What should I set up before January if my team adopts AI agents next year?
Set up five things before the first agent runs: the recording connection as the default route for agent work, the access inventory with named owners, the approval gates for consequential actions, the delivery standard for client-facing output, and the weekly review rhythm. Review the receipt and actual deliverable before accepting a handover.
The order on this page is its argument. Teams typically adopt agents in January enthusiasm, discover the record problem around the first incident or the first client question, and retrofit under pressure, with the early months left unrecorded for good. December is when the retrofit is avoidable, because the five setups are each small while no agent depends on them yet.
The connection and the inventory come first because they define the world the agents will inhabit: work routes through the recording connection by default, and every agent's reach is inventoried with an owner before it is granted, not discovered after. Gates come third, and December is when they are cheapest to design, because the team can decide calmly which action classes (sends, spends, customer-data changes, deletions) require a recorded human decision, before an agent has made any of them urgent.
The delivery standard and the review rhythm complete the posture: client-facing agent output ships as handovers from day one, so no client ever receives the unrecorded format and has to be migrated later; and the weekly scan-and-sample review starts with the first week of agent work, so the habit forms when the volume is small enough to forgive beginner mistakes. January then becomes what it should be: agents doing useful work inside a system that was already watching.
Checklist: the pre-adoption setup
1. Connection: the recording route installed and set as default for agent work; a test job run and checked. 2. Inventory: every planned agent, its systems, its data scope, its named owner, entered before access is granted. 3. Gates: the action classes needing a recorded human decision, listed with thresholds. 4. Delivery standard: client-facing output ships as handovers; the client note updated. 5. Review rhythm: the weekly scan and sample calendared, owner named, first session scheduled. 6. Policy one-pager: the five decisions above written on one page the whole team has read, filed with the inventory.
Worked example
The test job in item 1 should end in the shape of any accepted record, for example https://zambo.dev/handover/45434a17-4afc-486a-9646-e689bcee4297 : steps recorded during the work, a decision recorded on it, receipts checking out. If December's test job looks like that, January's real jobs will too, because the route, not the memory of the team, produces the record.
Frequently asked questions
We are adopting one small agent for one task. Is all this proportionate?
Run the checklist at one-agent scale: it fits on the one-pager in item 6. Proportionality shrinks the document, not the decisions; the first incident does not check how small the deployment was before arriving.
What if leadership wants to move faster than the setup?
The setup is an afternoon to a day at small scale, and it is the difference between a pilot that can be evaluated and one that cannot. Offer the trade honestly: speed now, with the record retrofit and the unrecorded quarter priced in, or a day of setup and a pilot whose results a board can inspect.
Which setup is most often skipped and most regretted?
The gates. Teams discover their approval classes by watching an agent take an action nobody would have approved, and then design the list in the incident's shadow. December's calm version of the same list is strictly better and takes an hour.
By Brennan Zambo, founder of Zambo.
Last verified: 2026-10-17.
Try a handover
A handover gives your client one proof link with verifiable receipts to review, then Accept or Flag.
try it: say 'use zambo to hand this over to <client>' in your ai