Nine verbs, one jagged frontier.

“Understanding AI” series, note 4 of 12. You now know what a language model is, how it works inside and how it differs from software and process robots. This note answers the question I am asked in the first meeting: “Fine, but what exactly does it do in my company?”

Monday, 8:15 am, a furniture factory near Bern, eighty employees. The managing director has asked every head of department for a list of “where AI could help us”. On his desk are forty-two proposals. Sales wants it to draft the quotes. Administration wants it to read suppliers’ invoices. Production wants it to predict when the edge-banding machine will break down. Human resources wants an assistant for questions about holidays. Marketing wants videos. And the head of quality, who is the most sceptical, has written a single line: “And who checks what it does?”

The scene is imaginary. The forty-two proposals are not: that is what happens when you ask without a map. And the forty-two are not worth the same. Some are solved by a software rule. Others are exactly what a language model does better than anyone. Others are, today, a catalogue promise. This note is the map I would put on that desk: nine verbs, seven departments, four studies that have measured real results and one warning.

The question is not “where can AI help?”. It is “which verb do I need here, and what happens when it conjugates it badly?”.

The nine verbs.

A language model does not “do AI”. It does concrete things with text, and more recently with voice and images. Once you know them by name, every proposal in your company becomes comparable with the others.

• Classify. Read something and say what kind it is: this email is a complaint, this invoice is for transport, this call is urgent. It is the most reliable verb and the one that adds the most silent value, because it tidies what ordinary software will then process.
• Extract. Pull concrete data out of messy text: the order number from an email, the amount and date from a scanned invoice, the penalty clauses in a contract. What a person used to type in.
• Summarise. Turn thirty pages into one, an hour-long meeting into five lines, a hundred reviews into three themes. Very good, with one risk: a summary leaves things out, and you do not see what it leaves out.
• Draft. A first draft of a quote, a reply to a customer, a job description. It is what most people already do in secret. We will see the figures.
• Translate. Between languages, and also between registers: technical into plain language, informal into formal. For a Swiss company serving customers in four languages, it is often the first profitable use.
• Answer from your documents. Answer questions by reading, for every query, the manuals, rates and policies approved by management, and citing where each fact comes from. It is the basis of any serious assistant. The technique is called retrieval and I explain it in note 5.
• Transcribe and speak. Turn voice into text and text into voice. Calls, meetings and notes dictated in the car become text on which to apply the previous verbs.
• See. Read scanned documents, photos of delivery notes, drawings, labels. What until recently needed a specialised character-recognition program is now done by the same model that writes.
• Code. Write and fix software. It is the verb that has changed the technology sector most, and the reason custom automations have become much cheaper to build.

Notice what is not on the list: decide. A model can propose a decision, rank the options and argue. But the decision, on anything with consequences, is a verb that stays with a person, for the reasons in note 2: the model predicts the probable, not the correct, and tends to tell you what you want to hear.

Seven departments, with their verbs.

Back to the factory’s forty-two proposals, sorted with the map. This is not a list of what is possible; it is a list of what I have seen work or what the studies have measured.

• Customer service. Classify and answer from documents. The model sorts the inbox, proposes the reply and the person reviews or sends it. It is the department with the best evidence, as you will see in a moment, and also the one with the Klarna warning.
• Administration and finance. Extract and classify. Suppliers’ invoices read and booked with a final review; reconciliation of transactions; automatic chasing of overdue payments. Here the model tidies and the software executes. It is by far the place where the most money is recovered per franc invested, and the one that usually gets the least budget.
• Sales. Draft, summarise and extract. Preparing a meeting from everything the company knows about the customer; a draft quote from an approved template; a summary of the call in the sales system without the salesperson typing anything.
• Human resources. Answer from documents and draft. An internal assistant for the usual questions, which reads the collective agreement and the staff regulations and cites the article. And a warning: recruitment with AI is strictly regulated in the European Union and delicate ground in Switzerland. It is not a verb to start with.
• Legal and compliance. Extract and summarise. Comparing a contract received with your own template and marking the differences; summarising a new regulation. Always with a person who signs, and with the lesson of Mata v. Avianca from note 2 firmly in mind.
• Marketing. Draft and translate. First drafts, variants, adaptation into four languages. It is where the most is spent and the least measured; we will see it in the note on costs.
• Operations and production. Classify and see. Incident reports dictated by voice and classified; photos of a defect compared with the defect catalogue. Predicting when the edge-banding machine will break down is another family of AI, the one of sensors and time series, and it needs years of data that almost nobody has.

What has actually been measured.

This is where I separate what I know from what I am told. There are four studies I cite in almost every meeting because they measured with control groups, in real companies or tasks, and published the full numbers.

The first is from customer service. In April 2023 Brynjolfsson, Li and Raymond published with the National Bureau of Economic Research the study of 5,179 support agents at a software company who were given, in stages, an AI assistant that suggested replies. Productivity, measured in queries resolved per hour, rose by 14% on average. And the figure that interests me most: the least experienced agents improved by 34%, while the most expert barely changed. The assistant, in practice, passed on to the newcomers what the veterans knew. Staff turnover fell. It is the best argument I know for AI as a covert training tool.

The second is about writing. In July 2023 Noy and Zhang published in Science an experiment with 453 university-educated professionals doing writing tasks typical of their jobs. With ChatGPT they took 40% less time and quality, rated by other professionals, rose by 18%. As in the previous case, those who wrote worst at the start gained the most. The tool closes the gap upwards.

The third is the one that changes how you think. In September 2023 a team from Harvard, MIT, Wharton and the Boston Consulting Group had 758 BCG consultants do real tasks, some with GPT-4 and some without. On tasks within what the model does well, those who used it completed 12.2% more tasks, 25.1% faster and with 40% higher quality. But the researchers deliberately included a task that looked just as easy and sat just outside that capability, one that required cross-checking data with interviews and applying judgement. There the consultants with AI were right considerably less often than those who worked without it: the group without AI got it right in 84% of cases and the groups with AI stayed between 60% and 70%. The authors called it the jagged frontier: the model’s capability is not a circle that grows, it is an irregular line on which two neighbouring tasks can fall on different sides, and from the outside you cannot see where the line is.

That is the warning almost nobody includes in the presentation. The same tool that makes your team 25% faster on one task makes it worse on the next, and worse with confidence, because people tend to accept what the model proposes when it is well written. The study describes two ways of working that avoid this: the centaur, who divides the work and does some tasks with the machine and others without, and the cyborg, who mixes both at every step but checks relentlessly. Neither of them consists in delegating.

The fourth study is really a story in two parts, and it is worth telling in full. In February 2024 the Swedish payments company Klarna announced that its AI assistant had held 2.3 million conversations with customers in its first month and was doing the work of 700 agents. It was the most cited case of the year in presentations. In May 2025 its chief executive publicly acknowledged that the bet had produced lower quality and that the company was hiring people again for customer service, with AI as support rather than replacement. It is not a failure of the technology. It is the jagged frontier applied to a whole company: it worked where the model does the verb well, and failed where the customer needed someone with judgement.

What is already happening in your company without your knowing.

There is one last figure I recommend keeping in mind before deciding anything. In May 2024 Microsoft and LinkedIn published a survey of 31,000 people in 31 countries. 75% of knowledge workers were already using generative AI at work, and 78% of them were using their own tools, not ones provided by the company. Most told nobody. This means the question “do we adopt AI?” comes too late. Your team has already adopted it. The real question is whether they do it with the company’s tools, data and rules or with a free account and your customers’ data. I go into detail in the note on security.

1. This week

   Ask for the proposals, but with a verb

   Have every head of department write their idea with one of the nine verbs. What fits none of them is usually classic software or a promise.

2. This week

   Mark the frontier

   Next to each proposal, write what happens if the model is confidently wrong. Where the answer is “nobody would notice”, a review is needed.

3. This month

   Start with administration

   Invoices, reconciliation, collections. Less showy than a video, more money and less risk.

4. This month

   Ask your team what they already use

   Without punishment. You will find out which verbs they need and which data they are conjugating them with outside the house.

How I approach it.

When I walk into a company I do not ask where AI could help. I go through the departments with the nine verbs and a notebook, and I write down three things per process: how many hours go into it, which verb is missing and what happens if it is conjugated badly. At the end there is a map in three colours: what a rule solves, what a model with review solves and what is not touched today. The Strategic Diagnostic I offer is exactly that map, made with your numbers.

Do the exercise with your own week. Write down the five tasks that take most of your time and put the matching verb next to each. Some will have none, and that is information too.

Now look at the one you would most like to get rid of and ask yourself the question that really matters: if a model got it wrong with complete confidence, how long would it take for someone to notice?

Next note: how the verbs connect to each other. What retrieval is, what a tool is, what an agent really is and why a fourteen-step system that gets each step right 95% of the time fails half the time overall.