Gruow

Thinking

The models are good enough. The gap is knowing where they fit.

The model is ready. What it doesn’t have is your business.

A line has been going around AI circles: you don’t need God to write your email. The joke is that today’s models are already far more capable than most of what people ask of them. A smarter model won’t fix a vague request.

The evidence backs the joke. Models now match experienced professionals on about half of real work tasks, and the price of using them has fallen more than 280-fold in two years. Yet most companies report little or nothing from AI.

The difference isn’t the model. It’s knowing where AI fits in the work, and building it in there. Ask it for small, generic things and you get small, generic results.

The models are good enough

In September 2025, OpenAI published GDPval, a test built from 1,320 real tasks across 44 occupations in nine industries. Professionals with an average of 14 years’ experience wrote the tasks, which produced real files: documents, spreadsheets and slide decks. Other experts graded the results. The best models’ work was rated as good as or better than the experts’ in nearly half the tasks, at about a hundredth of the time and cost.

The price has collapsed too. Stanford’s 2025 AI Index found that running a model as capable as the original ChatGPT cost $20 per million tokens in November 2022 and $0.07 by October 2024.

The GDPval tasks came with one important limit: each model got one attempt, with no back-and-forth and no background on the client or the company. That’s the gap this piece is about. Anthropic reached the same view from how businesses actually use its models. In its September 2025 Economic Index, it found that “capabilities seem to matter more than cost” in what companies deploy, and that using AI for complex tasks “might be constrained more by access to information than on underlying model capabilities.”

In plain terms: the model is ready. What it doesn’t have is your business.

Same tools, very different results

Almost every company now has access to the same models. The results couldn’t be further apart.

The consulting firm BCG surveyed 1,250 senior executives in 2025. The few companies getting value at scale reported 1.7 times the revenue growth of the rest and a 1.6 times higher operating margin.

McKinsey’s 2026 survey of 1,719 people in 97 countries found the same shape. Only about 6% are “high performers”, meaning AI accounts for at least 5% of their operating profit. 37% say AI has added to profit at all.

The clearest picture of generic use comes from Denmark. Economists Anders Humlum and Emilie Vestergaard linked survey answers from 25,000 workers in 11 occupations exposed to AI chatbots with official payroll records. Average time saved was 2.8%, and they found no measurable effect on earnings or hours worked. But where employers invested in training and their own tools, people used AI more and found new tasks for it. Access alone did little. Fit did the work.

Where it fits, and where it doesn’t

AI isn’t good at everything to the same degree, and the edge between what it does well and badly isn’t where people expect. Researchers from Harvard and BCG called this the jagged frontier, and tested it on 758 of BCG’s own consultants in 2023.

The same AI made consultants better on some tasks and worse on others

Consultants with GPT-4 against those without, 758 BCG consultants, 2023

Source: Dell’Acqua and others, Harvard Business School and BCG, 2023, via HBS Working Knowledge. Redrawn by Gruow.

On a product-design task inside the frontier, consultants with GPT-4 did more work, faster, and better. On a business-analysis task just outside it, the same tool made them more likely to get the answer wrong, because it gave a confident, plausible answer that missed what the task needed. Same people, same model. The only thing that changed was the fit between the tool and the task.

Fit also depends on who’s using it. In a study of 5,179 customer support agents, Erik Brynjolfsson, Danielle Li and Lindsey Raymond found that an AI assistant raised issues resolved per hour by 14% on average, and by 34% for new and less-skilled agents. For the most experienced agents, it made almost no difference. The assistant had learned from the best agents’ conversations, so it passed their know-how to everyone else.

Knowing where AI fits means knowing both: which tasks sit inside the frontier, and which people gain most from help with them.

What it looks like when it fits

The companies with real results didn’t ask AI for small things. Each picked one costly, repeated piece of work, gave the model what it needed to do that work, and built it into how the work already ran.

CompanyWhere they fitted AIWhat changedReported by
DuolingoCourse creation: one strong base course, adapted by AI for dozens of languages, checked by learning designers148 new courses in about a year. The first 100 took about 12 years.Company, April 2025
WalmartProduct catalog data, the details shoppers search and filter byOver 850 million data points created or improved. Executives said it would have taken 100 times the staff.Earnings call, August 2024
Morgan StanleyAdvisors’ research library and client meeting notes, tested against expert answers before rolloutOver 98% of advisor teams use it. Advisors’ access to the firm’s documents rose from 20% to 80%.OpenAI case study
A Fortune 500 software companySupport agents’ replies, with suggestions drawn from its best agents’ conversations14% more issues resolved per hour, 34% for newer agentsBrynjolfsson, Li and Raymond, independent study
KlarnaCustomer service chat in 35+ languages2.3 million conversations in the first month, two-thirds of all chats. Resolution time fell from 11 minutes to under 2.Company, February 2024

Most of these figures come from the companies themselves, so they show the best case. Klarna also shows the limit. In May 2025, its CEO said the AI-first approach had led to “lower quality” service, and the company began hiring people again so customers could always reach a human. Knowing where AI fits includes knowing where it stops.

What the companies that get value do differently

Put the cases and the surveys side by side and the same pattern shows up.

Generic use stays small. Fitted use changes how the work runs.

Generic use anyone, any task, one chat at a time

  1. Ask a chatbot a general question
  2. Get a generic answer
  3. Paste it into the work

Fitted use one costly workflow at a time

  1. Pick a costly, repeated task
  2. Give it your data and rules
  3. Build it into existing tools
  4. Test it against your best people
  5. A person approves what matters

Measure the result, then adjust or move to the next task

The pattern behind the cases above. Drawn by Gruow.

They change the work, not just the tools. In McKinsey’s 2026 survey, nearly three-quarters of high performers had fundamentally redesigned workflows around AI. About a quarter of everyone else had.

They aim at the core of the business. BCG puts 70% of AI’s potential value in core functions such as sales, marketing, manufacturing, supply chain and pricing, and only 13% in IT. Walmart’s catalog and Duolingo’s courses are the product itself, not a side task.

They give the model what it lacks. A model knows the world, not your company. Morgan Stanley pointed its assistant at the firm’s own research library. The support tool in Brynjolfsson’s study learned from the company’s best agents. That’s the missing context Anthropic’s data points to.

They test before they trust. Morgan Stanley graded every use case against expert answers before rollout, and ran daily tests after. Advisors still review the meeting notes before they go to clients.

How to find where AI fits in your business

You don’t need Walmart’s scale for any of this. The same questions work for a five-person firm.

  1. Start from the work, not the tool. List the tasks that happen often, follow rules, involve a lot of text or data, and keep someone waiting. That’s where the frontier usually runs in your favour.
  2. Put a cost on each one. Hours a week, delays, errors, customers lost. Pick the one where a fix would matter most.
  3. Write down what a good result looks like. Use real examples from your best work. These become the test.
  4. Give the model your context. Your documents, your rules, your past answers. This is the step generic use skips, and the one the evidence says matters most.
  5. Put it where the work already happens. In the inbox, the CRM, the catalog, the support queue. A tool people have to remember to open is a tool they stop using.
  6. Keep a person on what matters, and measure. Compare before and after on the cost you named in step two. If it doesn’t move, change the fit or drop it.

The models will keep getting better, and that won’t close the gap on its own. The businesses that pull ahead won’t be the ones with the smartest model. They’ll be the ones that know their own work well enough to show the model where it fits.

Sources