Every company with customers and suppliers has a contract queue. NDAs, vendor paper, order forms, renewals, amendments. Someone reads each one, compares it to the positions the company is willing to accept, pulls out the dates and obligations, and sends it to whoever has to approve it. After signature, someone is supposed to track what the company promised. Often nobody does.

That queue is a reasonable place for an AI system. The documents are text, the rules are mostly written down in a playbook, and the cost of the work is easy to count. It is also a place where a careless build does real damage. So this note stays on four questions: what the work costs today, where the documents go, who reviews the output, and what the system should never do.

Baseline first

Before anyone picks a tool, measure the queue as it runs now. Two weeks of logging is usually enough to tell whether there is a case at all.

A baseline sheet for one contract type: contracts received per month; reviewer minutes per contract, including back-and-forth; days from receipt to signature; share sent to outside counsel; how many deviations from the playbook reviewers caught; renewals or notice dates missed in the last year.

If the answer is twenty NDAs a month, a better template and a shared inbox may beat any model. If it is several hundred vendor agreements a quarter with a two-week turnaround, there is something to work on, and you now have the number a build will be judged against.

Decide where the documents stay

Contracts carry pricing, counterparty details, and sometimes personal data. Some arrive with privileged notes attached. Decide where processing happens before you choose a product, because that choice rules some products out.

  • If the documents must stay in the EU, the model has to run inside the EU too. An EU region of a US provider is a different arrangement, and your counsel should know which one you have. Sovereign AI is how we deploy the model and the workflow inside that boundary.
  • Keep privileged communications out of the system unless counsel has decided how they are handled. The simplest rule is that only the contract and its attachments go in, never the email thread around it.
  • Log what the system read and what it produced, per document, so anyone can answer "where did this contract go?"

Keep human review in the path

The system drafts and the reviewer decides. In practice it does three jobs: it marks where an incoming contract departs from your playbook, it extracts terms such as renewal dates, notice periods, liability caps, and price changes into the register you already keep, and it routes the contract to the right person. A named reviewer accepts or rejects each flag before anything leaves the team.

To know whether the system is good enough, build an evaluation set from contracts you have already reviewed. Your reviewers' past mark-ups are the answer key. Run the system against that set before it sees live work, and again after every change.

Then agree a failure budget with the people who own the queue. Write down what the system may get wrong and what stops it. A renewal date read wrong and caught at review is a nuisance. A non-standard indemnity or a change-of-control clause that goes unflagged is a miss, and one miss of that kind pauses the system until the cause is fixed.

Professional rules apply whatever the tooling. The American Bar Association's Model Rules of Professional Conduct, in particular Rule 1.1 on competence and Rule 1.6 on confidentiality, are where a US legal team will start. The lawyer who relies on the output stays responsible for it.

What not to automate

  • Legal advice, and the decision to accept a deviation from the playbook.
  • Negotiation positions, and anything sent to a counterparty without a person reading it.
  • Privilege calls.
  • Predicting how a dispute or a court will turn out.
  • Unusual or high-value deals. Route those to a lawyer at intake and keep them out of the automated path entirely.

A sensible first build

Pick the contract type with the most volume and the most stable playbook. Vendor agreements on your own paper are a common choice. Build the playbook check and the term extraction into the register you already use, with a reviewer approving every output. After a few weeks, compare turnaround and reviewer minutes against the baseline. Widen to the next contract type only when the system passes its evaluation set on that type too.

That build is a Business Process AI engagement: one workflow, a measured baseline, and a running system your team operates. If the documents cannot leave the EU, it runs on Sovereign AI. If you want to know whether your queue is worth it, book a scoping call and bring the baseline numbers, or a rough guess at them.

Sources: American Bar Association, Model Rules of Professional Conduct.