. On a multi-year subsurface programme, USD 10,000 is a rounding error against the contract value. That is the point. The cap converts an unbounded, existential downside into a bounded, survivable one.
The instrument below makes the cap's job visible. Drag a hypothetical damages claim upward. One line is what a standard uncapped master services agreement would let that claim reach; it tracks the claim one-to-one, straight off the top of the chart. The other line is the provider's real exposure under this contract, which climbs with the claim only until it meets the cap, then stays flat forever. The vertical gap between the two lines is the ruin the cap absorbs.
Why does an operator agree to a cap this low? Because the alternative is worse for the operator too. An uncapped research vendor either prices in catastrophe or, if it is a small specialist firm, simply cannot sign, leaving the operator with only large incumbents and their overhead. A bounded cap is what makes a lean, expert counterparty contractable at all. The operator trades away a remote right to a large recovery, which it would likely never collect from a small firm anyway, for a real reduction in the price of expertise.
IP stays with the provider, and that is deliberate
The agreement has the provider retain the intellectual property in everything developed, unless agreed otherwise in writing, and has the customer indemnify the provider against third-party IP claims arising from client-supplied data. For an operator used to work-for-hire software contracts, provider-retained IP reads as a giveaway. It is not. A method that generalises beyond one field is worth more to everyone if the people who built it can keep improving it across engagements, and the operator gets a licence to use the result on its own wells regardless. The indemnity is the matching guard: the provider trained on the operator's proprietary logs, so the operator, not the provider, warrants that feeding those logs into a model infringes nobody's rights.
The data warranty that stopped being hypothetical
The customer warranted the correctness and completeness of the data it supplied, and agreed to bear the delay costs of late or deficient data. When the draft was written this looked like boilerplate. It did not stay that way. Partway through, a required orientation column was missing from a batch of interpreted picks, and without it the apparent-to-true correction could not run, which stalled supervised training on those wells. The warranty is what made that a client-owned delay rather than a provider failure. We have written up the mechanics of that specific stall and the clause that decided who paid for the wait separately; the point here is only that the data-quality warranty is the clause that turned a mid-project data wall into a defined, bounded event instead of a dispute.
Two more triggers sit alongside it. If the customer is more than 14 days late on payment, the provider may suspend all services. If either party wants out entirely, 60 days notice ends the agreement. Neither is exotic. Together with the phase-exit right and the cap, they form a complete set of orderly ways to stop.
Publication was negotiated before a line of code
One clause has nothing to do with failure and everything to do with why a research vendor takes the work at all. The proposal secured the right to publish at proposal stage: an approved joint case study, a video interview, progress blogs, joint conference talks, framed around a 96-hour client approval SLA on what went out. A research firm's currency is its published record, and negotiating publication after the results are in hands the operator a veto it can use as a bargaining chip. Fixing it up front, with an approval SLA rather than open-ended consent, meant the work could become papers and talks without a second negotiation. The peer-reviewed output that followed years later ran on rights agreed before the first well arrived.
What a standard MSA gets wrong
Line up the clauses and the pattern is clear. A standard software MSA assumes a knowable deliverable, so it writes a result obligation. It assumes the vendor should carry the downside, so it leaves liability uncapped or capped at the contract value. It assumes work-for-hire, so it takes the IP. It leaves the data warranty thin and treats publicity as a marketing afterthought negotiated late.
Every one of those defaults is wrong for research. The document that let our engagement survive its failures inverted all of them: effort not result, a hard cap, retained IP, an explicit data warranty, publication fixed up front, and exit rights at every phase gate. None of it is glamorous. All of it is what a procurement and legal team should reach for the next time the deliverable is a bet rather than a build.
Limitations
The terms described here come from a specific draft service agreement and its governing proposal for one engagement, anonymised. Contract law and enforceability vary by jurisdiction; a USD 10,000 cap, a 14-business-day exit, and provider-retained IP are choices that suited this research relationship and its parties, not universal recommendations. This is an account of how one contract allocated risk, not legal advice, and no clause here should be adopted without counsel reviewing it against your own facts. The claim-size sweep in the instrument is an illustrative range drawn only to show the capped-versus-uncapped divergence; it does not represent any real dispute, which did not occur.
