Can the assistant accept a term our committee has not approved?

Not within the mandate mechanism, no. Positions outside the mandate are not negotiated — the run stops and escalates. Clauses you mark as gated stop every time regardless of autonomy level, so the most permissive setting still cannot carry an audit-rights change past your reviewer.

How do we evidence the mandate that was in force at the time?

Each run records the mandate it executed under, so the boundary is part of the agreement history rather than a configuration someone has to reconstruct. When a mandate changes, earlier runs keep pointing at the version that governed them.

Our own systems need this data. How do they get it?

Over the Model Context Protocol surface, with scoped credentials. Your agents and services read rooms, agreements, clauses, obligations and run transcripts directly and operate under the same gates as the in-product assistant. We do not claim named connectors, because we would rather give you a surface you control.

What happens if something goes wrong mid-negotiation?

There is an organization-wide kill switch that halts every agent at once, and any individual run can be stopped and taken over by a human in place. The partial run remains in the record — a halted negotiation is still evidence of what was said.

A deviation you cannot explain is the expensive kind

Financial institutions do not struggle to negotiate; they struggle to account for it afterwards. Paqt makes the mandate explicit before the negotiation starts, and the run log becomes the explanation — which clause moved, against which approved position, accepted by which named person.

The agreements that attract scrutiny

Each of these ends up in front of someone who did not negotiate it and needs to understand why it says what it says.

Counterparty onboarding — High-volume paper where the terms are largely settled and the work is confirming that this particular counterparty's version matches the one your committee approved.
Outsourcing and service providers — Agreements where audit rights, exit provisions and subcontracting are the clauses a supervisor will ask about. Each is configured to stop the run and wait for a named reviewer.
Vendor and technology contracts — The ordinary commercial volume — licences, data feeds, professional services — negotiated inside policy so review attention is not spent on payment terms.
Distribution and intermediary agreements — Long-lived arrangements with commission schedules and conduct obligations that must stay knowable years after the person who agreed them has moved on.
Confidentiality and information sharing — NDAs and data-sharing arrangements at volume, where the handling and retention clauses are the ones that matter and the rest is form.
Renewals and repapering — Amendment cycles where the operative terms must remain readable rather than reconstructed from a stack of side letters.

A mandate is your risk appetite, written down

Most institutions already have approved positions — they live in a playbook, a committee minute or a senior lawyer's memory. A mandate turns that into something the negotiation is actually bounded by.

Approved, tradeable, forbidden — Each clause gets one of three states and, where relevant, a band. The assistant negotiates inside it and the boundary is enforced rather than advisory.
Gates that outrank autonomy — Audit rights, exit provisions, subcontracting and anything else you name always stop for a named reviewer, regardless of how permissive the run's autonomy level is.
The right function, by clause — Risk, compliance, information security and the business owner each receive the clauses that are theirs, with the deviation and the approved position attached.
The justification is the record — Every acceptance carries the clause, the version, the mandate in force and the identity that approved it — so explaining a deviation later is a lookup, not an investigation.

The questions internal audit arrives with

These are lookups against the agreement history rather than exercises in reconstruction, because the history was produced by the negotiation rather than assembled after it.

On a single agreement: Which clauses deviate from the approved position
On a single agreement: What the counterparty originally asked for
On a single agreement: Who approved each deviation, and at which version
On a single agreement: Which mandate was in force when the run executed
Across a portfolio: Which clauses are conceded most often
Across a portfolio: Where the approved position no longer survives contact
Across a portfolio: Which counterparties hold non-standard terms
Across a portfolio: Which agreements carry gated clauses still under review
After execution: Obligations extracted from the executed agreement
After execution: Notice and exit windows surfaced from the document
After execution: Amendments resolved against the clauses they change
After execution: The complete record exportable as one package

A rollout your risk function can sign off

Nobody should turn an agent loose on regulated counterparty paper on day one, and the autonomy dial exists so that nobody has to.

Write down one playbook — Take a single agreement type and state the approved position per clause, the tradeable bands and the clauses that always escalate. This is work your institution has usually already done informally.
Run it in shadow mode — The assistant drafts positions and prepares escalations, nothing reaches a counterparty, and your team reads the transcripts against what they would have done.
Move to supervised — The assistant takes turns inside the mandate and escalates the gated clauses. Turn and spend caps bound each run, so a stubborn counterparty produces a pause rather than a surprise.
Review the deviation pattern — After a batch, read what was conceded and where. Either the approved position holds, or you have evidence the committee should revisit it — which is itself a useful output.
Expand by agreement type — Each new type gets its own playbook, mandate and gates. Autonomy is per mandate, so the settled paper can run freely while the sensitive paper stays supervised.

What we do not claim here

Financial institutions are evaluated on what they can evidence, and we are not going to give you something to evidence that does not exist. We describe the controls that are in place and leave the regulatory judgement with the people who answer for it.

An automated test in our build fails if a certification claim appears in any marketing copy. If a completed certification is a hard procurement requirement for you today, we would rather you know that at the start of an evaluation than at the end.

Not claimed: No compliance certification and no completed third-party audit report of our controls
Not claimed: No assertion that using Paqt satisfies any supervisory expectation that applies to you
Not claimed: No named integrations with your core banking, risk or reporting systems
Not claimed: No regulatory advice — the approved positions in a mandate are yours, written by your own functions
True today: Mandates that bound a negotiation, with clause-level gates no autonomy level can override
True today: Timestamped, tamper-evident history with every approval bound to a version and an identity
True today: Agent actions attributed to the agent and the mandate in force at the time
True today: A Model Context Protocol surface with scoped credentials, so your own systems read the record directly

What risk and compliance ask first

Can the assistant accept a term our committee has not approved? Not within the mandate mechanism, no. Positions outside the mandate are not negotiated — the run stops and escalates. Clauses you mark as gated stop every time regardless of autonomy level, so the most permissive setting still cannot carry an audit-rights change past your reviewer.
How do we evidence the mandate that was in force at the time? Each run records the mandate it executed under, so the boundary is part of the agreement history rather than a configuration someone has to reconstruct. When a mandate changes, earlier runs keep pointing at the version that governed them.
Our own systems need this data. How do they get it? Over the Model Context Protocol surface, with scoped credentials. Your agents and services read rooms, agreements, clauses, obligations and run transcripts directly and operate under the same gates as the in-product assistant. We do not claim named connectors, because we would rather give you a surface you control.
What happens if something goes wrong mid-negotiation? There is an organization-wide kill switch that halts every agent at once, and any individual run can be stopped and taken over by a human in place. The partial run remains in the record — a halted negotiation is still evidence of what was said.