A year ago, most AI conversations I had with leadership teams started in the same place: “is the technology really good enough?” Good enough to trust with real work, good enough to put anywhere near customers, good enough to justify the spend?
In the last six months, that question has quietened. Correctly configured, model capability is now well past the bar for a large share of everyday knowledge work, and it’s still rising. Fast. Yet for many, especially across the mid-market, the value hasn’t followed. Glean’s Work AI Index this year found 87% of employees using AI while only 13% of organisations could say their performance had improved.
If capability had been the constraint, those two numbers would move together, but they don’t. From what I see across organisations, three other things now gate the value. Economics, Permission and Absorption.
What will this cost at scale?
The subsidised era of AI is ending. The headline tools have been cheap because investors were covering part of the bill, but that arrangement is winding down at the same time as usage patterns get more expensive. When a company moves from chat (a person asking questions) to agents (systems doing multi-step work autonomously), consumption jumps: an agentic task consumes five to thirty times the tokens of a chatbot query, and usage-based pricing means the bill rises with it. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, with escalating cost a leading reason.
There’s a second cost that fewer plans anticipate. The platforms that hold your data (your CRM, your ERP, your ticketing system) are starting to meter the access that agents need to it. SAP began enforcing an API policy in June that routes third-party agents through a metered gateway (Forrester’s advice to CIOs: push back), and Workday and Snowflake now charge for agent activity through credit systems. You can own the valuable asset and still pay rent on the route to it.
The consequence is unglamorous. AI now needs the discipline that cloud computing needed a decade ago: budget it as a running cost rather than a one-off project, route work to the cheapest model that can do the job, know the unit economics before you scale, and treat access to your own data as a commercial term in itself. Many organisations are yet to master this (it doesn’t help that it sits awkwardly across teams: the ones doing the work, IT and Finance), but it is fast becoming an operational necessity.
What is our freedom to operate?
For the first time since this wave began, what governments and communities will allow is a live variable. June made this a concrete reality: a US export-control order suspended Anthropic’s newest model days after launch, for every customer worldwide. Firms that had moved work onto it fell back overnight (Stripe halted a codebase migration; Mozilla paused a security review); access returned three weeks later when the order was withdrawn. Meanwhile Europe’s AI Act timetable has moved more than once this year, and the two regimes are diverging, which matters to any business operating in both. Further down the stack, the data centres all of this runs on need local consent to get built, and communities have started withholding it.
None of this is a reason for alarm, but it is a reason for planning. The lesson many took from June is that sovereignty, optionality and fallback planning are now table stakes for operating AI at scale: a fallback model you have actually tested, and a clear view of whose decisions, in which jurisdiction, your capability depends on. Permission has become a variable to manage, like any other operating condition, and needs to be planned for, not assumed.
Can we absorb it?
The third gate is the one company leadership controls most directly: whether the organisation can actually absorb the capability. The 87/13 gap from earlier mostly sits here. The common pattern is that tools get deployed but work doesn’t get redesigned. Often this means every AI output needs checking, processes need babysitting (6.4 hours a week of it), and the checking of outcomes lands on people as unpaid extra work. In my experience, companies in that position end up slightly slower, at higher cost, with better-looking documents but grumpier staff.
The companies we see getting through this gate do a few things differently. They pick a small number of workflows and rebuild them properly rather than sprinkling tools everywhere. They decide who checks what, and they codify what good looks like, so the checking lessens with every pass. And they make sure that what one part of the business learns is available to the rest of it, rather than living in one person’s chat history. Beneath all of it sits the deepest part of this gate, the data layer the whole thing runs on, the learning loop. That one deserves its own piece (coming soon…)
Almost none of that is technology work in the usual sense, which is why waiting for a better model doesn’t help. The next model will be absorbed, or not, by the same organisation, depending on how they reorganise around it.
What we take from this
The practical shift: we evaluate AI plans on the gates rather than on capability alone.
1. What will this cost at scale?
2. What is our true freedom to operate?
3. And can we absorb it?
We expect the advantage to go to the businesses that clear all three gates because the best model with none of the gates cleared is not durable.
The good news is where the leverage sits: economics and absorption are substantially within a leadership team’s control, and freedom to operate can increasingly be planned for. All of which is more than could be said when the constraint was the technology itself.
Things with AI change; self-learning and self-optimising processes are already being hinted at. For now, the binding constraints sit inside firms and around them, and most are within the leadership team’s gift to affect. The deepest of them, the learning loop your firm builds around AI, is where we’ll go next time.
