For most of business history, execution was expensive.
Analysis required people. Content required people. Software required engineers. Coordination required meetings, managers and process. Turning an idea into something real consumed meaningful time, capital and organizational capacity.
That cost shaped how companies were designed.
AI is changing the equation faster than most operating models are changing with it.
Stanford’s 2026 AI Index reports that AI was used by 88% of surveyed organizations in 2025, while AI-agent deployment remained in the single digits across nearly all business functions. That gap matters: adoption is moving faster than operating-model redesign.
I think that gap is where the more important story sits.
A company can put AI inside the existing work without changing the work itself. It can add copilots, generate more drafts, summarize more meetings, produce more code and automate more isolated tasks. All of that can be useful.
But it does not answer the harder question:
Cheaper execution changes the bottleneck.
When production capacity is scarce, producing more is valuable. When production becomes abundant, the constraint moves.
The organization can suddenly generate more analysis than executives can absorb, more ideas than teams can pursue, more campaigns than markets need, and more software than anyone should maintain.
The bottleneck starts moving upward toward judgment.
Which problem deserves attention? Which signal matters? Which initiative should stop? Which capability is strategically important enough to own? Where should human accountability remain? What should leadership fund before the evidence is complete?
In a large field study of 5,179 customer-support agents, researchers found generative AI increased productivity by about 14% on average, with much larger gains among novice and lower-skilled workers. The evidence suggests part of the effect came from transmitting practices associated with stronger performers to less experienced workers.
Productivity measured as issues resolved per hour. The comparison matters because the gains were not evenly distributed; the strongest effect was among workers with less prior experience.
That is interesting to me for a reason beyond productivity.
If some forms of experience can be encoded and delivered at the point of work, then the value of expertise changes. Routine expertise becomes easier to distribute. The people who can define the system, recognize exceptions, make tradeoffs and update the logic become more important.
AI will not make judgment less valuable. It will expose the difference between repeatable expertise and real judgment more clearly.
Capability ownership needs a reset.
For years, many companies framed capability decisions around a relatively simple choice: build internally or buy externally.
That is now too narrow.
A business can own a capability, build it, buy a platform, partner, outsource delivery, automate parts of it, augment a team with AI—or combine several of those approaches.
The choice should not be driven by ideology. It should be driven by where advantage and learning actually compound.
I look at strategic differentiation, economics, proprietary data, learning value, talent, speed, frequency and risk.
One of the more important lessons I have learned is that capabilities create more than output. They create institutional learning.
During a broader digital transformation, the real shift was not simply adding digital products. Acquisition, CRM, analytics, technology and revenue had to move from isolated initiatives into the core operating model. That work also included restructuring that produced approximately $2.8 million in annualized operating-cost reductions while the digital revenue mix moved from roughly 20% toward 70%.
That is a stronger example of capability design than “in-house versus agency.” The question was what the company needed to become good at in order for the business model itself to change.
AI should change the workflow, not just accelerate it.
A lot of AI adoption still begins with a task-level question: how can we make this existing step faster?
That is a reasonable place to start. It is not where I think the larger value ends.
Research from Harvard and BCG describes AI performance as a “jagged frontier”: on some knowledge-work tasks AI dramatically improves speed and quality, while on other tasks outside that frontier it can make performance worse. Stanford’s 2026 AI Index shows the same unevenness at a broader level—agents have advanced rapidly, but still fail roughly one in three attempts on structured computer-use benchmarks.
That is why I do not believe the right design question is “human or AI?”
The better question is: what combination of people, systems and AI gives the business the best economics without losing accountability, learning or control?
OECD research on AI adoption notes that managers often treat AI as plug-and-play technology, even though meaningful adoption can require company-level changes across business operations and departments. Earlier NBER research on AI adoption makes a similar point: when tasks are interdependent, the value of AI can depend on redesigning the system rather than optimizing isolated tasks.
That distinction matters.
Automating a weak handoff preserves the handoff. Generating a report faster does not mean the report should still exist. Adding AI to a twenty-step workflow can make individual steps cheaper while leaving the underlying operating model untouched.
Sometimes the right answer is not to automate the work.
It is to redesign it.
The company that learns faster may have the stronger advantage.
As execution gets cheaper, I think the feedback loop matters more.
Can the organization see what happened? Can it separate activity from signal? Does the result change the next decision? Does customer behavior change the next allocation? Does the knowledge stay inside the business?
That loop is not an AI system. It is an operating system.
AI simply changes the economics enough that weak learning systems become easier to see.
What becomes more valuable?
My current answer is four things.
Judgment. Choosing the right problem, deciding what not to do and knowing when the evidence is sufficient to act.
Allocation. Moving capital, talent and executive attention toward the constraint that matters most.
Capability design. Deciding what the business should own, build, buy, partner, outsource, automate or augment.
Learning. Building an operating model where faster execution produces better future decisions instead of simply producing more activity.
What I may be wrong about.
It is possible companies absorb AI the same way they absorbed many earlier technologies: meaningful productivity gains, some role redesign and far less structural change than futurists predict.
Large organizations are durable for a reason. Functional specialization solves real problems. Hierarchies coordinate risk and accountability. Not every lower coordination cost should lead to a flatter company.
What makes me think this wave could be different is the combination of three things: the cost of producing knowledge work is falling, software can increasingly participate in multi-step workflows, and the tools are beginning to encode and distribute parts of organizational expertise.
I do not think that guarantees a new organizational form.
I do think it makes “we have always done the work this way” a weaker argument every year.
Questions I think leadership teams should be asking now.
Evidence & sources
I use external research to pressure-test the argument, not as a substitute for the argument. The interpretation above is mine.