AI policy.
How AI is used in client delivery, what never goes into a model, how AI work is reviewed, and what governance applies to the agents we build. Migrated from /ai-policy; 301 in place.
How does Impulse use AI in client delivery?
AI runs throughout our delivery: research, drafting, code, and the agents we operate internally. We describe this at capability level, and every deliverable passes human review before it ships to a client.
We sell the operating model we run. What we don't do is disguise generated work as artisanal, or bill AI-speed work at hand-speed rates. The pricing page carries that commitment in commercial terms.
What never goes into a model?
Client confidential data is never used to train models, never pasted into consumer AI tools, and never leaves systems governed by the engagement's data terms.
Model access in delivery runs through enterprise agreements whose terms exclude training on our inputs.
How is AI work reviewed before it ships?
A named human owns every deliverable. Generated work is reviewed for correctness, for voice, and for claims, and the reviewer is accountable for it exactly as if they'd typed it.
Internally that review is itself instrumented; the tooling stays internal, the accountability doesn't.
What governance applies to agents we build for clients?
Every client agent ships with four things: a named owner at the client, a written scope, logging on by default, and a review cadence sized to its blast radius. No exceptions, including for demos that go to production.
Agents that touch customers get the heaviest loop. Agents we can't govern, we don't ship. The agents services pages describe the practice; this page is the commitment.