Sample report — free to read

Local agents at client desks

September 2026 · a Prism Folio Signal Brief

The pattern this month: small AI agents that used to live in your terminal are landing on client machines and client processes — drafting email, filling CRM fields, prepping proposals — without a data scientist in sight.

If you consult, freelance, or build for other people, this is less a hype story and more a scope and liability question for your next engagement.

Three concrete takeawaysWhat changed

Client-side agents are now commodity to run
A single laptop can run a capable local model for drafting and routing tasks. The moat has moved from "can we run a model" to "who owns the workflow, the data, and the failure modes."
The paper trail is the product
What clients pay for is not the agent — it is a documented loop: what the agent was asked, what it returned, what a human reviewed, and what ships. Clients are starting to ask for this artifact in scoping documents.
The checklist is a deliverable, not a footnote
Teams that ship credible agent workflows attach a one-page runbook: boundaries, test cases, rollback, and a named human owner. A 20-minute brief teaches you to write that page; the page is what survives the engagement.

One-page action checklistDo this next

Before your next client call

Pick one bounded task — drafting, extraction, or triage — not "automate everything."
Write the input/output contract — what the agent receives and what it must return, in plain language.
Define the human gate — which outputs ship unreviewed, which require a named reviewer.
Stub a test set — ten real examples the agent must pass before the client sees it.
Name the failure mode — what the agent is allowed to do when it is confident, and who it asks when it is not.
Save the runbook — one page: boundaries, tests, rollback, owner. That page is the deliverable.