The Best AI Operating Model Still Leaves the Lights On for People

Full automation is a clean idea. Real organizations are not clean systems.

Customers bring unusual circumstances. Data arrives incomplete. Policies conflict. Markets change faster than a workflow. The closer a process gets to consequential human needs, the more often an exception matters.

At Indeed FutureWorks 2026, Wharton professor Ethan Mollick contrasted the “dark factory,” a fully autonomous operation with no people inside, with a more useful “twilight factory.” In the twilight model, AI performs much of the routine production while people remain present for variance, specialized expertise, and final sign-off.

The metaphor offers a practical alternative to arguing whether AI will replace or assist workers in the abstract. The better question is where autonomy works, where exceptions concentrate, and where a person must remain accountable.

Scale changes the work around the work

AI agents are not only producing content. They are becoming customers, researchers, and intermediaries. Jessica Sibley explained during “The Next Era of Work” that TIME now has more AI agents reading its content than human readers. That shift affects how material is written, published, discovered, and sold.

Marc Benioff described another form of scale at Salesforce, where agents were resolving millions of support inquiries autonomously. The human role did not disappear. It moved toward complex cases and the design of the system handling the rest.

These examples show why adding an AI tool to an old process is not enough. When machine volume becomes significant, upstream and downstream work changes too. Someone must decide what the system can do, how performance is observed, and which situations require interruption.

Exceptions are where human value concentrates

Automation performs best when inputs, rules, and acceptable outputs are stable. Human expertise becomes most valuable where those conditions break.

A customer may be technically ineligible for a remedy but facing circumstances the policy did not anticipate. A generated analysis may be statistically sound but conflict with an important market reality. A candidate may look unconventional because the system has never seen a comparable path.

These are not merely edge cases to eliminate. They often contain strategic information. Repeated exceptions can reveal a changing customer need, a flawed policy, or an opportunity the standard process cannot see.

A twilight factory keeps people close enough to recognize those patterns. It also preserves a route for affected people to reach someone capable of exercising judgment.

Human sign-off has to be substantive

A person technically present at the end of a workflow does not guarantee oversight. If they lack time, evidence, or authority to disagree, approval becomes ceremonial.

Meaningful sign-off requires clarity about what the reviewer owns. They need to understand the domain, see the inputs and reasoning relevant to the decision, and know when escalation is expected. The organization must accept that careful review may sometimes slow the process.

The rigor should match the consequence. A machine can autonomously categorize low-risk internal requests under light monitoring. A decision affecting employment, health, safety, or financial access needs stronger review and a clear appeals path.

So what does a twilight operating model require?

Begin by mapping the workflow at the level of decisions. Identify stable, repeatable steps where autonomy adds value. Then identify points with high variance, incomplete information, serious consequences, or a need for empathy and negotiation. Those are candidates for human ownership.

Define the handoff in both directions. The system needs criteria for escalating an exception, and people need a way to correct the system so the same failure becomes visible. Track not only automation rate, but override quality, unresolved exceptions, customer outcomes, and new patterns found through human review.

Workforce design should follow the same map. If people will handle the hardest cases, their jobs require deeper expertise and stronger support, not simply fewer colleagues. Training must include how the automated process works and where it is likely to fail. Career paths need to preserve routes into that expertise.

The twilight factory is not a compromise made because technology is temporarily imperfect. It is a recognition that efficient repetition and responsible judgment are different kinds of work. The strongest operating model gives each to the system best equipped to perform it, then designs the relationship between them with care.