AI for Contract Review in Government: Call-Offs, Model Clauses and Extension Options
Contract review in a public body is not the same job as contract review in a company. You are rarely negotiating from a blank page. The terms are largely set for you: a framework agreement someone else competed, a model services contract you are required to use, mandated schedules you cannot quietly drop. The difficulty is checking, at volume, that what is about to be signed matches what you are obliged to sign, and proving later that it did.
The second difficulty is memory. A ten year contract outlives the category manager who awarded it and the lawyer who reviewed it. Extension options pass unexercised, or get exercised by default because nobody noticed the alternative while there was time to compete. Both problems are about reading carefully across a large document set, which is where an agent earns its place.
The call-off against the framework it sits under
A call-off is supposed to be the easy case: the framework fixed the terms, so the call-off fills in the blanks. In practice the blanks are where the risk sits, and the framework permits only a narrow band of adjustment that is easy to exceed without noticing.
An agent reads the framework agreement and its schedules alongside the draft call-off, then reports clause by clause: which framework terms the call-off reproduces unchanged, which it has varied, whether each variation falls inside the permitted band, and which mandatory schedules are present, absent or substituted. It also checks the mechanical things that cause disproportionate trouble: whether the services ordered fall inside the lot and scope description, whether pricing matches the rate card, and whether the term plus every extension option exceeds the maximum call-off duration. The output is a deviation list with both wordings side by side, plus whatever the agent could not resolve, quoted not guessed.
Checking a draft against the model clauses you must use
Most public bodies maintain, or are given, a library of model clauses and standard terms. The failure is not disagreement with them. It is drift: a clause copied from a contract that was itself copied from an older template, a schedule that was correct two versions ago.
An agent holds the current library as its reference set and compares a draft against it, reporting each mandated clause as present and identical, present but altered with the alteration shown, or missing. For altered clauses it states the effect rather than the diff: a liability cap moved, an audit right narrowed, a step-in right removed, an indemnity reduced to a warranty. It also flags internal inconsistency, where hand-assembled contracts fail most often: defined terms never defined, cross-references to the wrong schedule.
Supplier-proposed amendments to standard terms
When a supplier returns your standard terms with proposed changes, the cost is not the clauses they flagged. It is finding the ones they did not. An agent produces the complete change list, including single-word edits: "shall" to "may", "will ensure" to "will use reasonable endeavours to ensure", a deleted "without limitation".
It then sorts that list by effect on your position, and marks which changes touch terms you have no authority to concede because they come from a mandated template. Some positions are commercial judgement; others are not yours to trade at all.
Obligations that must pass to the lower tiers
Your contract with a prime supplier usually requires obligations to pass down: data protection terms, security requirements, audit rights, payment terms to the lower tiers, transparency duties, sometimes social value commitments. An agent reads the head contract and whatever subcontract documentation you have a right to see, then builds a flow-down map: for each obligation that must pass down, whether it appears, whether it appears weakened, and where the chain stops. It also tracks whether subcontractor appointments that required your consent received it. Where you have no visibility of a tier, it says so rather than reading silence as compliance.
Data protection and data processing schedules
Processing schedules are both highly templated and routinely wrong, which suits automated checking. The agent examines the schedule for what is either present or not: controller and processor roles for each activity, nature and purpose, categories of data subject and personal data, retention and deletion periods rather than placeholders, international transfers and the safeguard relied on, security measures stated specifically rather than as "industry standard", and breach notification timescales.
It compares that against your required position and against the processing the specification actually describes, which is where the real mismatches appear: a schedule promising anonymised analytics attached to a service that will handle identifiable records.
Security and assurance schedules
Security schedules fail quietly because the obligations are deferred: a security management plan within thirty days of commencement, an annual penetration test, a certification maintained throughout the term, notification of changes to hosting location. An agent extracts each as a dated obligation with an owner and a required artefact, so you can ask which security deliverables are outstanding across the portfolio and which certifications lapse next quarter.
Social value and local employment commitments
Social value is won in the tender and lost in the contract. The bid promised apprenticeships, local supply chain spend or a carbon reduction trajectory. Whether those promises became binding obligations, with measurement definitions, reporting frequency and a consequence for shortfall, is a question most organisations cannot answer.
The agent reads the award documentation and the executed contract together and reports where a bid commitment became an obligation, where it became an aspiration with no reporting duty, and where it disappeared. In delivery it tracks reports received against reports due, and figures reported against figures committed.
Variations and change control over a long contract
The contract you are managing is the original document plus every variation since, and the two are rarely read together. An agent maintains the consolidated position: which clauses have been varied, by which instrument, when, and what the operative wording is now. It checks each proposed variation against the contract's own change control procedure, and against the limits the procurement rules you work under place on modifying an awarded contract. Where a change looks material, it flags it for legal review with the cumulative value of prior variations attached, because the aggregate is what people forget.
The award record that has to survive audit and disclosure
Public contract records are read years later by people who were not there: an internal auditor, an external audit body, a committee, a challenger's solicitors, a journalist making a freedom of information request.
An agent checks completeness against your records management requirements: the executed contract with all schedules and signature evidence, the award decision and its rationale, evaluation records, conflict of interest declarations, the approval that authorised the spend, and the variation instruments in sequence. It reports missing items and inconsistencies, such as a contract value in the award report that does not match the executed contract. It can also draft a redaction proposal for a disclosure request, marking sensitive pricing, personal data and security detail with the exemption relied on, for a human to decide.
Expiry dates and extension options across the portfolio
Almost every public body has a contract register with gaps, and almost none can state confidently what expires in the next eighteen months. An agent reads the contract documents rather than trusting the register, extracts commencement date, initial term, expiry, each extension option with its length, and the window and notice period for exercising it, then reconciles that against the register and reports the differences. That converts a retrospective scramble into a forward plan: what expires when, what can be extended, what must be re-competed, and when each procurement has to start to land before expiry.
Where AI fails here
A flagged clause is not legal advice and cannot be. The agent can tell you an indemnity has been narrowed or a cap moved. Whether the resulting position is acceptable, whether a clause is enforceable as drafted, and how a court would read an ambiguity are questions for a commercial lawyer. Treat agent output as a legal opinion and you have replaced review with reassurance.
Public bodies also carry a transparency burden that changes what counts as a usable answer. If you cannot explain why a decision was taken, you cannot defend it to an auditor, a committee or a challenger, so an unexplainable recommendation is worthless even when it is right. Every output needs a clause citation and the source wording, and the decision must be recorded as a named person's judgement.
Procurement challenge risk makes a wrong consistency call expensive in a way it is not in the private sector. An agent that reports a call-off as compliant when it exceeds the framework's permitted scope has contributed to a decision that may be challengeable. So the honest configuration over-reports uncertainty: a flag that turns out to be nothing costs ten minutes, a missed deviation costs far more.
Model clause libraries drift between versions, and an agent will compare against whatever you gave it, including a superseded template, without complaint. Version control of the reference set, and of the framework documents the owning authority keeps amending, is a live operational responsibility, not a setup task.
Historic contracts are frequently scanned or already redacted, and extraction accuracy on a photocopied signature page is materially lower than on clean digital text. Measure that cohort separately and report it honestly rather than averaging it into one reassuring number.
Finally, an agent cannot fix a contract register that is a spreadsheet with gaps. It can tell you what the documents it was given say, and where the register disagrees with them, but it cannot find a contract nobody filed. If the records are incomplete, the output is an accurate map of an incomplete estate, which is useful but not the same as a complete one.
Where to start
Start with portfolio mapping and expiry and extension extraction. It is bounded, it uses documents you already hold, and every output is checkable against the source clause. Nothing in it requires the agent to judge whether a term is acceptable, so you learn how reliable the extraction is before anything consequential depends on it.
The alternative narrow start, if your expiry position is already sound, is checking call-offs against a single current framework: one reference set, one repeatable comparison, with a commercial lawyer reviewing every flag for the first few months so your accuracy picture is real rather than assumed.
Adopt AI Partners builds these agents inside your own contract and records systems, so the audit trail sits where your auditors already look, and stays until your commercial team runs it without us. If you want to talk through how contract review works in your organisation, describe the process to us at hello@adoptaipartners.com.