Who Pays When the Model Is Wrong? Change Futurist | September 26, 2026
Ask who in your organization pays when a model makes the wrong call. If the honest answer is whoever clicked approve, you already know your override rate, whether or not anyone has measured it.
Four authors took up that question from different directions in pieces published over the last 48 hours. Adi Ignatius at Harvard Business Review previewed new work from Das Narayandas on why employees override AI systems that have already proven themselves. Buse Aras at Gartner forecast how few organizations will let AI make even a tenth of their planning decisions by 2030. Marco Steecker, also at Gartner, named low AI literacy as the leading barrier finance teams face. And Josh Stephenson at Marketing Week reported Ipsos findings that consumers want more say over who benefits from AI.
Employees, planners, finance teams, and customers all run the same calculation. They rely on AI in proportion to who absorbs the cost of a wrong answer.
On Thursday I argued that someone has to own the decision about what a team stops producing. This edition follows that owner into the moment a model's recommendation lands on someone's screen. In most marketing organizations, the person who acts on the recommendation and the person who carries the consequence are two different people, and nobody has reconciled them.
Move the cost of a wrong AI call off the person who followed it
In the HBR Executive Agenda published September 25, Ignatius gave subscribers early access to an article by Narayandas, a Harvard Business School professor. Narayandas argues that employees often override proven AI systems to protect themselves, because their careers absorb the cost when the AI is wrong. He proposes a set of fixes leaders can put in place.
Read that as a compensation design problem. An employee who follows the model and misses takes the hit in a quarterly review. An employee who overrides the model and misses can say they used their judgment, which still reads as defensible in most performance conversations. Overriding is the rational career move, and your people are rational.
Nobody has ever landed on a performance plan for trusting their gut.
In the AI capability framework I use, this sits in the Trust layer, in the form that layer takes under Augmentation: calibration. The question there is whether the person knows when to trust the contribution and when to override it. I have usually described the failure mode as silent over-reliance. Narayandas documents the opposite failure, and incentive design produces it just as reliably.
Calibration only works when an override gets judged on evidence. That requires the cost of a sanctioned model call to sit with whoever sanctioned the model for that decision. The principle I am building into Book 2 holds that humans stay accountable for direction even when the analysis runs fully automated. Accountability for direction belongs to the person who set it, and the analyst at the end of the workflow almost never set it. An AI governance board charter is the right place to write this down, alongside decision rights, exception handling, and documented risk acceptance.
It reminds me of what John Kim, CEO of Delight.ai, said when I interviewed him on The Agile Brand podcast: "I wouldn't trust an AI simply because it says it made the right decision. I would trust a system that's designed to surface potential mistakes and continuously improve based on evidence."
Kim's test depends on errors surfacing. In most marketing organizations, the people closest to the work have every reason to keep them buried.
Commentary: Narayandas relocates the adoption problem from training to consequences, which gives marketing leaders a lever they control directly: who owns the downside of a sanctioned model call.
Classify every decision before you hand any of it to a model
Aras published a Gartner prediction on September 24 that only 5 percent of organizations implementing supply chain planning automation will make at least 10 percent of their planning decisions autonomously by 2030. Gartner surveyed 243 senior leaders at organizations with at least $500 million in revenue. Eighty-three percent had spent at least $3 million automating planning, and 51 percent had spent between $3 million and $10 million.
Aras puts the gap plainly: "investment alone does not create AI readiness." Aras names decision ownership alongside data quality, workforce capability, and architecture as the work organizations have to finish before they entrust more decisions to AI.
The Gartner recommendations translate to marketing almost word for word. Classify each priority decision as strategic, tactical, or operational. Decide whether AI should support, augment, or automate it based on value, complexity, risk, and the need for human judgment. Then measure whether decisions improve, tracking whether planners use the new capabilities as intended, cut manual workarounds, and stop reverting to legacy tools.
Reverting to legacy tools is Narayandas's override under a different job title.
Run the classification on your own function. Send-time optimization for a lifecycle program is operational and belongs in the automate column. Discount depth for a win-back offer is tactical and belongs in augment, where a model recommends and a named person owns the range. Brand positioning stays in support. Each row needs exactly one accountable owner, which is the same discipline a RACI matrix enforces, applied to decision classes.
I wrote in Coherence at Velocity that AI and automation make ambiguous decision rights more expensive. Aras's respondents supplied the invoice: millions spent per organization, and a single-digit share expecting meaningful autonomy by the end of the decade. Aras also describes planners moving from data administrators to plan orchestrators. In my new book, Marketing Operations 3.0, I put the same shift this way: humans define goals, guardrails, exceptions, and accountability, and agents handle scale and speed inside them.
Commentary: Aras gives marketing operations a ready-made taxonomy for AI decisions and a metric, reversion to legacy tools, that any team can start tracking this quarter.
Train people to read the model before you ask them to trust it
Steecker published findings on September 24 showing that CFOs must take a more disciplined approach to finance AI investment. Gartner surveyed 160 senior finance leaders from January through April 2026. Data extraction, accounts payable and receivable automation, and report creation generally return value within nine to ten months. Data management, insight generation, and forecasting take longer to mature.
Steecker warns CFOs not to let the appeal of quick productivity wins push aside the slower use cases that improve decision making and support revenue growth. Steecker also reports a shift in the main barrier. Talent used to top the list, and agentic coding tools have eased that constraint. "Low AI literacy is now the most significant barrier finance leaders must address," Steecker says. The remedy is practical: project assignments, sandbox experimentation, and short on-the-job activities.
Literacy is what lets an override rest on evidence. A marketer who can read a model's confidence, its training window, and its known blind spots overrides when the evidence says to. A marketer who cannot read those things overrides out of fear or approves out of deference, and both habits cost you money.
It reminds me of what Patricia Rollins, VP of Marketing at Thryv, said when I interviewed her on The Agile Brand podcast: "The biggest mistake small businesses can make is blindly trusting AI. Use it only for tasks you already understand, otherwise you run the risk of moving faster in the wrong direction."
Rollins was speaking to small business owners, and the advice scales cleanly to enterprise teams. Understanding the task is the precondition for judging the model that performs it.
One more note from Steecker's findings. Gartner analysts spent this week telling your CFO to cut underperforming AI initiatives. Bring your own evidence on which marketing uses return value before finance builds that list for you.
Commentary: Steecker's time-to-value data warns marketing leaders that the highest-value AI uses, including forecasting and next best action, sit in the slow-return group, so defend them explicitly in 2027 planning.
Give customers the control you want your own teams to have
Stephenson reported on September 24 on the latest Ipsos Global Trends study. Ipsos researchers found that 56 percent of respondents across 38 markets believe AI has a positive impact on the world, with 35 percent disagreeing. In the United States, one-third of respondents see a positive impact. In the UK, the net balance of AI views sits at minus 14. The skepticism concentrates among younger consumers: agreement that AI has a positive impact fell seven points among 16- to 24-year-olds and four points among 25- to 34-year-olds, while older groups held steady.
Ipsos researchers conclude that consumers want more control over who benefits from AI, and they identify agency as a valuable currency for modern consumers. They also flag a framing problem. A CFO hears productivity claims as good news. The average consumer hears them as a reason to pile on more work with little benefit to their own life.
Customers run the same calculation your employees run. When someone else captures the benefit of an AI decision and they absorb the cost, they override it, by abandoning the chatbot, rejecting the recommendation, or leaving.
The fifth Agile Brand principle asks you to focus on the relationship over the individual transaction. In an AI-mediated experience, the relationship depends on the customer's ability to adjust, decline, or escalate what the system decided on their behalf. A recommendation engine with no visible way to say no feels like surveillance to someone who already doubts who benefits.
Commentary: Stephenson's report gives marketing leaders a consumer-side reason to design visible controls into every AI-facing experience, and a warning against selling AI to customers with the same productivity language used in the budget deck.
Price the overrides in a workflow you run every week
A $900 million specialty retailer runs a win-back program that sends 400,000 discount offers a quarter to lapsed customers. The team licensed a next-best-offer model eighteen months ago. For one large segment, the model recommends a 10 percent discount. The campaign managers' habit, formed long before the model arrived, is 20 percent.
The managers override 62 percent of the model's recommendations back to 20 percent. Nobody told them to. Their reasoning follows Narayandas exactly. When a win-back campaign misses its reactivation target, the campaign manager's quarterly review takes the hit. When margin erodes from over-discounting, the loss rolls into a finance line nobody traces back to a specific campaign.
Price it. The overrides cover 248,000 offers a quarter. At an 8 percent redemption rate and an $85 average order, that produces 19,840 orders carrying an extra ten points of discount, or $168,640 a quarter in margin given away. Across a year, the habit costs about $675,000. The team's own holdout tests put reactivation within three-tenths of a point at either discount level.
The repair involves no new software. The lifecycle director becomes the named owner of the discount-depth decision class and signs off on the model's recommended range. Campaign managers keep the right to override, and every override requires a one-line reason code. Misses inside the approved range go to the director's review, and the manager who followed the model faces no penalty for doing so. Override rates go into the weekly operating report next to reactivation.
Within two months, the override rate falls to 15 percent. Annual margin leakage drops to roughly $163,000, a recovery of about $510,000.
The reason codes deliver the second payoff. A cluster of overrides in one product category shares the same explanation: a price increase three weeks earlier that the model's training window missed. Those overrides were correct. The team retrains the model on current pricing, and the overrides that remain start working as a defect report.
The retailer kept its model and its team, and changed who pays for a miss.
Take these four steps into Q4 planning
Measure your override rate by decision class. Pick the three highest-volume decisions where a model already makes a recommendation, and record how often people accept, modify, or reverse it. You cannot manage a number you have never seen.
Move the cost of a sanctioned model call to whoever sanctioned it. Name one accountable owner per decision class, write down the approved range, and make clear in performance reviews that following the model inside that range carries no personal penalty.
Require a reason code on every override. Overrides with reasons become model defect reports and training data. Overrides without reasons stay what they are today, a quiet tax.
Build the customer's override into every AI-facing experience. Give customers a visible way to adjust, decline, or reach a person, and track how often they use it as a trust metric.
I have had more than 900 conversations with senior enterprise leaders on The Agile Brand podcast, and when guests describe an AI program that stuck, the story often includes a moment when a senior leader stood in front of the team after a model-driven call went wrong and said the miss was theirs. The rollout plan and the vendor shortlist rarely come up in those stories.
Find out who pays when the model is wrong in your three highest-volume decisions. If the answer is the person closest to the keyboard, move it before you license another model.
Featured this cycle:
Why Your Employees Override AI. Adi Ignatius, previewing new work by Das Narayandas, Harvard Business Review, September 25, 2026
Gartner Predicts Only 5% of Organizations Will Make At Least 10% of Supply Chain Planning Decisions Autonomously by 2030. Buse Aras, Gartner, September 24, 2026
Gartner Says CFOs Must Take a More Disciplined Approach to Finance AI Investment. Marco Steecker, Gartner, September 24, 2026
Personal consumer optimism holding firm against global pressures, study finds. Josh Stephenson, Marketing Week, September 24, 2026