What the McDonald's CEO Is Actually Saying About AI in 2025 (and What He Isn't)

I spent Thursday evening doing the thing I always do six trading days before a megacap QSR print: I re-read the CEO’s most recent public comments and built a two-by-two on a legal pad. On one axis: how aggressive is the AI roadmap. On the other: how big is the downside if a piece of it goes wrong on a busy Saturday morning. McDonald’s sits in a corner of that grid that almost no one on the sell-side is willing to underline in their preview notes. The corner is “deliberately slow, and correctly so.”

Chris Kempczinski spoke at an industry conference late last week. The trade-press recap — Restaurant Technology News’ week-of-April-27 wrap is the cleanest version — pulled out his three forward-looking calls for the rest of 2025: more protein on the menu, more AI inside the system, and more innovative sauces. The headline writers, predictably, grabbed the AI line. What they left on the cutting-room floor is the part that actually matters for valuation: how Kempczinski framed the build, who he named as partners, and what he very pointedly did not commit to.

McDonald’s is the largest single buyer of nearly every input it touches. When it leans into a technology, the entire QSR supply chain has to make a capex decision within a quarter. So when its CEO talks about AI, the right question is never “is he excited?” It is “what is the pace, and what is the dependency stack?” The pace, on the evidence of the last ten months, is slow. The dependency stack runs through Google Cloud for foundation models and analytics and Cognizant for the systems-integration layer. Neither is a moonshot partner. Both are the boring, defensible, enterprise-grade choice.

My base case going into the May 1 Q1 call: AI gets one or two paragraphs in the prepared remarks, zero hard dollar figures attached to it, and a Q&A answer that uses the words “test,” “learn,” and “thoughtful” more than once. That is the right script for this management team to read. The market, I suspect, will not love it. The market will be wrong.

What Kempczinski actually said vs. what got the headline

The headline coming out of the conference was “McDonald’s CEO predicts AI integration in 2025.” That is technically accurate and analytically useless. Every QSR CEO is predicting AI integration in 2025. The question is what kind, where, and on whose balance sheet the failure modes sit.

What Kempczinski actually flagged, per the trade press, was a three-part operator-facing thesis for the year: protein, AI, sauces. Two of those three are menu calls. The third is an infrastructure call. He grouped them together because, from the inside of a 40,000-location system, they are all the same kind of decision: small per-store cost, large aggregate exposure, and a payoff curve that depends on every general manager executing the same way at 7:45 a.m. on a Tuesday.

The AI portion of the message, as the trade press paraphrased it, was anchored to two named partners: Google Cloud on the platform side, Cognizant on the integration and operational rollout side. That partner pairing is not a hot take. It is a Fortune 50 CIO’s pairing. Google brings the foundation-model layer, Vertex tooling, and an existing analytics relationship McDonald’s has been building for years. Cognizant brings the unglamorous middle of the stack — the bit where you have to make a model talk to a forty-year-old POS schema in three thousand franchisee variants without breaking the lunch rush.

Notice what is absent from that pairing. No marquee voice-AI vendor. No headline-grabbing autonomous-ordering bet. No “we will have AI taking orders at X locations by Y date” number. That absence is the story.

Why McDonald’s downside is bigger than Wendy’s upside

This is the part of the analysis that almost no one is writing up, so I’ll be explicit about it.

McDonald’s serves roughly 26 million US customers a day. Wendy’s serves a fraction of that. The asymmetry matters because the failure mode of a voice-AI drive-thru system is not “a little bit worse than a human.” The failure mode is “the line stalls, the order is wrong, the customer drives off, and a thirty-second clip of the bot misunderstanding ‘no pickles’ is on every social platform by lunch.”

Wendy’s announced its FreshAI rollout last year with a target of 500 to 600 locations. CNBC’s July 2024 explainer on the QSR voice-AI landscape is the cleanest single source on the state of play across McDonald’s, Yum, and Wendy’s. The piece is worth re-reading because it dates the moment McDonald’s wound down its IBM voice-AI pilot — June 2024 — and contextualizes it against the chains that doubled down.

The framing inside that piece treats the three trajectories as roughly comparable: everyone is “testing.” That framing is, in my view, the wrong unit of analysis. Here is the right one. Compare the asymmetry on each side.

For Wendy’s, a 500-store FreshAI deployment that works has a meaningful margin and throughput story attached to it. A 500-store deployment that misfires costs the company some headlines and a write-down on a few million dollars of capex. The downside is bounded.

For McDonald’s, a system-wide voice-AI deployment that works has a smaller incremental margin story per store because the system is already operationally tighter. A system-wide deployment that misfires can damage brand trust in a way that takes years to rebuild. The downside is not bounded — it is reputational, and it is denominated in same-store traffic for multiple quarters.

In options-pricing language: Wendy’s is long a relatively cheap call. McDonald’s is short a relatively expensive put. Those are different trades. They should not be discussed in the same sentence as if both companies are “deploying AI.”

This is exactly the kind of asymmetry the rebrand-as-distributor-M&A frame in our later piece on US Foods and the QSR supply chain keeps returning to: scale that looks like an unambiguous advantage on the way up becomes an unambiguous fragility on the way down. McDonald’s AI roadmap is being built by people who have internalized that.

The Google + Cognizant build-vs-buy decoded

Read the partner choices as a build-vs-buy answer in disguise.

If McDonald’s wanted to own the AI stack outright, the partner list would look different. You’d see in-house ML hiring waves, a public commitment to a foundation-model training run, and a CTO doing the conference circuit. None of that is happening. Kempczinski is not the CTO, and the CTO is not the one talking.

If McDonald’s wanted to fully outsource the AI stack, you’d see a single integrator-of-record taking the risk. That is roughly what Yum has done with Nvidia in its public messaging. McDonald’s has not gone there either.

What Google + Cognizant actually buys is something more interesting: control of the data layer and the integration layer, rented compute and rented model capability. McDonald’s keeps the customer data, the menu data, the operational telemetry. Google provides the models and the infrastructure to run them. Cognizant provides the army of engineers and PMs needed to make that work across markets, franchisees, and regulatory regimes.

The strategic implication is that McDonald’s is preserving optionality. If the foundation-model landscape consolidates around one provider, McDonald’s can lean in. If it fragments, McDonald’s can swap providers without ripping out its operational data. If a regulator in Europe decides drive-thru voice capture is a GDPR problem, the data architecture lets them carve out a region without unwinding the global program.

My base case: this is what a $200 billion enterprise IT decision actually looks like. It is not a moonshot. It is not exciting. It is the right answer.

The IBM pilot ending in June 2024 is the empirical proof. McDonald’s ran a real, multi-location voice-AI trial with a credible partner, decided the unit economics and accuracy curve weren’t ready, and walked. That is the behavior of a system that has internalized the asymmetric-downside framing. A system optimizing for headlines does not walk. A system optimizing for the next ten years of free cash flow does.

What the May 1 print won’t tell you

The Q1 print drops next Thursday morning. Here is what I expect, and here is what to actually pay attention to.

What I expect on the print itself: low-single-digit US same-store sales, with international weighted toward the developed markets segment, a continued narrative on value-menu reset, and a margin line that’s roughly in the area of consensus. Currency will be a footnote. Commodity inputs will be flagged as benign-to-favorable. None of this will move the AI valuation discussion.

What I expect in the prepared remarks: one paragraph on technology investments, with Google and Cognizant named, framing the spend as a multi-year operating-efficiency program rather than a revenue-acceleration story. The word “AI” will appear, but it will be paired with words like “operational” and “back-of-house” more than with words like “drive-thru” or “ordering.”

What I expect on the Q&A: at least one sell-side analyst will try to pin management on a voice-AI rollout date. Management will not give one. The analyst will press. The CFO will redirect to operating margin guidance. The transcript will not contain a hard number on AI deployment.

What the print will not tell you, and what genuinely matters: the actual run-rate of cloud spend, the head-count picture in the technology org, and the contract structure with Cognizant. Those data points live in the 10-Q and in segment commentary, and they are the ones that tell you whether the build is real or whether it’s slideware.

My base case for the stock through the print: the AI narrative is a flat-to-discount overlay on the McDonald’s multiple right now, not a flat-to-premium one. Investors who want AI exposure in QSR are putting it on through Yum or, increasingly, through ghost-kitchen and back-of-house infrastructure names. McDonald’s is not in the AI premium bucket and probably won’t be in 2025. That is a feature, not a bug.

The companies in the bucket — the ones the market is rewarding for being aggressive — are the ones whose downside is bounded. Wendy’s qualifies. So does Domino’s on the digital-ordering side. McDonald’s, by virtue of its share and its brand surface area, does not qualify and should not try to. Kempczinski, to his credit, is not trying.

The contrarian read is straightforward. The CEO of the largest QSR in the world spent the back half of a panel last week telling a roomful of operators that he is going to be careful. He named the two most credible enterprise partners he could have named. He did not commit to a date. He let the IBM pilot’s June 2024 wind-down stand as a quiet data point about how the company evaluates these decisions. If you are pricing McDonald’s as a company that is “behind” on AI, you are pricing the wrong risk.

The Q1 print on May 1 will not resolve any of this. It will, if I’m reading the script right, simply confirm that the management team in Chicago understands the asymmetry better than the sell-side does. Watch the language. Count the named partners. And, most importantly, count the dates that are not given. The absence of a date is the most bullish thing this company can say about AI right now.

— Oliver writes The Bottom Line on M&A and valuations. Tips: [email protected].