Waiting a year on AI costs you a year of corrections you can't buy back

On Sunday, Microsoft CEO Satya Nadella published a post about what companies give up when they use AI models they don’t control. His argument, as TechCrunch summarized it, is that buyers pay twice: once in money, and again in the knowledge they feed the model through prompts, tool use and corrections. One line from the post, quoted by The New Stack, stuck with me: “Every correction is distilled into institutional know-how.”

Nadella was writing about who should own that know-how, and he has an obvious interest in the answer. But read the sentence again from the point of view of a company that hasn’t started. If corrections become know-how, then a company that waits a year has a year fewer corrections. That is the real cost of waiting, and it is the one owners underestimate.

The case for waiting sounds reasonable

The argument for waiting is easy to make, and it isn’t foolish. The models get better every few months. Prices fall. Tools that need an engineer today will be point-and-click next year. Why pay to be early?

There is also evidence that early money gets wasted. Last Wednesday Gartner published a customer service survey showing that service leaders put a median of 12% of their 2025 budgets into AI, more than any of the ten business functions Gartner looked at. Only 24% of those leaders showed positive financial returns across their AI use cases.

By the numbers

24%

of customer service leaders showed positive financial returns from AI

So yes, you can spend badly. My answer is to spend small and specific, on one department and one task. Waiting doesn’t fix spending badly. It just postpones the lesson.

Small means something concrete. Pick one task that a person does every day, such as answering order status questions or drafting the weekly client report. Use a tool you already pay for, or one that costs about what a part-time contractor would. Name one person who reviews every output for the first month. Measure the task before you start, so you can tell at the end whether it worked. A company that spends that way can be wrong about the tool and still come out ahead, because the instructions, examples and review habits carry over to the next tool.

What the tools won’t give you next year

Next year’s model will be better than this year’s. It will not know that your company never ships on Fridays, that your biggest client hates bullet points, or that “net 30” means net 45 for three of your accounts. It will not know which of your support macros are out of date, or which customer gets a phone call instead of an email when something goes wrong.

None of that comes in the box, no matter which vendor you pick. It comes from running an agent on real work, watching it get things wrong, and writing down the fix. Every correction becomes a line in the instructions or a new test case. After six months, a company that has done this has a document no vendor can sell it: a precise, tested description of how its own work gets done.

The same goes for people. A team that has spent a year reviewing an agent’s drafts knows what good output looks like, where the agent is weak and how to write an instruction that works. That is a skill, and like any skill it takes reps. You cannot hire it in a hurry when you finally decide to start, because the people who have it will be working for the companies that started earlier.

LoopHow a company gets good at this

1. 1Agent does the task
2. 2A person reviews it
3. 3The fix is written down
4. 4The rules get sharper

Your customers are not waiting

The same Gartner release has a finding that should worry anyone who plans to sit this out. Customers are about three times more likely to use third-party AI tools than a company’s own chatbot to sort out a service problem. Their use of outside AI tools for service nearly doubled in the past year. Among customers who use generative AI, 58% have used it to complete a task on their behalf. Among business buyers, it is 74%.

DataCustomers who let AI act for them

• All GenAI users58%
• B2B buyers74%

Think about what that means in practice. Your customers are already asking an assistant to compare your price to a competitor’s, draft the complaint email, or figure out your return policy. If your business is slow, unclear or hard to reach, their assistant will find that out faster than they would have. The bar your company is measured against is moving whether or not you move.

Gartner’s own advice to service leaders follows from this. It recommends shifting money away from standalone chatbots and toward letting customers complete actions, such as booking an appointment, submitting a document or updating an account. That is agent work, and it is exactly the kind of thing that takes months of tuning before it runs well.

Add up what a year costs

I would put the cost of waiting a year in plain terms, and I’d ask any owner to write their own version of this list.

• A year of hours your team spends on work an agent could do, which you can estimate today by timing the task.
• A year of corrections and test cases you didn’t collect, so your first agent next year starts as raw as this year’s would have.
• A year of your people not building the skill of directing and checking AI, while competitors’ people do.
• A year of hiring decisions made on old assumptions about how much one person can get done.

Against that, the savings from waiting are the difference in price and quality between this year’s tools and next year’s. For a first agent in one department, that difference is small. In my view, the tools that exist today already handle order status, lead follow-up, call summaries and report drafts well enough to pay for themselves when the rules are written carefully.

I have built companies for twenty years across television, VR and games, and every platform shift I’ve lived through rewarded the people who started learning while the tools were still awkward. Their advantage was a pile of mistakes already made and fixed.

What to do this week

1. Pick one task in one department that someone does every day, and time it for a week.
2. Turn on whatever agent you already pay for, on that task only, with a person reviewing every output.
3. Start a correction log: every time the reviewer changes something, write down what and why.
4. At the end of the month, turn the log into updated instructions and ten test cases.
5. Put the hours saved and the correction count on your leadership meeting agenda every month.

Sources

• Satya Nadella has issued a shocking warning to companies using AI (TechCrunch, July 13, 2026)
• Nadella’s reverse information paradox (The New Stack, July 13, 2026)
• Gartner survey finds customers are three times more likely to use third-party GenAI than company-provided chatbots for customer service (Gartner, July 8, 2026)