The agent decides. The workflow makes it land.

What a workflow is made of

Each part answers a question that otherwise gets answered by whoever happens to be in the thread.

• Intake that captures the facts — A request arrives with the counterparty, the paper, the value and the ask already structured — so triage never starts with “can you send me the document?”.
• Classification before people — Inbound paper is compared against your standard and sorted: matches, acceptable variation, material deviation. Only the last category consumes human attention.
• Routing by clause — Approvers are attached to clause families and thresholds, so the indemnity question reaches legal and the discount reaches finance without anyone forwarding anything.
• Triggers after signature — Obligations extracted from the executed agreement schedule their own follow-ups — renewal reviews, reconciliations, expiring certificates.

Approvals that find their own owner

Routing is the difference between a policy that exists and a policy that holds. Gates are conditions the platform enforces, not notifications someone may act on.

• Clause-family routing — Indemnity, data, insurance, IP, pricing and term each have an owner, so a reviewer only ever sees the clauses they are accountable for.
• Threshold rules — Contract value, discount depth and spend tier change the approval path, so a small deal does not collect the approvals a large one needs.
• Blocking conditions — Execution can be blocked until approvals are complete, insurance is verified or a consent is received — enforced rather than remembered.
• Delegation and cover — Approvers have delegates, so an absence slows a decision by a person rather than stopping the workflow entirely.
• Parallel where possible — Independent clauses route concurrently instead of queueing behind each other, which is what lets a long agreement close in days rather than rounds.
• Everything attributed — Each routing decision, approval and override is recorded with its identity and timestamp, so the path a request took is reconstructable.

The workflow continues after signature

Most contract tooling stops at execution, which is where the obligations start. Because terms are extracted from the executed document, the schedule maintains itself.

• Obligations as structured terms — Notice windows, renewal triggers, caps, reconciliation dates and expiring requirements come out of the agreement as data, each linked to the clause it came from.
• Triggers with runway — A renewal review fires on the notice period in the document, with the lead time you specify — so the conversation starts while lapsing is still an option.
• The agent picks up the trigger — A fired trigger can open a run: summarize the agreement, recommend a position and start the conversation inside your policy, escalating as configured.
• Amendments update the schedule — An amendment attaches to the clause it changes, so dated obligations reflect the operative terms rather than the original agreement.

Routing is not a job for a prompt

The agent should decide what position to take. It should not be the thing deciding whether finance needs to approve a discount — that is policy, and policy wants determinism.

• Asking a model to route: Who approves what varies between runs
• Asking a model to route: A threshold is a suggestion the model may weigh differently
• Asking a model to route: An omitted approver is invisible until an audit finds it
• Asking a model to route: Changing policy means changing a prompt and hoping
• Asking a model to route: Nobody can state the current rules precisely
• Asking a model to route: The same request can take two different paths
• Routing as configuration: The same request always takes the same path
• Routing as configuration: Thresholds are enforced boundaries, not weighted considerations
• Routing as configuration: A required approver cannot be skipped by a run
• Routing as configuration: Policy changes are explicit, versioned edits
• Routing as configuration: The current rules are inspectable in one place
• Routing as configuration: The agent's judgement is applied inside a fixed structure

What teams ask when configuring

• How much setup does a first workflow need? Less than a full policy exercise. One template, the clauses that always need an approver, and the thresholds that change the path are enough to run a narrow workflow — usually NDAs or order forms — and the rest accumulates from what escalates.
• Can different teams run different workflows? Yes. Playbooks are per workflow, so sales paper, vendor paper and employment paper carry their own intake shapes, routing rules, gates and autonomy levels.
• What drives a workflow from outside Paqt? The agreement surface is available over the Model Context Protocol, so an agent or internal service you run can open a room, read state, submit a request and act on a trigger directly.
• What happens when a trigger fires and nobody acts? It escalates on the path you configured rather than expiring quietly, and the agent can be authorised to open the conversation itself so the window is not lost to an unread notification.