Most companies know which AI tools they pay for: a writing assistant in marketing, a meeting-notes app, a few ChatGPT subscriptions on expenses, whatever came bundled with the Microsoft licences. Far fewer can say where AI actually sits in the business, or what it has changed there.
That gap shows up in the numbers. In McKinsey's State of AI 2026 (1,719 respondents in 97 countries), 80% say AI has made them personally more productive. Only 37% see any effect on their company's operating profit, the same share as a year earlier.
Last week I presented a simple framework for that gap to about forty business owners and professionals at an event in Munich. It has five things AI can do, four levels, three things that decide whether it works, and five steps for next week.
- Five things AI can do in a business
- The four levels
- Most AI projects don't fail. They get orphaned.
- Three things that decide whether it works
- Five steps for next week
Five things AI can do in a business
"Use AI" is too vague to act on. In a business, AI does five things. In the talk I called them flavors, because that is the whole menu.
- Read. Invoices, contracts, forms, each in a different layout. It pulls out the fields so nobody has to type them.
- Find. You ask in plain words. It answers from your own documents and shows you the page.
- Draft. The first version of anything you write over and over. A person approves it before it leaves.
- Check. The delivery note against the order, the form against the rules. It flags what doesn't fit.
- Decide. Inside a boundary you set. The routine case gets handled, and the odd one goes to a person.
So where should your company start? That depends on how far it already is. Nobody starts with all five. Companies climb into them in an order, and the order has four levels.
The four levels
Level 1: Everyone on their own
Individuals use free or cheap tools for their own tasks. The company isn't using AI; people are, privately, and nobody is keeping track. Emails written with it every day. A summary of the forty-page PDF nobody wanted to read. A picture for the marketing post. The colleague who drafts every awkward message in ChatGPT first.
That matches how people use ChatGPT in general. In the study OpenAI published with NBER economists in September 2025, based on more than a million consumer messages, over 70% had nothing to do with work. Writing was the biggest share of the work messages, about 40%, and two-thirds of all writing requests asked ChatGPT to edit, translate or critique text the user already had.
One flavor, draft, on your own material, in your own browser tab. The productivity is real. It is not a strategy yet. It is a hobby that happens at work.
Check question: can you say who in your company uses AI, for what, and on whose account?
Level 2: Every department its own AI
Six departments, six tools, each one an island. In sales, AI inside the CRM writes the follow-ups. Support has a chatbot on the website, marketing a copywriting tool, the office the assistant in the document suite, operations a meeting-notes app, and HR a CV screener.
Find arrives at this level, officially and with a budget behind it, one island per department. Each tool works, and none of them talk to each other. In the ServiceNow/ThoughtLab Enterprise AI Maturity Index 2026 (4,500 executives and 2,000 employees in 19 countries), the share of companies whose AI workflows were integrated across departments fell from 30% to 16% in a single year, while AI spending rose 110%. More tools, and fewer connections between them.
Check question: how many separate AI subscriptions does the company pay for, and do any two of them share what they know?
Level 3: AI that knows your company
Level 3 is one place where the AI can read everything your company knows. For the first time, it knows who you are. It reads the current manual, not every old version of it. It knows what you agreed with this customer. It gives one answer from the price list, whichever department asks. It turns years of tickets and meeting notes into plain words, and a new hire who asks "how do we do this here?" gets the real answer.
Read arrives here, on your own knowledge, not the internet's. The AI knows nothing your company didn't already know. The difference is that it can finally read it.
There is one condition, and it comes back below: your documents have to be in order.
Check question: ask it something only your company knows, a pricing rule or a client's exception, and see whether the answer is right and shows where it came from.
Level 4: AI where it belongs, and nowhere else
Here is the twist: the most mature company uses AI in fewer places than the one below it. On purpose.
All five flavors are available at this level, and check and decide are new. Each one is used only where it belongs. A process with ten steps might call AI in three of them; the other seven stay as plain code, because code does the same thing every time and can be tested. AI to understand, code to execute. If it has to add up to the cent, it is not a job for AI. If you need the same answer twice, you need code, not AI. If the rule was never written down anywhere, there is nothing to automate yet.
The same goes for what your own people build. Chances are someone in your company built an app last month without telling anyone. The question isn't whether to allow it. It's whether it runs on rules you can trust.
The mature company learned where AI was a liability, and took it back out.
Check question: can you name one place where you decided against AI, and say why in one sentence?
That is the ladder. Level 1 uses draft. Level 2 adds find. Level 3 adds read. Level 4 adds check and decide, each only where it belongs. Choose the work by where you are on the ladder. A company at level 2 that starts with a project where AI decides on its own is skipping two levels, and the data and processes that project needs are not there yet.
Most AI projects don't fail. They get orphaned.
Whatever level a company is at, the record so far is thin. In Deloitte's State of AI in the Enterprise (January 2026, 3,235 leaders at companies that already use AI daily), three in four had moved fewer than 40% of their AI experiments into production. In McKinsey's State of AI 2026, the share of respondents who attribute 5% or more of their operating profit to AI is about 6%, unchanged from last year.
Most of those projects did not fail. They were orphaned: they worked, and then nobody was holding them. Three things decide whether yours ends up as one of them.
Three things that decide whether it works
None of them is a tool: the data the AI reads, the way the work is done around it, and the people who have to change what they do. Each one needs work before the tool arrives. Skip any of them, and the project ends up with nobody holding it.
Get the data ready
In the CIO-Agenda 2026 survey of decision-makers in Germany, Austria and Switzerland (Foundry/CIO-COMPUTERWOCHE with Lufthansa Industry Solutions, 294 respondents on this question), 44.2% graded their own data quality for AI agents a five or a six on the school scale, the two failing grades. Only 29.6% called it good or very good.
The AI can only read what exists, in one place, that people agree is correct. If a company keeps ten versions of the same manual, a chatbot on top will answer from all ten, with the same confidence. Before anyone builds it, someone has to decide which version is true and retire the rest. That decision is the work. The chatbot is the easy part.
Redesign the work, not just the tool
In the same ServiceNow/ThoughtLab index, only 5% of companies are redesigning the work itself. Most of the others are automating the workflows they already had, step by step, with the approvals, handovers and waiting times still in place. Many of those workflows were built around what people could manage by hand: forms that exist because someone had to retype the data, a second check because the first one was unreliable. Automate a process as it is, and it keeps every one of those flaws. Faster. Before the tool arrives, ask which steps should still exist at all.
Bring the people with you
In BCG's AI at Work 2026 (9,923 people who use AI at work), 80% reported measurable impact when leaders had said clearly what should change, even with limited tools. With strong tools and no clear direction, only 60% did. Direction beats tools. A licence for everyone is not direction, and neither is a training day. Direction is saying which work should look different in three months, and where the time that comes free should go: to the backlog, to the customers waiting, to the work that keeps getting postponed. If nobody decides that, the hours AI saved have nowhere to go. They get absorbed back into the day, and nobody can point to what changed.
Five steps for next week
None of them is a purchase.
- Measure your level. Answer the four check questions above, honestly, one per level. Your level tells you which flavors your company can use today. Then have the people who do the work answer the same questions. Directors and staff rarely land on the same level, for reasons I described in Your AI programme looks better from where you are sitting.
- Map where the hours actually go. Ask the people who do the work, not the org chart. What you want are the messiest processes: the ones where the same data gets retyped twice, where a case sits for three days waiting on an approval, where everyone knows the workaround and nobody has written it down. Those leak the most time and money, and your candidates are among them. The process you would name today rarely comes out on top.
- Pick one piece of work your level can do. Choose it from the candidates in step 2, and use only the flavors your level allows. A good first candidate passes four tests. It comes up every week, so within a month you have enough cases to see whether anything changed. The person responsible for it wants the help, so someone will keep it going. The information it needs is in one place, not spread across five systems. And when the AI gets it wrong, the cost is a correction, not a loss: a draft someone has to rewrite, not a payment sent to the wrong supplier.
- Walk through it, and write down what should change. Which steps go, which stay with people, and where the saved hours will go. If nobody writes it down, the time goes where it always went.
- Give it an owner, a number and an end date. The owner is a person with a name, not a department and not a committee. The number is your metric: how long a finished case takes, start to finish, measured once before anything changes and again afterwards, so you have something to compare. The end date is a date in the calendar, close enough that someone has to report back. A pilot without those three keeps running indefinitely, and six months later no one can say whether it helped.
And pick just one: one process, one owner, one number, one date. Three pilots at once is how companies end up with none.
The talk I gave ended on one line: AI isn't a purchase. It's a decision. Don't buy another AI tool next week. Decide where AI belongs in your company, and where it doesn't.