When I released my feature documentary Beer Wars, I struck a deal with three of the largest theater chains in the country to bring the film to 450 screens for a one-night event, broadcast live by satellite before streaming made that kind of release routine.
I received spreadsheets from the theater chains showing which theaters, which markets, and how many seats they expected to sell in each. I pushed back. Data from my promotional partners showed real demand building in specific cities, not spread evenly across their standard grid. I asked the chains to move seats out of theaters that wouldn’t fill and into the cities where we were about to sell out.
They wouldn’t budge. Their response was consistent: their experience running one-night events told them their numbers were right.
We sold out New York and Boston weeks before the event and turned away paying customers. In the suburbs, where their model had confidently assigned tickets, theaters sat empty. I had identified the mismatch, but their own history made them unwilling to reconsider.
I’ve caught this pattern more than once throughout my career: an authoritative source with a real track record, delivering an answer with total confidence, unwilling to adjust it even when given specific evidence that this situation was different. I usually recognized it right away. This time I didn’t. I was new to LLM discoverability and schema markup. As I rebuilt my online presence, I was using several AI models as an advisory board. I assumed they understood their own terrain better than I did. I kept looking for consensus among them instead of trusting my own read. Four models reached the same wrong conclusion about a technical problem on my own website. That felt like confirmation, until it wasn’t.
The theater chains had formal authority, and they exercised it openly. Authority drift is harder to see. Often, no one explicitly decides to give an AI system more authority. Its recommendation simply becomes the default because it arrives quickly, sounds certain, and is easier to accept than to challenge.
Most organizations have already asked one question: what can AI do? Far fewer have confronted the harder one: what should AI decide? In the rush to show results, the first gets answered and the second often gets skipped.
Those two questions don’t move together the way people assume. The percentage of a task AI performs doesn’t tell you how much decision authority people have allowed it to exercise. A system can do most of the visible analysis while a human still makes the final call. Or it can do very little of the visible work and still end up deciding the outcome, quietly, by shaping which options look credible enough to consider in the first place. Work allocation and decision authority are not the same variable. Treating them as if they move together is how drift happens without anyone noticing.
I give executives a simple exercise to find where this may already be happening inside their organization: take the ten most consequential decisions from the last quarter. For each one, ask honestly: who actually made the call? If the answer is unclear, or if the honest answer is something close to “the system decided,” that’s worth taking seriously.
Once leaders run that audit, the follow-up questions reveal what the list alone cannot. Was AI used as evidence, as a recommendation, or as the default? Was a human owner named before the decision was made? Could that person explain why the recommendation was accepted or rejected? Those questions don’t come from the audit itself. They’re what the audit is for.
This is exactly the gap The Human + AI Equation addresses. The Equation asks leaders to determine what stays human-led, what should be AI-augmented, and what can be fully automated. Authority drift is what happens when the work mix changes but the authority decision is not revisited. A change in what a system can do may justify a change in how the work gets done. It doesn’t automatically justify a change in who is allowed to decide.
AI capability is moving faster than most organizations’ systems of accountability. That gap won’t close by slowing the technology down. It closes when leaders get as deliberate about ownership as they are about adoption: who is responsible, who is allowed to override the recommendation, and who answers for it when the system is wrong. That responsibility doesn’t get easier as AI gets more capable. If anything, the human side of the equation becomes more valuable, not less.
This is the second in a series exploring The Human + AI Equation. Read Part 1: Why I Created The Human + AI Equation. Read Part 3: What Makes You More Valuable as AI Becomes More Capable.
Related Coverage
The Trust Problem That Could Doom Smartglasses and AI Adoption — Forbes
Complete This 10-Decision Audit to See Whether AI Is Secretly Running Your Startup — Inc.