Your 2027 AI budget has no line for deciding what to finish
On September 15, Teresa Torres and Petra Wille published a podcast with a title that reads like a correction. "I don't think delivery is free," Torres says. "I don't think delivery will ever be free." The first sixty to seventy percent of building a thing has genuinely gotten faster over the past eighteen months, but closing the last thirty percent, in her words, is still "months to years of work."
Three days later, Stefan Wolpers put that episode second in his weekly roundup, next to four people making the same argument in four different vocabularies inside the same seven days. Josh Elman: demos are cheap, working products are not, and judgment still rules. Mike Cohn, on AI-generated backlogs producing false confidence. Marily Nika, on defining minimum viable quality by ranking the failure modes you are willing to recover from. Addy Osmani, on unsupervised agents quietly loading technical debt into legacy systems.
Every one of those pieces is written for practitioners and are worth reading as you consider how much time you’re going to invest in your next AI-powered build process. None of them, however, is written for the person who has to defend a budget line item and associated number in a planning meeting.
The assumption in your 2027 plan
The AI enablement budgets being drafted this month were mostly sized against a belief that building had become nearly free and therefore the only scarce thing left was knowing what to build. Half of that held up. Producing a first working version did get dramatically cheaper. Finishing that idea to production-level code did not. The planning documents we see with our clients have no line item to account for this difference.
When building was expensive, the estimate did governance work implicitly. “Two engineers and a quarter” was a number serious enough to provoke a question, and powerful enough to see the weak ideas get filtered out regardless of product “sense.” Now the working demo shows up before anyone has decided it should exist, and the cost filter that used to delay this work until some validation had taken place no longer happens. Nothing has replaced it, which is why so many teams went into this year measurably faster and came out of it with a backlog of half-finished plausible things.
What actually costs money now
The expense did not disappear. It moved. It moved into reviewing what the tools produced, evaluating whether it holds under real conditions, rejecting most of it, and deciding which of the remaining things the organization is actually going to finish. That is labor, it is skilled labor, and across our work it lands almost entirely on the middle layer of the organization, on the product managers and team leads who were handed a faster pipeline but no additional capacity to judge what comes out of it.
So when the 2027 budget gets written, three items belong next to every initiative in the plan.
Who decides this is finished, and on what evidence?
What the team stops doing when drafts start arriving faster than anyone can review them?
What it costs to say no nine times in order to say yes once? (given that saying no now requires someone to look carefully at nine plausible working things instead of nine estimates.)
Those three answers are the costs of AI enablement. Everyone in the field now agrees that judgment is the scarce resource, but agreement is not a budget. A judgment capability that is not scheduled, staffed and paid for is optional at best.
This is most of what we end up working on with teams these days. The companies we see arrive wanting to talk about tooling and less about the impact of the new tools on the working process and ultimately the capability to ship high-quality, completed products to customers. If that is the conversation your team needs to have before the plan is locked, our upcoming workshops for product managers, on strategy, OKRs, discovery, and building AI products, are listed and open at ti.to/sense-respond-learning. (Use code SRFALL26 for 15% off across the board).
Whatever you decide to fund, make sure it includes the “deciding work” as a cost. The teams that finish well next year will be the ones whose budget acknowledged that judgment takes time, and then put that time into the plan.