Skip to content
Book a consult

Insights  /  Operating model

The AI operating cadence: a rhythm teams can sustain

Intake, prioritization, review, measurement — and how to keep it lightweight.

Sopra Labs AI  ·   ·  7 min read
SopraLabs Hero Cadence

Key takeaways

  • AI programs rarely fail in one meeting. They drift: requests scatter across inboxes, pilots run with no decision attached, and the committee meant to steer them stops meeting.
  • Most pilots stop short of production. In Deloitte’s survey of 3,235 business and IT leaders, only 25% of respondents had moved 40% or more of their AI pilots into production (Deloitte).
  • A cadence that lasts has four routines and one document: a single intake path, a short scoring session every two weeks, a monthly review that ends in scale, fix or stop, a quarterly refresh, and a decision log.
  • Size is the design constraint. A rhythm that asks three people for thirty minutes survives a busy quarter, while a standing hour for eight senior leaders usually does not.

The committee that stopped meeting

The AI steering committee launches in January with a full room. Eight leaders, a shared folder, a standing hour every other Thursday. For two months it works. Ideas get heard, three pilots get approved, and the minutes go out the same day.

Then quarter-end arrives and a session is canceled. The next one is moved, then shortened, then held without the CFO. By June the committee has not met in seven weeks.

The work does not stop. It just loses its address. Requests now arrive in the COO’s inbox, in hallway conversations, and in a chat thread nobody owns. The three pilots are still running. One is clearly useful. One has a vendor invoice due and no sponsor in the building. Nobody can say what the third is for.

In the operations review, the COO asks which pilots are moving forward. People look at each other. Nobody has decided, because no meeting exists where deciding happens.

Pilots drift when nothing forces a decision

Nothing in that room went wrong on purpose. The committee ran on enthusiasm, and enthusiasm is a poor scheduling system. Eight senior calendars every other week is an expensive habit, and the first quarter-end exposed the price.

Pilots that stop short of production are common enough to measure. Deloitte surveyed 3,235 business and IT leaders in August and September 2025 and found that “only 25% of respondents have moved 40% or more of their AI pilots into production” (Deloitte). Gartner says, without describing its method, that by the end of 2025, at least 50% of generative AI projects were abandoned after proof of concept, for reasons that include escalating costs and unclear business value (Gartner).

Smaller companies sit on the harder side of the gap. In McKinsey’s 2026 survey, 54% of respondents from organizations with at least $1 billion in annual revenue report scaling AI across the enterprise, compared with one-third of those from smaller organizations (McKinsey). All three firms sell advice on this problem, so read the figures as direction, not precision.

Stopping a pilot is not the failure. A weak pilot that gets stopped is the system working. The failure is the pilot nobody stops and nobody scales. A pilot has a natural beginning, because someone is excited. It has no natural end. Unless a date exists on which someone must say scale, fix or stop, the default is to continue, and continuing is not a decision.

What separates the organizations that keep AI work alive looks like routine. In a Gartner survey of 432 respondents, 45% of leaders in organizations with high AI maturity said their AI initiatives remain in production for three years or more, against 20% in low-maturity organizations. Gartner credits “choosing AI projects based on business value and technical feasibility” along with governance structures and engineering practices (Gartner).

That echoes an older finding from another setting. The U.S. Census Bureau’s 2010 management practices survey drew responses from approximately 37,000 manufacturing establishments. A Census paper on it reports that firms adopting more structured practices for monitoring, targeting and incentives “are more productive, more profitable, and grow faster” than firms that adopt fewer (Census Bureau). Regular review is not overhead on top of the work. It is how the work gets finished.

Build a rhythm small enough to survive

The fix is not a better committee. It is four small routines and one document.

Open one front door. Every AI idea enters through the same short form: the problem, who has it, how often it occurs, what data it touches, and who would own the result. Requests that arrive by email or in a hallway get the link. The point is not to ration ideas. It is to make them visible and comparable.

Score every two weeks, in thirty minutes. Three people are enough. Each new request gets a simple score on value, feasibility, adoption and risk, and one of three answers: start, wait or no. A rubric makes the decision faster and takes most of the politics out of it. The answer and one sentence of reasoning go in the decision log.

Review monthly, and end every item with a verdict. Each initiative in flight gets one page and one of three outcomes: scale, fix or stop. A fix carries a date. A second fix with no progress becomes a stop. NIST’s AI Risk Management Framework describes a similar checkpoint: “A determination is made as to whether the AI system achieves its intended purposes and stated objectives and whether its development or deployment should proceed” (NIST AI RMF, MANAGE 1.1). Record the verdict, the reason, the decider and the date. The log keeps the meeting short, because nobody reargues last month.

Refresh every quarter. Rescore the whole portfolio, look at which tools and models have changed, and adjust priorities. New product announcements go onto a change backlog and wait for this session, which spares the team a reshuffle every time a vendor ships something. NIST asks organizations to plan periodic review, “including determining the frequency of periodic review” (NIST AI RMF, GOVERN 1.5). It names no interval. Quarterly is frequent enough to keep up with the tools and slow enough to let a decision play out.

Shorten the meeting before you cancel it. A ten-minute session that makes two decisions keeps the habit alive. A canceled one teaches everyone that the rhythm is optional, and the requests go back to email.

The usual objection is that nobody has time for more meetings. This rhythm replaces meetings. Three people for thirty minutes twice a month, plus one monthly review, takes less leadership time than the original eight-person committee did. It also removes the unscheduled version: the inbox threads, the vendor calls and the hallway lobbying that fill the gap when no decision point exists.

The same question, with a log to answer it

The COO asks again which pilots are moving forward. This time someone opens the decision log. One pilot was scaled in July. One has thirty days to fix a data problem. One was stopped in August, with the reason written beside it.

The committee did not fail for lack of commitment. It was sized for the launch, not for the ninth month. The rhythm that replaced it is smaller and duller, and it is still meeting.

The opening scenario is a composite of common patterns, not a client engagement. SopraLabs designs this rhythm with leadership teams as the Operating Cadence module of our services: intake → prioritization → build → deploy → measure, delivered as an intake process, rituals and meeting templates.

Frequently asked questions

What is an AI operating cadence?

An AI operating cadence is the recurring schedule on which an organization takes in AI ideas, prioritizes them, reviews work in flight and revisits its tools and priorities. In practice it is a few short, fixed meetings and a decision log. Its purpose is to make sure every initiative reaches a decision instead of drifting.

How often should an AI steering committee meet?

NIST sets no interval. Its AI Risk Management Framework expects periodic review to be planned and leaves “determining the frequency of periodic review” to the organization (NIST). A workable pattern is a thirty-minute scoring session every two weeks, a monthly review of work in flight, and a quarterly refresh of tools and priorities.

How do we decide whether to scale or stop an AI pilot?

Decide on a fixed date, against the purpose the pilot was approved for. At each monthly review a pilot gets one of three verdicts: scale, fix by a named date, or stop. A pilot that misses its fix date twice should stop, and the reason goes in the decision log.

What belongs in an AI decision log?

Each entry records the decision, the reason, who made it and the date. That covers intake answers (start, wait or no) and review verdicts (scale, fix or stop). One shared document is enough, and it is what lets a new executive or an auditor see why something was funded or ended.

Sources

Keep reading

Browse all insights