Three months later, sales were down 40%.

This story is becoming common. And the problem it reveals is not technical.

Most executive teams today are asking the right questions about AI: What can it automate? Where does it generate productivity? Which functions can it augment? These are important questions. But they miss the more consequential one: What is your AI already deciding on your behalf and do those decisions reflect the company you want to be? 

The real risk of AI pricing is not that machines will set the wrong price. It is that companies will accidentally turn pricing policy into an optimization problem and discover too late that the algorithm has made strategic, ethical, and political choices the leadership team never debated.

The shift that few executives have fully registered

For most of the past decade, AI in business operated as a copilot. It recommended, flagged, and ranked options. Humans decided. This architecture, however imperfect, preserved a clear line of accountability.

That architecture has quietly changed. Agentic AI systems, the next generation now being deployed across pricing, customer management, and revenue optimization, don’t recommend. They act. They pursue objectives autonomously, remember past interactions, learn from outcomes, and adjust continuously without requiring human validation at each step.

In pricing, this shift is already visible. Salesforce has made the first move. With Agentforce, it introduced action-based pricing: Customers can pay not just for access to software, but for work an AI agent completes. The unit of value is moving from the human seat to the machine-executed action. This is not an incremental optimization. It is a fundamental change in what gets decided, by whom, and at what speed.

McKinsey’s latest AI survey shows that while AI use is now widespread, most organizations remain in experimentation or pilot mode, and only a minority are scaling agentic AI. That makes governance even more urgent: Companies are giving machines more autonomy before they have mature routines for deciding when human validation is required.

Three ways AI is making pricing decisions you may be missing

First, velocity. Pricing agents now operate at speeds that outpace any human review cycle. Prices can shift multiple times per day based on signals (demand, inventory, competitor behavior, time of day) that no human team could process in real time. The decisions are individually defensible. But their cumulative effect on customer trust, brand positioning, and long-term margins is rarely audited at the same frequency.

Second, personalization. Agentic systems can create, in practice, a different market for each customer. Two people visiting the same platform at the same moment may see different prices, different bundles, different urgency signals calibrated to what the system has inferred about their willingness to pay. The revenue upside is real. So is the exposure: when customers compare notes, the gap between “personalized offer” and “discriminatory pricing” can collapse overnight.

Third, memory. Unlike earlier AI tools, agentic systems build a cumulative model of each customer relationship. Past decisions shape future ones. A pricing agent that learned six months ago that a particular client segment accepts 15% premiums will keep applying that logic, refining it, extending it, unless someone explicitly changes the objective. No individual decision triggers a review. The drift is invisible until it isn’t.

The real risk isn’t the wrong price. It’s the unexamined decision

Ticketmaster is the cautionary tale, and shows it is not new. A U.S. Senate investigation found that between 2019 and 2022, the number of dynamically priced tickets sold by Ticketmaster for North American concerts rose by more than 700%. The system did not fail because it could not optimize. It failed because optimization collided with consumers’ sense of fairness.

The subcommittee opened an inquiry into Ticketmaster’s role in rising ticket prices. The algorithm hadn’t malfunctioned. It had performed exactly as designed—and no one had defined how far was too far.

The commercial cost of perceived unfairness is not theoretical. Bain has long argued that online retailers need to offer customers a fair, not necessarily the lowest, price, and pricing research shows that fairness perceptions and loyalty are deeply intertwined. Yet the companies building the most sophisticated pricing engines are often the least equipped to explain, defend, or correct what those engines are doing. The precision of the machine outpaces the governance of the institution.

The failure mode is almost never a rogue algorithm. It is an organization that delegated decisions without defining their limits.

A test for what should and shouldn’t be delegated

Not all pricing decisions carry the same institutional weight. A useful distinction separates transactional decisions—frequent, local, reversible—with bounded impact, from institutional decisions, those that touch brand promise, perceived fairness, long-term relationships, or the social license to operate.

Transactional decisions are ideal candidates for full automation. Adjusting prices based on real-time inventory, local demand signals, or channel dynamics falls squarely in this category. Speed and consistency genuinely outperform human judgment here.

Institutional decisions are different. They are the ones where getting the optimization right but the relationship wrong produces outcomes like Ticketmaster. For these, a useful test is what we call the 4Rs. Before delegating any pricing decision to an AI agent, executives should ask four questions: Is it reversible quickly without creating a toxic precedent? Is it sufficiently routine to be standardized? Is it rule-bound by explicit internal rules? And is any error repairable without a public crisis? If the answer to any of those questions is no, the decision should not be fully delegated. When a single one of these conditions fails, full delegation to the machine becomes a governance risk, not a productivity gain.

The deeper question is structural. Every agentic pricing system encodes a vision: of what matters, what can vary, what is negotiable, what is not. Companies that have not made those choices explicitly have made them implicitly by default, through the objective functions their data scientists optimized for last quarter.

The question every CEO should be able to answer

The executives we have spoken with who have navigated this transition most successfully share one characteristic: they treat AI pricing governance not as an IT question but as a leadership question. They can articulate, without consulting a dashboard, what their pricing agents are permitted to do on their own, and what requires a human decision.

They have, in effect, written a pricing constitution: the rules that will produce all future prices, before those prices are produced. Such a constitution should define which variables an agent may use and which ones are off-limits, how much price variation is acceptable across customer segments, when a human must approve a change, and which long-term metrics like trust, churn, repeat purchase or complaints can override short-term margin.

The question is not whether your AI will make pricing decisions. It already is. The question is whether the rules it is following are the ones you would stand behind, this in front of your customers, your board, and, if necessary, a congressional hearing.

If you cannot answer that today, the time to define those rules is before the next Monday morning notification arrives.

Francois Candelon is a partner at private equity firm Seven2 and executive fellow at the HBS AI Institute (formerly known as the D^3 Institute). Read other Fortune columns by François Candelon.

Paul-Louis Andres is a director at Seven2.

Augustin Manchon is a professor at HEC and Paris-Dauphine and Pricing Strategy and Governance advisor to CEOs at Manchon & Company.