The AI Operating Model: Redesign Work Beyond the Tools

An AI operating model is the set of decisions about work design, decision rights, review, accountability, and learning that determine whether AI changes what an organization can actually deliver. Buying seats and running a training session is AI adoption — it changes what tools sit on people's desktops. It does not touch who is allowed to decide, who has to sign off, what counts as evidence of good work, or how anyone gets better at judgment over time. Those are the things that determine outcomes, and they don't change on their own.

What actually has to be redesigned

When producing a first draft of an answer becomes cheap, the scarce resource stops being production and becomes the set of conditions under which an answer deserves to be used. That is a different kind of work, and it lives in five places most technology rollouts never touch.

• Work design. Which steps in a process still require a person to originate the work, and which now only require a person to evaluate it? Those are different skills, different pacing, and often different people than the roles currently assigned.
• Decision rights. Who is allowed to accept, reject, or escalate AI-assisted work, and at what threshold? If that answer hasn't changed since before AI, the people closest to the work have gained speed with no matching authority to act on it.
• Review. A reviewer who checked five drafts a week cannot meaningfully check fifty. Review has to shift from reading everything to sampling with intent, checking the riskiest categories of output, and building automated checks for the rest.
• Accountability. When an AI-assisted recommendation turns out wrong, whose name is on it? If the answer is vague, people will quietly stop trusting the tool's output rather than clarify who owns the call — a problem that shows up first in performance conversations.
• Learning. Junior staff have historically built judgment by doing the slow, effortful version of the work themselves. If AI does that work for them, the organization needs a deliberate substitute, or it will run out of people capable of doing the senior job in ten years.

Products, services, and customer expectations will move too

Operating model redesign isn't only internal. Once turnaround time on routine requests drops from days to minutes, customers recalibrate what "reasonable" means, and competitors who haven't redesigned anything will look slow by comparison regardless of quality. Some offerings that used to be justified mainly by the labor required to produce them stop making economic sense at their old price. Others become viable for the first time because the marginal cost of a custom version approaches the cost of a standard one. None of this is optional to think about once adoption is underway — it happens whether or not leadership planned for it.

Signals the operating model has genuinely changed

Adoption metrics — logins, usage rates, self-reported time saved — tell you almost nothing about whether the operating model moved. Better signals look like this:

• Decision rights have visibly shifted: more calls are made by the person doing the work, fewer by escalation, and that shift is documented rather than informal.
• Review has been redesigned around risk, not volume — the highest-stakes categories get more scrutiny than before, and low-stakes categories get less, on purpose.
• Someone can answer, without hesitation, who is accountable when AI-assisted work is wrong.
• Junior staff have a defined path to building judgment that doesn't depend on doing the slow version of the work AI now does.
• Delivery timelines, not just task times, have moved. If throughput hasn't changed, the constraint wasn't touched.

This is the same argument the book makes about the shift in human contribution more broadly: as production gets automated, the durable work becomes deciding what's worth producing and under what conditions it can be trusted — the move from production to direction.

Questions for the executive team

• Which of our approval, review, and audit processes were designed for a volume and pace of work that no longer exists?
• If an AI-assisted decision goes wrong next quarter, can we name the accountable person today, before it happens?
• Where have we let people work faster without giving them any more authority to decide — and what queue is that speed piling up behind?
• How will the next generation of senior staff build the judgment that used to come from doing the work themselves?
• Has anything about what we offer customers, or how fast we offer it, actually changed — or has our internal process just gotten more crowded?

None of this requires a large program office or a new framework with a proprietary name. It requires leaders willing to change decision rights and review processes that have been settled for years, which is a harder and less comfortable job than approving a software purchase. Beyond Doing is written for that job. Read more about the book.