Deviation lists are long. Ranking is what saves the reviewer time
A mechanical comparison against a playbook produces 30 findings, most of them inconsequential wording differences. For a reviewer, going through thirty items of the same kind is the same as reading the document once, plus extra steps.
Whether a difference matters is a judgment of substance, not of form. An indemnity that is reworded but ends in the same place is not a finding. A definition that is quietly changed to broaden the scope is a finding, and that kind of change does not announce itself.
Clause interactions are where this becomes genuinely difficult. A liability limit that is acceptable together with a wide indemnity is a different position from either clause on its own. A sentence-by-sentence comparison cannot detect that.
This is not a legal opinion. The agent compares the contract with your stated positions and highlights those that are worth looking into. Your legal adviser decides whether any of it is acceptable.
How the agent is built in Agent Studio
Your playbook is the boundary. Prompt skills include the preferred wording for each type of clause, your acceptable range, your stated red lines, and the reasons for those red lines, where you have recorded them.
The system prompt makes clear what the agent does not do. It compares the contract with your positions. It does not evaluate legal risk, and it never says whether a term is enforceable, common in that market, or advisable. That limitation is stated in those exact words because that is the sentence a reader will look for.
It also asks the agent to put clause interactions in a separate section, because a clause-by-clause review does not capture them.
The review flow provides the clause map to the Agent Chat node, so the agent reasons about structured clauses rather than a PDF.
Notion receives the ranked deviations, the interaction notes, and a separate list of clauses that the agent could not confidently classify. That final list matters because an unclassified clause is not a clean one.
Slack tells the reviewer that the first pass is ready. No agreement has been accepted, signed, or negotiated. The prompt has no authority to do any of those actions.
What this agent is built from
- Prompt skills: hold your playbook, including preferred wording, acceptable range, red lines, and their reasons.
- System prompt: compares against your positions only. Never assesses risk, enforceability, or market norms.
- System prompt: reports clause interactions separately from clause-by-clause findings.
- Agent Chat (Util): receives the structured clause map from the review flow.
- Notion (Integration): receives ranked deviations, interactions, and unclassifiable clauses.
- Slack (Communication): tells the reviewer that the first pass is ready.
- When an ActionFlow is enough: If you are checking presence and absence against a clause list, build the workflow. That is the right first build.
Frequently asked questions
Start building AI workflows
Create a free account, open a template or a blank canvas, and run your first ActionFlow.
Newsletter
Get product updates
New nodes, agents, and product notes. We send mail only when we have something worth opening.