Quick answerRun it as its own dedicated cycle, timed to your fiscal year, with three fixed inputs every year: the prior year's cost and ROI numbers, a fresh model and vendor landscape review, and a prioritized backlog of capability gaps sourced from support, sales, and product. Output a funded roadmap and a named owner for each committed initiative, not just a budget line, since a number without an owner rarely survives contact with the rest of the year's priorities.
Why this deserves its own recurring ritual
A one time decision about when to invest more headcount or budget in an AI agent program answers a single moment. An annual planning cycle is different: it is the standing mechanism that decides, every year, whether last year's bet paid off and what the next year's bet should be. Programs that skip this and only revisit funding reactively, when something breaks or a stakeholder complains, tend to drift toward whatever the loudest recent incident demanded rather than what the data actually supports.
Input one: last year's real numbers, not the projection
Every planning cycle should start by closing the loop on what was projected the year before against what actually happened, using the same ongoing ROI tracking cadence that already runs throughout the year rather than a fresh one off analysis built for the planning meeting. If the tracking cadence and the annual planning cycle pull from different numbers, one of them is wrong, and it is worth finding out which before the rest of the cycle builds on either.
Input two: a fresh model and vendor landscape review
The model landscape moves fast enough that a roadmap built on last year's assumptions about cost, capability, and available providers is often stale before the year starts. Build a standing agenda item that forces an explicit answer to whether staying on the current provider and model tier is still the right call, rather than letting that decision persist purely by inertia because nobody put it back on the table.
From the team
We build production AI systems for startups.
LLM pipelines, RAG, and agent workflows that hold up under real traffic — not just in the demo.
Input three: a prioritized capability backlog from outside the AI team
The AI team's own backlog is necessary but not sufficient. Support, sales, and product each accumulate a list of things they wish the agent could do that never gets formally submitted anywhere. The planning cycle should have a defined intake window where those groups submit and rank requests, and the ranking should be visible to everyone who submitted, so people can see why their request did or did not make the cut.
Sizing the initial team versus scaling an existing program
Note that this recurring cycle is a different exercise from the original team sizing and budget decision made when an engagement first starts. That initial sizing answers how to staff a brand new build. The annual cycle answers how to keep staffing and funding an already running program correctly matched to its actual scope, which usually means a much narrower, more incremental set of decisions than the original build required.
Turning the plan into commitments, not just a document
The output of the cycle should be a small number of funded, owned initiatives with named accountable people, not a long wish list ranked by enthusiasm. A roadmap with fifteen items and no owners quietly becomes a roadmap with zero completed items by the next cycle. Better to fund four things properly with clear ownership than to nominally approve fifteen and deliver none.
FAQ
Should the annual cycle be tied to the calendar year or the fiscal year?
Fiscal year, so the funding conversation lines up with the budget process it actually has to compete inside, rather than running on a separate clock that finance has to reconcile after the fact.
Who should own the planning cycle itself?
Whoever owns the program's roadmap day to day should also own convening the cycle, with finance and the functional stakeholders as required participants rather than optional reviewers.