top of page

Remembering is no longer the edge. What gains value when AI can prepare a briefing in minutes?

10 minutes ago
12 min read

For many years, in meetings and committees, and simply whenever decisions were made, you could see the same picture. The person who was valued was the one who kept all the numbers in their head: revenue by business line, the deadline of every project in the portfolio, what was left of the budget, the number of the right order. People believed them faster than others, because memory worked as a signal: if someone remembers everything, they know the subject and you can make a decision with them.

This is changing now, for two reasons. First, data and regulations change more and more often, uncertainty grows, and there are more grey zones. A decision has to take more and more factors into account, and you physically can't hold all of it in memory any more. And if you try to memorise everything, you lose exactly the ability to think abstractly and to stay flexible. These are the biological limits of our brain.

Second, an AI assistant prepares a briefing on any metric in minutes, together with links to the sources, and everyone with a chat window open gets the same briefing. What matters more now is how the workspace is organised, whether AI has access to the right context and how good the data in that context is. So memory for numbers and facts is no longer enough on its own. You need a combination of human and machine that gives accuracy, speed, flexibility, breadth and creativity.

So memory for numbers and facts is losing its value as an advantage. At the same time, I don't think numbers and facts in your head are no longer needed. You need them to notice where AI got it wrong: a model makes mistakes smoothly and confidently, and only someone who knows what order of magnitude is normal here, and who owns the context, can catch it. But in my experience, something else gains the most value. It is the ability to think systemically and structurally, that is, to understand how facts and numbers are connected and what follows from what, and knowing what AI can do and where it can't be trusted. In other words, a leader has to be able to read what AI produced in minutes, decide what to do with it next, and understand what AI may not know.

What AI took from the "walking encyclopedia"

The reference part of knowledge went to AI first. By reference I mean everything that can be looked up: the exact value of a metric for a period, a date, a document number, the wording of a clause in a contract or a regulation. You used to go to a report for this, to an analyst or to a colleague who remembers. It took hours, sometimes days, and the person who had the answer right away saved time for everyone in the meeting. That's why they were valued.

Now this reference is available to anyone who asks, and together with the speed, the signal is gone. Memory showed that a person was deep in the work, and today anyone can sound competent if five minutes before the meeting they asked the model a couple of questions and loaded a knowledge base. It has become harder for a leader to tell from an answer who really understands the question and who just prepared a good briefing.

At the same time, AI prepares a good briefing only when it has something to build it from. If reports sit in different people's mailboxes, past committee decisions are recorded nowhere, and the current version of the budget is only with the CFO, the assistant will answer from whatever it finds, and the answer will be incomplete, even though it looks just like a complete one. So the speed of access to information is now provided by the organisation, and the memory of one person decides little here. You need to lay out the data properly, watch its quality, give AI access to it and set up how the model works: write instructions on how to check the answer and what accuracy is needed. This is the classic task of RAG projects (RAG is the approach where the model searches the company's knowledge base before answering). How well it is solved directly determines what every model answer in the company is worth.

Where a good AI briefing comes from: lay out the data, watch its quality, give access, set up the model

A similar split is discussed for the models themselves. Andrej Karpathy argues that models don't have to store all knowledge about the world inside: the thinking core can be small, and it turns to tools for the facts (the No Priors podcast, September 2024, and Dwarkesh Patel's podcast, October 2025). I see a direct analogy with a leader here: you don't have to keep everything in your head if you understand where the rest is and how to work with it.

Why you still need numbers in your head

From all this it's easy to conclude that there is no need to learn and remember any more. I don't agree. Reference numbers can really be handed to the assistant, but the anchor numbers you'll have to keep yourself.

By anchor numbers I mean the ones that show the business is in order: the order of revenue and margin, headcount, unit cost, capacity utilisation, the limits of what is acceptable. A leader usually has ten to fifteen of them. Why do you need them if the assistant gives you the briefing? To have something to check its answer against, because the model's mistakes look as neat as its correct answers. If the assistant writes that a workshop's output tripled in a month, only someone who knows that workshop's load and capacity will see it and understand that this doesn't happen. If the briefing quotes a regulation with a number that doesn't exist, only someone who has worked with that regulation will notice.

Here I'm closest to the American cognitive psychologist Daniel Willingham. He explained to teachers why thinking doesn't work without knowledge in memory: in his words, "background knowledge is decisive even in chess, which we might consider the prototypical game of reasoning" (American Educator, Spring 2009). My conclusion for a leader follows from this: without your own anchors in your head, you have nothing to check someone else's answer with.

Besides numbers, there is context that AI simply doesn't have. AI doesn't know what was agreed at the last committee, why a project was stopped last year, which contractor misses deadlines and which of the neighbouring departments really makes the decision. If this isn't in the data open to it, it won't take it into account. And digitising all the context, all relationships, decisions and communications, and also making sure of their quality, is simply impossible, as practice shows. So even with a well-built knowledge base, part of the context always stays only with people. Someone who owns the context sees that a neat answer misses the main point and can correct the decision before it goes further. That's why I call numbers and context in your head insurance.

What gains value: three things

If AI gives the briefing and memory is the insurance, then in my experience the main growth in value goes to those who can do three things.

Systems thinking. This is the ability to see how facts and numbers are connected, what follows from what, what changes next and what breaks first. Take a project portfolio. The assistant will quickly tell you that one project's deadline is slipping by a quarter. Then come the questions whose answers are nowhere written down in full in the company. What does this mean for next year's budget and for the contractor's workload? What happens to the two projects waiting for its results? What about the promise made to the owner? A human puts this chain together, and the better they see it, the earlier the company learns about the problem. The breadth I mentioned at the start is needed exactly for this: links often cross the borders between fields, from finance to production, from IT to people, from a decision in one area to consequences in another. Employers have similar expectations. According to the World Economic Forum's report on the future of jobs, analytical thinking remains the top core skill, seven out of ten companies consider it essential. And 54% of employers expect systems thinking to be used more by 2030 (WEF, Future of Jobs Report 2025).

Structured thinking. By this I mean the ability to break a task into parts and the links between them before you take it to AI. It's easy to get confused here. AI can sort ready data onto shelves by itself, and does it well: summaries, classifications and tables come easier to it than to most people. But breaking down the task, that is, understanding what exactly has to be solved, what questions follow from it, what data each one needs and in what order to look at it, remains human work. It decides what the model will be asked at all. To the question "how do we cut repair costs" the assistant will give a list of general advice. If the question is broken down (which three types of equipment cause the most unplanned downtime, what an hour of downtime costs on each, what share of repairs are repeat repairs), the model's answer becomes material for a decision. And if the question is set badly, the leader gets a neat answer to a different question and may make a decision on it without noticing the swap.

Understanding what AI can do and where it can't be trusted. The model is strong at reference on the data open to it, at a draft, a comparison and a summary of a well-described field. It is weaker where facts are few or fresh, where a number is needed that the model didn't find and so "filled in", and in everything about your company that isn't written down in the data available to it. It is also weak at understanding the real world and at the creative part of the work. Hardest of all for it is to navigate uncertainty and grey zones, where different constraints and rules apply, and they often contradict each other.

A person who understands this checks selectively: they know which place in the answer to double-check and which to trust. Someone who doesn't either checks everything and loses the time gained, or checks nothing and one day brings a beautiful mistake to the committee. Interpretation is connected to this too. AI produces material, and what follows from it for the decision and what to do with it is decided by a human. I explained in detail why the decision stays with them in the article "Why AI Won't Replace the Executive: 4 Reasons".

To make this easier to apply, here is how I would split the work between AI and a leader. In practice, this is the combination of human and machine I talked about at the start. You can hand the briefing to the assistant, but on one condition: next to every number and every document number there must be a source you can open. It's in a briefing without sources that the model makes things up most often.

Hand to the assistant (briefing, always with a link to the source)

Keep in your head (anchors)

Understand (links)

exact values of metrics for a period

order of magnitude: revenue, margin, headcount, unit cost

which metrics pull others along and with what delay

numbers of orders, contract clauses, regulations

10-15 numbers that show the business is in order

what breaks first if a metric moves by 20%

wording, dates, reference data

the limits of what is acceptable: what doesn't happen

where the feedback loops are and who owns them

summaries, drafts, comparisons

agreements, commitments, the history of decisions

how a decision in one area will echo in another

When you can't rely on an AI briefing

There are situations where the idea "the assistant will give me the briefing" works poorly or doesn't work at all, and a leader needs to know about them in advance.

Three situations when you can't rely on an AI briefing

The first: AI isn't at hand, or there is no time for it. A negotiation, a conversation with an investor, a talk, an accident on site. For a negotiation or a talk, it's worth preparing a briefing with AI in advance and keeping it at hand, it helps a lot. But in a live conversation, half a minute for a query is already too long, and the conversation often goes where the briefing didn't prepare you. If the other side names their own number or compares with a period that suits them, only someone who keeps the main numbers in their head will notice. The hardest example of this kind comes from aviation. In 2009, an Air France plane flying from Rio de Janeiro to Paris lost reliable airspeed readings at high altitude, the autopilot disconnected, and the crew couldn't keep control of the aircraft. The French investigation bureau spent three years on the crash. The final report says that an understanding of the general physics of high-altitude flight could have helped the pilots considerably. The bureau recommended checking that pilots had really understood the theory of flight mechanics (BEA, final report on AF447, July 2012). Note that it's about understanding how the system works, that is, about the right column of the table above. In business the stakes are usually lower, but the conclusion for a leader is the same: anchor numbers and an understanding of the links have to be in your head in advance, because at the moment of failure there will be no one to ask.

The second: professions where knowing by heart is the competence itself. A power grid dispatcher, a surgeon, a pilot, a lawyer in a court hearing. The idea "AI will give me the reference" can't be applied to them, otherwise it becomes an excuse not to learn what the result depends on, and sometimes someone's life.

The third situation is the most unpleasant. If AI takes over the reference, people may stop thinking about the answer and stop keeping the anchors in their head. I haven't seen direct data on anchor numbers, there is indirect data on critical thinking. Michael Gerlich surveyed 666 people and found a negative link between how often a person uses AI and their critical thinking. Younger participants depend on AI more and score lower on critical thinking. Part of this link is explained by people offloading mental work to the tool (Gerlich, Societies, January 2025). The study is correlational, and it can't prove that AI is what weakens thinking. But the conclusion for a leader doesn't change: the anchors in people's heads will have to be grown on purpose, especially in juniors. I looked at how to do this when the model takes the simple work in the article "The career pyramid is turning into a diamond". One of the moves there is exactly about this: a training mode without AI until a person has their own anchors.

What this changes in hiring and committees

If value has shifted, the way we assess people has to change too. I would stop checking who remembers how many exact values. Instead, I would check the anchors and the three things discussed above.

Four questions for hiring and committees: anchors, systems and structured thinking, understanding AI

Anchors are the easiest to check. Ask the person to name, without a briefing, the order of the three or four main numbers of their area: revenue or budget, headcount, the main efficiency metric. Accuracy to the percent isn't needed here. What's needed is the order of magnitude by which the person will notice a mistake.

For systems thinking, the question "what happens if this metric moves by 20%?" works. Someone who remembers the numbers will name the current value. Someone who understands the system will say what changes next, how soon and what breaks first. It's good if they also name what they don't know.

For structured thinking, the request "break this task down" works. Give a candidate or a presenter a real task and ask them to say what exactly needs to be solved, what data will be needed for it and in what order they will look at it. From the answer you can see right away whether AI will get a good question from this person.

For understanding AI, the simplest test works. Show the person a model's answer on their topic and ask what they would double-check here and why. A strong candidate will find a couple of places where the model could be wrong or not know something. A weak one will either accept everything or say that AI can't be trusted at all.

The same works in committees. Anyone can gather the numbers now, so from a presenter you should expect the links: what changes next, which decisions depend on this number and what the briefing doesn't take into account. And a leader should ask themselves separately where the briefing came from and whether AI saw the current data. If not, you have to start with the workspace, not with the people.

Where to start this week

Take a sheet of paper and write down 10-15 anchor numbers of your business or your area, without the assistant and without reports. Next to each one, write what it is connected to and what changes if it moves by 20%. Where there is no answer, you have a gap: either in the anchors or in understanding the links. Gaps in the anchors show up right away, and regular work with your own reports, without go-betweens, closes them. Gaps in the links are closed by talking to the people who run the neighbouring areas, and the assistant won't do this work for a leader. Then ask AI to put together a briefing on the same numbers and compare. The differences will show one of three things: the model lacks context and you need to open it up, the model made a mistake or filled something in, or you are the one who is wrong, and that's useful to know too.

Exercise for the week: 10-15 anchor numbers, links, comparison with an AI briefing

Further reading

More on how a leader can work with AI in practice is in Dzhimsher Chelidze's book "Artificial Intelligence. Freefall". Download the book for free.

bottom of page