Your AI Has an Objective Function. Did You Write It or Did Your Fear?

Your AI is working. The real story is who it’s working for.

Starbucks shut down its North American AI inventory system in May after nine months across 11,000 stores. It was sophisticated enough to count milk eight times a day. Not sophisticated enough to tell one milk from another.

McDonald’s had already pulled its IBM voice AI program after a series of widely reported failures, including orders like nine sweet teas on a single ticket. Two weeks after Starbucks pulled its program, McDonald’s announced plans for a global AI operating system across 43,000 restaurants. In announcing McDonald’s Next, the CEO framed the objective around the customer: we cannot ask our customers to choose between hospitality and speed. Whether the deployment architecture reflects that brief is the question this article is asking of every AI mandate in the industry.

Same industry. Same technology era. Same tools. The difference isn’t the AI. It’s the objective function.

In the U.S., AI keeps getting pointed at the operation: cut labor, plug gaps, protect the system. And it does. Perfectly. The customer becomes the tradeoff.

This is the part most organizations never examine. AI doesn’t “lean” toward efficiency -- it is aimed there. It doesn’t “accidentally” prioritize cost over customer value -- it is told to. Every AI system is an optimization engine. It executes against the objective function it is trained toward, at scale, without judgment, without pause. When the customer isn’t the target, the customer becomes the compromise.

Japan’s consumer market has always enforced a different standard. They call it Anshin: the assurance of safety and trust as a felt human experience, not a compliance metric. Japan’s Ministry of Economy, Trade and Industry named it explicitly in 2024. The market had been enforcing it long before anyone wrote it down.

Japan named it Anshin: the standard for a successful AI deployment is whether the person the system serves feels secure in its output.

American QSR brands spent the same year pulling systems that could not distinguish oat milk from whole milk.

The technology was equivalent. The objective function was not.

McDonald’s Ran the Same Experiment Twice. The Market Graded Both.

McDonald’s US built a technically defensible AI stack:

Supply Chain Control Tower 2.0, ingesting real-time data across 2,600 SKUs, 150 logistics partners, and 70,000 weekly supplier deliveries, replacing a legacy system that held $2.4 billion in static safety stock and still produced 2 million annual stock-out events
Ask Pickles, a retrieval-augmented generation model fine-tuned on 180,000 pages of operations manuals, delivering real-time procedural guidance to kitchen crew via headset in 32 languages
Edge computing infrastructure, built with Google Cloud, running predictive maintenance models on critical restaurant equipment

Every model worked. Every model was built to cut cost and improve throughput. The customer was never the target. The customer was the tradeoff. McDonald’s US same-store sales declined through 2024 and into 2025.

The pattern repeats in Japan, but the trajectory changes.

McDonald’s Japan deployed a self-order kiosk, designed by French manufacturer Acrelec and installed without UX localization, trained to maximize average transaction value. It required ten screen interactions to order a single coffee, withheld pricing until the final payment screen, and fired upsell prompts at the moment of payment. It did exactly what it was built to do.

Japanese customers pushed back immediately and publicly. The experience felt designed to extract rather than serve. A market that treats pricing transparency and frictionless service as baseline expectations does not forgive a system built to treat both as negotiable. Sales declined. The kiosk became a case study in what happens when the objective function is aimed at the operation’s revenue target rather than the customer’s experience of the transaction.

The kiosk was not a technology failure. It was a market verdict on the objective function.

McDonald’s Japan responded by building an app on entirely different terms. A cross-border product team set a single commercial objective: maximize customer return rate. The system ingests individual transaction history, loyalty signal data, and Japan’s local digital ecosystem inputs. It delivers transparent pricing at every step of the order flow, frictionless repeat-order pathways calibrated to individual purchase patterns, and zero pressure at payment.

The app did not create the growth streak. It earned the next phase of it.

Result: a 4.6-star App Store rating with millions of daily active users, operating income up 15.5%, and EPS up 44% year over year through Q1 2026.

Two deployments. One brand. One market. The objective function changed. The numbers followed.

Organizations running AI without that trajectory are not necessarily running the wrong platform. They are running the wrong objective function. That is a different problem with a different fix.

Starbucks Spent Millions Optimizing Inward. Japan Never Had To.

In May, Reuters confirmed Starbucks discontinued its North American AI inventory program. The story behind that decision has been building for years.

The system was built to count beverage inputs eight times more frequently than manual processes, targeting the milk shortages Niccol had identified as a measurable revenue drag. Sophisticated enough to count milk eight times a day. Not sophisticated enough to tell one milk from another.

Nine months. 11,000 stores. Pulled.

Deep Brew, Starbucks’ broader AI platform, runs inventory, labor scheduling, and store allocation across North America. The inventory program pulled in May was one application inside that stack. Both are well-built. Both were pointed at the same target: the operation’s problems, not the customer’s experience.

Every model was running correctly. The objective functions were aimed at the wrong thing.

North America reported comparable store sales declines of 2% to 8% across consecutive quarters through the same window.

Niccol said it plainly when he arrived: technology had been deployed to manage the organization’s internal complexity rather than generate customer value.

The comparison to Starbucks Japan is not direct. Japan operates under a licensed structure with a different cost base and consumer profile. The licensed model does not explain the menu philosophy, the hospitality standards, or the customer experience decisions made at market entry. But the contrast is instructive.

The International segment, which includes Japan, outperformed North America through every quarter of the decline.

Net income in Japan grew from $1.018 billion to $1.247 billion in a single fiscal year. A 22% increase. No AI remediation program required.

Starbucks Japan was built from market entry around Anshin.

The founding question was not what does this system need to run efficiently. It was what does this market require from this brand to sustain trust. That question drove every product and experience decision: portions calibrated to local consumption, a menu that centered Japanese consumer preferences rather than exporting Seattle’s, seasonal rotations that tracked Japanese cultural moments, and stores designed to Japanese hospitality standards rather than global brand templates. Starbucks Japan did not need AI to remediate its inventory. The operating model was built around the customer before the inventory problem had a chance to compound.

In many cases, the North American AI spend is not innovation capital. It is the cost of repairing an operating model that was never built around the customer. Anshin was not Japan’s AI strategy. It was Japan’s founding architecture. When AI arrived, the objective function was already there.

Yum! Brands: The Objective Function Is a Choice, Not a Geography

The cultural objection is worth naming directly because it is the most common exit ramp from this argument. Japan is Japan. The philosophy that produces Anshin runs deeper than a deployment brief. American organizations cannot acquire it by changing who writes the training objective. If that is true, the pattern above is interesting and nothing more.

Yum! Brands refutes it.

Yum! operates KFC, Taco Bell, Pizza Hut, and Habit Burger across 61,000-plus restaurants globally. 57% of revenue from the United States. American leadership. American quarterly reporting pressure.

Byte by Yum!, launched February 2025, was built on a customer behavior objective function from day one. The platform pulls consumer sentiment from social media and third-party review channels and runs it against store-level operational data.

Byte Coach puts specific commercial guidance in front of restaurant general managers based on what customers are actually responding to, not what the operation finds easiest to measure.

Voice AI, built on NVIDIA infrastructure, went from concept to 600 live restaurants in four months.

Over 200 million AI-generated customer communications ran in 2025, delivering up to five times more incremental sales than traditional marketing approaches.

The US results are specific. Taco Bell reached a record 41% digital mix driven by US loyalty growth. KFC US scaled Byte Coach to its domestic store base ahead of international rollout.

Q2 2025 systemwide: KFC digital sales up 22%. Taco Bell loyalty membership up 45% year over year. Digital sales $9 billion. Same-store sales up 2%. The digital and loyalty numbers are the signal. Same-store sales at 2% reflect an objective function that compounds over years, not quarters.

An American company, in American markets, with American shareholders and quarterly reporting pressure, built its AI around customer retention. The results followed.

The objective function is a choice. It is not a geography.

Why American Food Companies Keep Training the Wrong Model

The pattern is not hard to find. It is hard to admit.

Board mandates arrive before commercial outcomes are defined. The deployment architecture gets written by whoever holds the budget, typically operations or IT. The objective function defaults to what that function knows how to measure and defend in a quarterly review: cost, throughput, error rate, uptime.

Most executives recognize this dynamic the moment it is named. The CFO approving supply chain AI and the CMO who might have approved a customer lifetime value model are rarely in the same conversation before the training data is assembled. Operational efficiency is quantifiable in the current period. Customer retention plays out over years. Capital flows toward the metric the approver is accountable to.

In many organizations, leaders are managing that concern before they manage the strategic question of what the AI is actually being built to produce. The result is a mandate that gets translated into the most defensible use case rather than the most commercially valuable one.

Japan is running a different calculation.

Working-age population is falling toward 59 million by 2040. Food service is among the two sectors reporting the most acute labor shortfalls in Bank of Japan surveys. There is no workforce to protect from displacement.

The question is not whether AI will change how people work. It is whether AI can hold the standard of service as the people disappear.

That question produces Anshin. The American version has produced discontinued programs, viral failures, and a widening gap in commercial results between the brands asking it and the ones that are not.

The Finding Most AI Strategies Miss

The operating models winning with AI share one characteristic that has nothing to do with the technology they selected.

They knew what the AI was for before they built it.

Starbucks Japan and McDonald’s Japan were not built around Anshin because they had a better AI strategy. They were built around Anshin before AI existed as a commercial technology. This is Monozukuri at its core: the Japanese principle that making something well requires understanding what it is for before you build it. When AI arrived, the objective function was already written into the business. When the objective function is clear before deployment, the technology has a brief worth executing.

The American brands pulling programs are not failing at AI. They are succeeding at it. The objective functions were set to optimize the operation. The operation is being optimized. The customer is downstream.

This is not a technology problem. It is a leadership sequencing problem. It closes when the commercial objective function is owned by commercial leadership before the technology team inherits the brief. Which means the CTO, the CMO, and the CFO need to be in the same room before the first model is selected, not after the first program is pulled.

These are not complicated questions. They are uncomfortable ones. There is a difference.

Who holds commercial accountability for this deployment, and is their performance tied to customer lifetime value or to operational cost reduction? If it is the latter, the model will optimize correctly toward the wrong target.

Was the training objective defined before or after the deployment architecture was written? If the technology team set the objective function by default, the organization is running Anshin in reverse: the operation at the center, the customer downstream.

And the most clarifying question of all: if this program were discontinued in nine months, what would the postmortem say it was built to do? If that answer does not match what the board expects from this investment, the gap is already widening.

The brands that answered these questions correctly before deployment are not waiting for the $115 billion AI market to validate their approach. McDonald’s Japan, Starbucks Japan, and Yum! Brands are reporting the validation every quarter.

AI doesn’t drift. It obeys. The objective function is a choice. Most companies don’t realize the choice was already made for them.

The customer and the technology team are rarely in the same room when the objective function is written. That is where the gap begins.

Where that choice gets made -- and who you trust to make it -- may be the most important leadership question you face this year.