Run Your AI Like a Portfolio: Budget in Quarters, Not Years
The roadmap that died before its approval meeting
Earlier this year we reviewed an AI roadmap for a client in logistics. Fourteen initiatives across eighteen months, dependency arrows, swim lanes, the works. It was a genuinely well-made document.
It was also already wrong. In the seven weeks between the final draft and the board meeting, one vendor had shipped a feature that made initiative three redundant, and a round of model price cuts had quietly rewritten the business case underneath two others. Nobody updated the deck; boards do not appreciate deltas at approval stage. So the company approved a plan that both sides of the table half knew was obsolete, and delivery would now be governed against it for a year and a half.
None of this happened because anyone was careless. It happened because the planning instrument assumes a stable technology baseline, and there isn't one.
The ground moves quarterly
Consider what a planner has had to absorb in the past eighteen months alone. Model tiers appearing above what was previously the top of the range. API prices falling hard while total spend went up anyway. Vendors folding features into their platforms that six months earlier would have justified a build. We argued in Build, Buy, or Wrap that this cadence punishes long builds. It punishes long plans for the same reason: whatever you commit to in January is depreciating against a baseline that improves without asking you.
An eighteen-month roadmap in this environment is not a plan. It is a bet that the ground will hold still, placed at the exact moment the ground is at its least still in decades.
Positions, not projects
The alternative comes from a discipline that has managed uncertainty for a long time: portfolio management. Instead of a queue of committed projects, you hold positions of three kinds.
Production systems have graduated and earn measured returns. They get funded like operations, with the acceptance and monitoring regime that implies, and they stay in the portfolio only while the returns hold.
Pilots are candidates for production, running under the graduation criteria that most pilots never get given: defined thresholds, defined windows, decided before launch.
Spikes are small purchases of information, in the sense we described in Planning for a Number You Don't Have Yet. Capped budgets, a few weeks each, and a measurement as the deliverable. This is where new model capabilities get tested against your actual work, cheaply.
The quarterly rebalance
Once a quarter, one meeting, three possible calls per position: scale it, hold it, or kill it. Budget freed by a kill flows to whatever now looks strongest, which this quarter might be something that did not exist when the year started.
Finance keeps its annual process; that fight is not worth having. The envelope stays annual. What moves inside the year is allocation authority. The board approves "€2 million for the AI portfolio, reviewed quarterly against these rules" rather than fourteen line items with names and end dates.
And kills happen without ceremony. A spike that dies after six weeks and €30,000 is not a failed project to be explained in a lessons-learned session that everyone dreads. It is the portfolio doing its job. The moment killing a position requires a defensive memo, people stop killing positions, and the portfolio silts up with the walking dead.
Measure the portfolio, not the bet
This is the piece that changes the conversation with the board. Individual bets in an uncertain environment fail routinely, and judging each one in isolation teaches your organisation to only propose safe ones. Judge the portfolio instead.
A healthy one has a shape: roughly a third of spikes should die. If nothing dies, you are not exploring anywhere near the edge of what is possible. If everything dies, your filters are broken. In between sits a hit rate you can track, report, and improve, which is more than can honestly be said for a Gantt chart's milestone completion percentage.
What the board gets under this model: a one-page direction statement (where AI should change the company's economics, and where it shouldn't), the current portfolio with each position's status, and the hit rate over trailing quarters. Less impressive-looking than fourteen swim lanes. Considerably harder to be wrong with.
Still governing AI delivery against a roadmap nobody believes? Get in touch to talk through what a quarterly portfolio cadence would look like for your organisation.