The productivity promise around AI is real enough to shape capital allocation, but too conditional to justify a technology-only business case. The OECD’s January 2026 release shows adoption expanding across people and firms while remaining uneven by organization size. Access is spreading; advantage will depend on how well organizations redesign the work around it.
OECD’s 2026 work on AI and skills points to complementary investments in skills, data, process design, and organizational adaptation as conditions for productivity gains to appear. That is a serious signal, not a guaranteed return. A business case that cannot describe the changed workflow, the new control, and the evidence threshold is still a technology hope.
The model is not the unit of value
An AI tool can produce a faster answer and still make the enterprise slower. A drafting assistant may increase review load. A forecast may arrive without the authority to act on it. A coding tool may accelerate releases while increasing the number of insecure dependencies. A clinical assistant may surface more information while making the point of care harder to navigate.
The unit of value is the workflow: the sequence of signal, judgment, action, handoff, and feedback that produces an outcome. AI changes one or more steps in that sequence. Leaders therefore need to ask four questions before asking which model to use:
- What decision or service promise should improve?
- Where does human judgment remain essential, and what context does it need?
- What new exception, review, security, or training work will the capability create?
- How will the organization know whether quality improved rather than activity simply increasing?
This reframes the business case. Instead of claiming that AI will save a percentage of time, a team can describe the operating change: fewer manual reconciliations, earlier exception detection, shorter cycle time at a controlled error rate, or more capacity for a high-value decision.
The adoption gap is a design signal
The fastest adopters are not necessarily the organizations with the most sophisticated models. They are often the ones with a clear use case, a visible owner, enough data quality to begin, and a workflow where people can learn without carrying all the risk alone. OECD research on AI and skills makes the same point from the workforce side: training is associated with more positive outcomes, but skills are only useful when the work gives people a place to apply them.
That is why a serious AI portfolio should include workflow redesign and capability transfer as explicitly as it includes software. Each use case needs a service owner, a risk boundary, a measurement plan, and a path for the people closest to the work to improve the system. Otherwise, the organization creates a collection of impressive tools with no coherent operating advantage.
A practical diagnostic
For each proposed use case, map the current state in one page. Name the user, the decision, the source signals, the points of delay, the current control, and the consequence of error. Then sketch the future state with the same discipline. If the future picture contains a model but not a new role, measure, exception path, or management rhythm, it is not yet an operating design.
Track leading signals before financial outcomes arrive: time to a trusted answer, percentage of cases requiring rework, rate of human overrides, exception resolution time, adoption by the intended users, and the quality of the evidence available to a manager. These measures do not replace business results. They reveal whether the system is becoming capable of producing them.
What would change our mind?
AI will not create material productivity in every workflow. The economics may fail when the task is too rare, the data too weak, the review burden too high, or the risk of a wrong output too consequential. The right response is not to force adoption. It is to make the stop condition explicit and move investment toward the workflows where human judgment and machine capability genuinely reinforce each other.
The strategic question for leaders is therefore not whether their organization has AI. It is whether the organization can repeatedly turn a promising capability into a better way of working. That is an operating model question, and it is where the productivity case becomes real.