The AI Works. Your Company Is What's Slowing It Down.

Eric Vaughan, chief executive of the enterprise-software company IgniteTech, gave every Monday over to AI. Customer calls and budget work were off the table; as he told Fortune, "you had to only work on AI projects." The company paid for AI tools and prompt-engineering classes, brought in outside experts, and by Vaughan's account put 20% of payroll into the effort. The effort, Fortune reported, ran into mass resistance. Within about a year IgniteTech had replaced nearly 80% of its staff, according to headcount figures Fortune reviewed, and Vaughan is clear that this is no recipe: "I do not recommend that at all." His verdict on the whole thing: "Changing minds was harder than adding skills."

Vaughan had the tools, the training and a fifth of his payroll to spend, and the hard part was still his own company. Most firms are in a quieter version of the same place.

A common explanation is that the technology isn't ready yet. The next model will be smarter, the next release will plug into your systems, so wait. That explanation is mostly wrong. For a great deal of ordinary work, the tools are already good enough. The research points to what separates the companies getting value from the ones stuck in pilots, and most of it is inside the building. Three things stand out to us as the place to start. The leaders take part themselves instead of handing AI off to the IT department. Someone who actually runs the work is put in charge of it and, we'd argue, given the time to do it. And the company takes one painful process all the way from a demo that mostly works to something it can rely on, on a Tuesday when nobody is watching.

The surveys don't blame the software

The research is blunter than the vendors are. In October 2024 Boston Consulting Group reported that 74% of companies "have yet to show tangible value from their use of AI." When BCG sorted out what was in the way, it put "around 70%" of the trouble down to people and process, 20% to technology, and "only 10%" to the AI algorithms.

In 2025 a research group at MIT published a report called The GenAI Divide. It found that only about 5% of the AI pilots it studied were producing rapid revenue gains, while most delivered "little to no measurable impact" on profit and loss, Fortune reported on August 18, 2025. As Fortune summarized the report, the real issue was not the quality of the models but the "learning gap" for both tools and organizations. The gap runs both ways: tools that don't adapt to how the work actually runs, and companies that don't change the work to fit the tools.

McKinsey's State of AI report, published in March 2025 from a survey fielded in July 2024, tested 25 things companies do and asked which of them went with AI showing up in operating profit. Redesigning workflows had "the biggest effect." Yet only 21% of respondents whose organizations use generative AI said their organizations had fundamentally redesigned even some of their workflows. Two more factors on McKinsey's list: whether the CEO oversees AI governance, and whether senior leaders are "role modeling the use of gen AI."

Put the three side by side and they tell one story. What differs is what happens after the purchase.

The boss goes first

In April 2025 Tobi Lütke, chief executive of Shopify, made public an internal memo saying that reflexive AI use was now a "baseline expectation," that it would count in performance reviews, and that before asking for more headcount and resources, teams "must demonstrate why they cannot get what they want done using AI."

The headcount rule is the striking part. The more useful line was about Lütke himself. The expectation, he said, "goes for me as well as everyone else."

Whether the boss uses the tools matters because the decisions that turn a pilot into a process belong to leaders. Which job changes? Whose sign-off is no longer needed? How many errors are acceptable, and who answers for them? A leader who has never watched an AI get something subtly wrong can't make those calls well. They either trust the demo too much or don't trust it at all. Alexander Sukharevsky, who co-leads McKinsey's AI arm, puts it plainly in the March report: companies' instinct to hand the job to the IT department "over and over again... turns out to be a recipe for failure."

Time for the people who do the work

The second thing is time, and we think it is the one companies most often get wrong. Classes and a weekly AI day for the whole staff is one thing. Clearing part of the week for the person who runs the invoice queue, so that she can own one process until it works, is another. Vaughan bought the first, at great expense, and got resistance.

The second kind of time pays for dull work that no class covers: chasing down the cases the tool gets wrong, writing out rules the team has always kept in its heads, and deciding where the exceptions go. A pilot that runs on an enthusiast's spare hours, we'd argue, tends to get as far as a demo and stop there.

It also has to be the right person's time. According to Fortune's account of the MIT report, one key factor for success was "empowering line managers—not just central AI labs—to drive adoption." The person who runs the invoice queue knows which suppliers always send the wrong reference number. The AI team in head office doesn't.

From a demo to a process you can rely on

The third thing is the hardest. Aditya Challapally, the MIT report's lead author, explained why some young AI startups were succeeding: "they pick one pain point, execute well."

Klarna, the Swedish buy-now-pay-later company, shows both the gap and what working on it looks like. In February 2024 it said its AI assistant had handled 2.3 million conversations in its first month, doing the equivalent work of 700 full-time agents. In May 2025 its chief executive, Sebastian Siemiatkowski, told Bloomberg that the chatbots were cheaper but delivered "lower quality," and that customers should know "there will always be a human if you want."

It sounded like a reversal. The numbers Klarna gave Fortune in August 2025 tell a more modest story. Its contractor had rehired "just two people," for a pilot pairing highly trained staff with the AI.

So the software kept doing more. What changed was the arrangement around it. The chief executive has said publicly that customers can always reach a person if they want one. As of August 2025, the company was testing, with two rehires, whether trained people working alongside the bot do better. And a company that once advertised how many agents its AI could replace has since heard its chief executive tell Bloomberg that "investing in the quality of human support is the way of the future." Klarna hadn't finished that work; as of August 2025, the human pairing was a two-person pilot. But it is the right kind of work: not waiting for a better model, but deciding where the machine stops.

The objection: just wait for better models

The strongest case against this argument goes like this. Models have improved, and some of what struggled at launch works better now; Klarna says as much about its own assistant. If the tools keep getting better, maybe the smart move is to let others do the painful redesign and buy the finished product later. MIT's own numbers point partly that way: buying from specialist vendors and partnering with them succeeded about 67% of the time, Fortune reported, and in-house builds about a third as often. And the surveys deserve some scepticism. They rely on what managers say about their own companies, and by McKinsey's own count, the 25 factors together explain only part of why some companies do better than others.

Some of this is right. Tools that don't learn from feedback are a real technical limit, and the MIT researchers say so. Buying is often wiser than building. But even a bought tool still has to land in a process someone owns, with someone deciding what happens to the cases it gets wrong. A better model won't rewrite your approval chain or tell you which customers must always reach a person. That was a decision Klarna had to make, not a feature it could wait for. And the head start compounds. A company that has taken one process from demo to dependable has learned where its own rules were never written down, and no software update delivers that.

What changes

If this is right, the question for next year's budget is not which tool to buy. It is three questions. Have you used AI on your own work this month? Who, by name, owns one process, with time cleared from their week to make it work? And which painful workflow will you take all the way, so that a year from now, on an ordinary Tuesday, the routine cases go through without anyone checking, the odd ones land on a named person's desk, and someone looks each month at whether customers got worse answers?

Spend less of next year's AI money on licences and more on one person's time and one finished process. For a great deal of ordinary work, the software is ready. It is waiting on your company.