For the general counsel
A regulatory duty your organisation holds does not transfer to a supplier, and any arrangement implying it does is describing something your regulator does not recognise. And we do not offer an uncapped indemnity — cover a supplier could not fund is a promise rather than a remedy, and you would discover which at the worst possible moment.
The agreement arrives with a deadline and a redline from the other side that has been through their own counsel twice. Somewhere in the business, a colleague has already told an executive committee the date it goes live. You are the last person who can say the exposure is unacceptable, and saying it late is expensive for everybody including you.
The novel part is that the standard playbook does not fully cover it. Warranty language written for hosted software assumes deterministic behaviour, and the clauses you would normally rely on to allocate fault presuppose that fault can be located. When an automated step contributes to a wrong outcome, the question of who caused it does not resolve cleanly into the categories the precedent library was built around.
Underneath that sit the duties that stay yours regardless. Notification obligations run on statutory clocks that begin whether or not a supplier has answered the telephone. Retention schedules are set by statute and by the schedule your own records policy commits you to. A preservation obligation attaches to material the moment litigation is reasonably anticipated, including material sitting inside somebody else’s infrastructure that you now have to be able to reach.
Then there is what happens after signature, which is the part the negotiation almost never reaches. A subpoena arrives naming records held by a counterparty. A regulator asks how an outcome was arrived at and expects a legible answer rather than an assurance. Discovery requires a preservation hold applied somewhere you do not administer. Each of those is a Tuesday afternoon problem whose difficulty was decided months earlier, by a clause nobody argued about because it read as boilerplate.
The clauses that determine your actual position after something goes wrong are the ones least argued over during drafting, because they read as standard and their cost only becomes visible once they are needed.
Start with indemnity, because it is where the gap between comfort and remedy is widest. A supplier offering uncapped cover for artificial-intelligence outcomes is offering something no supplier of that size could fund, and an uncollectable indemnity is worse than a capped one — it stops the conversation about controls, which is the conversation that would actually have reduced your exposure. What is on offer here is a stated cap with insurance behind it, disclosed before drafting rather than discovered during it.
Then the duty question, which is settled rather than negotiable. Where your organisation holds a regulatory obligation, operating a function through a supplier does not move it. What a well-designed arrangement can do is make discharging it evidenceable: every action leaving a receipt that names the actor and the authority it relied on, so that when a regulator asks how something happened there is a record rather than a reconstruction.
On decisions about people, the boundary is drawn deliberately wider than most agreements draw it. Determinations of eligibility, employment, credit, housing or benefit are excluded — and so is merely ordering a queue of people, because ordering decides while appearing not to, which is precisely the shape a fair-lending or fair-housing examination is built to find. That exclusion is architectural rather than a policy commitment, and it is worth confirming as such during diligence rather than accepting the sentence.
And the clause we will not sign, stated before you draft it: deletion of a specific record from an already-written backup image. Our architecture cannot perform it — an image is written once and restored whole. A supplier who signs that clause is either mistaken about their own systems or does not intend to comply, and you would rather establish which now. What is offered instead is deletion from live systems on a defined timeline plus expiry of the images themselves on a stated schedule.
Whether the indemnity is collectable rather than comforting — measured by the stated cap set against the insurance position and the contracting entity behind it.
Whether a duty was mistakenly treated as transferred — measured by each obligation mapped to the party that holds it as a matter of law, in writing.
How long a preservation hold takes to apply — measured by elapsed time from instruction to hold in force, and whether your team can do it unaided.
Whether an outcome can be reconstructed for a regulator — measured by whether a receipt names the actor and the authority relied on, retrievable without a support request.
How quickly a notification clock can start — measured by the committed window from our awareness to your awareness, written as a contract term.
Which clauses will be refused — measured by the refusal list obtained before drafting rather than discovered in the third redline.
What survives termination — measured by return format, timeline, and which obligations continue past the end date.
that this reduces your legal exposure. A new counterparty is new exposure. What is claimed is that the positions determining that exposure are disclosed before drafting rather than discovered during it, which is a different and smaller thing.
Retention follows the schedule you set rather than a default we chose. That distinction matters because a supplier default that is shorter than your statutory schedule destroys material you were obliged to keep, and one that is longer keeps material you were obliged to dispose of. Both are findings.
A legal hold is applied by your own people and overrides deletion for the material it covers, including material that would otherwise have expired on schedule. Your team should confirm during diligence that they can apply one without asking us and without a support ticket, because a hold that depends on a counterparty’s response time is a hold with somebody else’s clock in it.
For discovery, governed state and the receipts attached to it are designed to export in documented forms, so material can be produced in a shape a reviewer can work with rather than as a set of screenshots. Where a specific production format is required by a court in your jurisdiction, that is a scoping conversation and not an assumption either side should make.
Every line below is a commercial or architectural position rather than a sentiment, and each says where to check it. Where something is designed rather than attested, the weaker word is used deliberately — a legal function is the worst possible audience for a word doing more work than it can carry.
The refusals appear alongside the capabilities because a supplier with no stated limits has either not examined the question or is not describing it fully, and counsel is the reader most likely to know which is more common.
Everything below is inexpensive for us to provide before drafting and awkward to extract once a deadline is in the room. That asymmetry is the whole of your leverage and it decays from the day the executive committee hears a date.
The refusal list is the unusual request and the most useful one. Knowing which clauses will be refused before you draft them removes two rounds from the negotiation and tells you something about the counterparty that no positive claim can.
A bounded first engagement — one function, low sensitivity — so that the agreement governing it can be short, specific, and unwound without an argument about what was implied.
Ask for the refusal list first. It is a single request, it costs nothing, and it removes two rounds of redline in which both sides discover positions that were fixed before either of you started. It also surfaces the supplier who has never thought about the question, which is information you would rather have in week one.
Then separate the duties from the drafting. Establish which obligations remain yours as a matter of law before negotiating the ones that are actually allocable. A great deal of negotiating effort is spent moving duties that were never movable, and the effort is not neutral — it consumes the attention that should have gone to the clauses that do allocate something.
And bound the first engagement so that the agreement can be short. A tight scope produces a short contract, a short contract has fewer places for a Tuesday-afternoon problem to hide, and an arrangement that can be unwound cleanly is worth more to your organisation than one negotiated to a standstill.
Then ask which entity signs it and what stands behind it, because uncapped cover from a supplier who could not fund a significant claim is a comfort rather than a remedy, and the comfort has a cost: it ends the conversation about controls, which is the conversation that would actually have reduced your exposure. A disclosed cap with insurance behind it is a smaller promise and a collectable one, and you can price a collectable promise.
Nobody can give you that, and a supplier who writes it into a warranty has written something they cannot perform. What is available instead is narrower and more useful in an examination: the decision types your rules reserve to a person are excluded by architecture rather than by policy, the exclusion extends to merely ordering a queue of people, and every action leaves a receipt naming the actor and the authority relied on. That produces an account of what happened, which is what a regulator asks for and what a warranty of correctness would not have provided anyway.
No, and this page opens with that rather than burying it, because a supplier who lets you believe otherwise has created your worst exposure while appearing to reduce it. The duty stays with you. What changes is your ability to discharge it — a documented authority grant, a per-action receipt, and an excluded-decision list you can evidence architecturally is a stronger examination position than most in-house processes can produce. It remains your duty in every respect.
Your own team applies it, and it overrides scheduled deletion for the material it covers including material that would otherwise have expired. Confirm during diligence that they can do that unaided and without a support ticket, because a hold whose application depends on a counterparty’s response time has somebody else’s clock inside your preservation obligation. That is a five-minute test in a bounded first engagement and it is worth more than the clause describing it.
We will not sign that, and the reason is architectural rather than commercial: a backup image is written once and restored whole, so removing one record from inside it is not an operation our systems can perform on request. What is offered is deletion from live systems on a defined timeline plus expiry of the images themselves on a stated retention schedule. A supplier who signs your clause as drafted is either mistaken about their own architecture or does not intend to comply, and establishing which one is worth more to you than the signature.