How I decide which AI initiatives to kill

Every AI portfolio review I sit in is missing the same verb. Executives will fund things all day. They will revisit things, pause things, deprioritize things, put things on the roadmap for next half. What almost nobody does is kill things, in writing, with a reason attached. Killing a project means telling someone their idea lost, and most leadership teams would rather let twelve initiatives starve slowly than tell one sponsor no.

That politeness has a price. Every initiative that should be dead is still consuming budget, engineering attention, and a slice of the only resource that really matters early on: the organization’s willingness to believe AI can work here. After three years shipping production AI at AnswerAI, the AI product company I run, I’ve watched more value destroyed by projects nobody would kill than by projects that failed honestly.

So the discipline I run with clients is blunt: once a quarter, every AI initiative gets one of three verdicts, on paper. Fund, hold, or kill. No fourth option, no “continue monitoring.”

Unkilled portfolios kill themselves

If deliberate killing sounds harsh, look at what happens without it. S&P Global Market Intelligence found that the share of companies abandoning most of their AI initiatives jumped from 17% to 42% in one year, and the average organization scrapped 46% of its AI proofs-of-concept before they reached production. Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, and unclear business value.

Read those numbers as what they are: kill decisions that got made anyway. Nobody escapes killing projects. The only choice is whether it happens deliberately and early, when the sunk cost is a workshop and a prototype, or by exhaustion eighteen months in, after the spend, when the abandonment takes the team’s credibility down with it. A 42% abandonment rate is not a technology statistic. It’s what a portfolio looks like when the kill decisions were deferred until they made themselves.

The quarterly memo

The mechanism is deliberately unglamorous. Once a quarter, every AI initiative in flight gets a short entry in a single memo, and each entry has to answer three questions before any verdict is discussed.

First, which P&L line does this move? Not “efficiency” or “productivity.” A line: labor cost in the tax practice, revenue per proposal, admin overtime. If nobody can name the line, that’s usually the whole review right there.

Second, what has been spent to date, and what value has come back to date, in the same units? Dollars against dollars, or hours against hours. Projected value doesn’t count in this column. An initiative eight months in with all of its value still in the projection column is telling you something.

Third, who owns it? A name, not a team.

Then the verdict, in writing, signed by the sponsor: fund, hold, or kill. The writing matters more than it sounds. A verdict delivered verbally in a meeting is renegotiated by the following Tuesday. A verdict in a memo that the whole leadership team saw is a decision. I’ve published a sanitized example of this memo — the sample AI Investment & ROI Memo from a composite client, assembled from patterns across real engagements rather than a disguised real firm — alongside the other sample deliverables.

What earns a fund

A fund verdict requires a baseline, a measured delta, and a named owner. All of them. Enthusiasm, usage stats, and stakeholder love earn nothing.

Here’s what that looks like with real math, from the composite memo above. An accounting firm’s engagement-letter drafting workflow: partners and managers were spending 3.5 hours per letter, timed against actual letters before anything was built. The build cost $28K. In the seven weeks after it shipped, 212 letters went through the workflow at an average of 45 minutes each. That’s about 2.75 hours returned per letter, roughly 560 hours, and at the firm’s blended $85 an hour, close to $50K of gross capacity in seven weeks.

I still called the payback about five months, not five weeks, for two reasons. I count from the day the SOW was signed, not the day the thing shipped, because the money left in March even though the value started in May. And I credit a saved hour at half its blended rate until it demonstrably turns into billable work or reduced overtime, because a saved hour that evaporates into slack is worth nothing. Even discounted that hard, the verdict wrote itself: fund, and extend the same approach to proposals.

That’s the standard. Notice what’s absent: no adoption percentages, no seat counts, no survey saying people like it. Those are inputs. The fund column is for initiatives that can show the before-number, the after-number, and the person accountable for the gap between them.

Hold is not a soft yes

Hold is where the discipline usually rots, because hold feels kind. It lets everyone leave the room without conflict. So I enforce a rule: there are exactly two honest reasons to hold, and every hold gets a re-entry condition and a date. A hold without both is a kill wearing a hold’s clothes.

The first honest reason is that a dependency isn’t ready. Same composite client: an internal knowledge assistant scored near the top of the backlog, aimed at roughly 1,800 hours a year that staff spent hunting for procedures and prior research memos. On paper, an obvious fund. But the document corpus it would draft from was a mess — in places, five live versions of the same tax template, and nobody could say which one was current. An assistant retrieving from that corpus wouldn’t save 1,800 hours; it would confidently serve stale answers with a citation. This failure mode is common enough that Gartner predicts organizations will abandon 60% of AI projects that aren’t supported by AI-ready data through 2026. So the verdict was hold, with a condition: corpus cleanup finishes, a named person confirms one canonical version per template, and the initiative re-enters review the following quarter. Not dead. Waiting on something specific.

The second honest reason is that a decision upstream is unmade. You can’t fund an expansion of a pilot tool while the leadership team hasn’t decided whether to consolidate on one platform. Holding until a dated decision gets made is legitimate. Holding because nobody wants to force the decision is not, and the memo should say which one is happening.

Anything else — the sponsor is attached to it, it might be strategic someday, we’ve already spent so much — is a kill.

The killed list stays published

The last piece is the one teams resist most: killed initiatives stay on the memo, visibly, with their one-line reasons, quarter after quarter. Website chatbot, killed in March, risk exceeds value. Fully automated tax prep, killed, regulatory exposure plus quality risk. The list doesn’t shrink.

The reason is that ideas don’t die in meetings. They die in meetings and resurrect in hallway conversations, usually reintroduced by someone who wasn’t in the room, pitched to someone who doesn’t remember the reasoning. Without a published kill list, your organization re-litigates the same five bad ideas every quarter, and each re-litigation costs senior hours. With one, the answer is a pointer: killed in March, here’s why, here’s what would have to change to reopen it.

There’s also a portfolio effect. BCG’s survey of 1,000 executives found that the companies actually generating value from AI pursue roughly half as many opportunities as their less advanced peers. Half. The winning behavior is concentration, and concentration is downstream of killing. Every visible kill is budget and attention flowing back to the two or three workflows that showed a real delta.

The quarterly memo takes about a day to prepare once the baselines exist, and one uncomfortable meeting to deliver. That’s the entire cost. Against it, weigh what the abandonment statistics describe: nearly half of AI work dying anyway, late, expensive, and demoralizing, because nobody was willing to write the word kill while it was still cheap. The verdicts get made either way. The only question is whether you’re the one making them.