Every multi-unit operator I talk to has been pitched an AI tool this year. A bot for the operations manual. A drive-thru voice. A lease reader. A candidate screener. The pitch always answers the same question: what can this do?
That’s the wrong question. The right one is: what will this amplify?
AI doesn’t install an operating system. It audits the one you already have.
The Oldest Rule in Business Technology
There’s a line usually credited to Bill Gates that every operator should tape to the wall before a software demo: “Automation applied to an efficient operation will magnify the efficiency. The second is that automation applied to an inefficient operation will magnify the inefficiency” (IT History Society). Even the attribution is shaky (the archive that carries it lists the source as unknown), but the rule has held up for decades, and AI is the strongest test of it yet.
The evidence so far points one way. McKinsey’s 2026 State of AI survey found that about 6% of companies saw significant financial impact from AI, almost exactly the same share as the year before, despite heavier adoption. The difference between that 6% and everyone else wasn’t which model they bought. About three-quarters of the top performers had fundamentally redesigned their workflows, compared with about a quarter of everyone else (McKinsey via WSJ Partners). McKinsey’s own summary is blunt: “Value flows only when organizations redesign their operating models to capture the new technology’s promise” (McKinsey).
You’ve probably also seen the claim that “95% of AI pilots fail.” Be careful with it, for the same reason you should be careful with the old line that 80% of franchises succeed. The MIT report behind it actually found that around 5% of custom enterprise AI tools reached production with sustained impact, drew on 52 interviews and 153 conference surveys, and was never peer reviewed (People Team AI). What’s worth keeping is its explanation: tools stalled because of “brittle integrations misaligned with day-to-day operations,” and the divide “does not seem to be driven by model quality” (AI Wiki summary of MIT NANDA). Different number, same lesson.
What It Looks Like in Our World
The drive-thru is the cleanest experiment our industry has run.
McDonald’s tested IBM’s automated order taker in more than 100 restaurants and shut it down in 2024 after it struggled to understand customers and get orders right (Yahoo Finance). Yum slowed its voice AI rollout across about 500 restaurants in August 2025 after complaints and the now-famous order for 18,000 cups of water (Restaurant Dive).
Now look at Bojangles. Its voice system runs in 452 locations, needs a human only 1 to 2% of the time, and cut drive-thru workload by 35%, according to the company. The detail that matters isn’t in the headline. The rollout came after six months of preparation, and the team added around 550 unique menu items to the system in the first month of testing (Franchising.com).
Read that again. The hard part was writing down, exactly, every way a customer can order. That isn’t AI work. That’s operating system work. Bojangles did it first, and the AI worked.
AI Is a Stress Test, Not a Fix
Here’s the pattern I’d expect in most multi-unit companies. Put a chatbot on top of the operations manual and ask it how to close the store. If three versions of the closing procedure are floating around (the binder, the district manager’s version, and the way the best store actually does it), the bot will pick one and deliver it with total confidence. You haven’t fixed the inconsistency. You’ve made it faster and harder to see.
The research on knowledge workers shows why that’s dangerous. In the Harvard and BCG study of 758 consultants, AI made people faster and better on tasks it was suited to. On a task outside its range, accuracy dropped from 84.5% without AI to 70.6% with it, and to 60% for the group that also got prompt training. The researchers’ warning: “polished output is not the same as sound reasoning” (Harvard Business School Digital Data Design Institute).
There’s a second trap: AI speeds up one step and exposes the next one. A study of more than 100,000 developers found AI coding tools raised coding activity by 180%, but software releases rose only 30%. The bottleneck moved to the human steps of reviewing, integrating, and releasing (Knowledge at Wharton).
Translate that to our four lanes:
| Lane | Where AI can genuinely help | What has to be true first |
|---|---|---|
| Field operations | Finding the right procedure fast, coaching newer managers, flagging exceptions | One current, owned version of each procedure; a clear rule for what happens when the answer isn’t in the manual |
| Development | Tracking opening checklists, chasing status, summarizing handoffs | Defined handoffs with an owner and a “done” standard at each one |
| Real estate | Abstracting leases, summarizing site packages, side-by-side comparisons | Someone who owns critical dates and verifies the output; approval criteria that exist before the summary does |
| Franchise infrastructure | Answering candidate questions, onboarding support, routing field questions | Approved answers, a disclosure-safe boundary, and a human path for anything outside it |
A lease abstract in 30 seconds doesn’t help if no one owns the renewal date. A beautiful site summary doesn’t help if the approval meeting has no criteria. MRI Software, which sells lease AI, makes the same point: success depends on “the controls and processes to verify its outputs” (MRI Software).
Where the Other Side Is Right
The strongest counterargument deserves a fair hearing, because parts of it are correct.
First, AI can be the thing that finally forces the process work. Bojangles’ six months of preparation was process repair. Arguably the AI project was the only reason it got done. If a tool purchase is what finally makes a company write down its menu, its closing procedure, or its lease calendar, that’s a real benefit.
Second, AI can capture know-how the company never documented. In a study by economists Erik Brynjolfsson, Danielle Li, and Lindsey Raymond of 5,179 customer support agents, an AI assistant raised productivity 14% on average and 34% for newer, less-skilled workers, with “suggestive evidence” that it spread the practices of the best performers (Stanford SIEPR). In other words, the tool did some of the operating system’s job.
Third, waiting for a perfect process is its own way to stall. The MIT report found that tools built for specific processes could deliver value “even without major organizational restructuring” (AI Wiki summary of MIT NANDA). And Taco Bell didn’t quit after the parent company’s stumble. It kept expanding voice AI through a separate partner, Omilia, which now runs in more than 890 U.S. Taco Bells (Restaurant Dive). Learning in public worked for them.
All fair. But notice what each example has in common: the value came when the company treated the project as process work, not as a software purchase. That’s the argument, not a rebuttal of it.
Five Questions Before You Buy
Before any AI tool touches your operation, ask:
- Is there one written, current version of this process, and does someone own it?
- If a new hire asked two of your best managers, would they get the same answer?
- Who checks the AI’s output, and how quickly would an error get caught?
- What happens when the situation isn’t covered? Is there a human path?
- Do you know how the process performs today, so you can tell whether the tool made it better?
If you can’t answer the first two, you don’t have an AI problem yet. You have an operating system problem, and the AI will find it for you, at scale and in front of your customers.
That’s where I land. If AI can strengthen an operating capability, use it. If it just automates a bad process, don’t. AI doesn’t install an operating system. It audits the one you already have.