Skip to searchSkip to main content
// the interface problem

Two Boxes.
Two Different Brains.

A familiar chat box, an old search habit, and two practical moves for leaders who want AI to do real work instead of expensive guessing.

// box one — search
ai adoption framework
Ten tabs, ranked
Two of them flatly disagree
You decide which to believe
retrieves · ranks · leaves the thinking with you
// box two — chat prompt
ai adoption framework
one answerno sourcesno rankingzero tabs open
generates · fluent · singular
// twenty years of training

"Just google it" became a full sentence

If you are of a certain age, you can still hear the dial-up tone. That whine and clatter as the modem shook hands with the world is one of those sounds you cannot unhear, and for me it came with a specific memory: lying on my bed with a laptop, a 20 metre LAN cable running up the stairs from the landline, one pin on the connector broken, quietly hoping none of my siblings would trip over it on a late night stroll to the kitchen. On the other end of that cable was Yahoo! Chat and a world that felt completely foreign. Then Google arrived, and within a few years it had changed how we spoke to each other. "Just google it" became a full sentence, a piece of advice and occasionally an insult.

That is roughly 20 years of training. Type a few words, scan a page of blue links, open the promising ones, decide which source you trust. We got so good at it that we stopped noticing we were doing it.

Then, at the end of 2022, ChatGPT arrived free and open to anyone, much as Google once had. It presented a rectangle with a cursor blinking in it. Our brains, conditioned by two decades of practice, looked at that rectangle and thought: familiar, I know what to do here.

I put two empty boxes on the screen: a search bar and a chat prompt. Same shape, same cursor, same instinct to start typing.

From an MBA and MPhil guest lecture at GIBS, South Africa.

// same interface, different machine

What each box is actually doing

01

What the first box does

Search retrieves documents and ranks them. That is the whole job, and it leaves most of the thinking with you. You open ten tabs, you compare what they say, you notice when two of them flatly disagree, and you decide which one to believe and why. You know where every claim came from, because you were the one who went and got it. The messiness is doing something useful: those ten tabs are the visible evidence of a judgement you are in the middle of making.

02

What the second box does

A language model is not looking anything up, unless you deliberately give it something to look at. It generates an answer from patterns compressed during training, and what comes back is fluent, well-structured and singular. One answer. No trusted sources, no ranking, nothing to indicate that three credible people would have answered differently. Zero tabs open.

Search is the thing that retrieves. Generative models generate. Retrieval only enters the picture when you attach documents, connect a system or use a tool that grounds the answer in real sources. Getting that the wrong way round in a meeting, project, or boardroom is an easy way to lose your audience.

// what fluency hides

Two boxes, two different brains, one user interface

And fluency is the part that catches us out, because an answer that reads well feels true, and that feeling tells you almost nothing about whether it is. For anyone who signs off decisions, the absence of visible disagreement is the thing to watch.

Weighing conflicting sources is where your judgement earns its keep, and a single confident paragraph quietly removes the occasion to use it.

// two practical moves

Two moves that change the work

Move 01

Write a brief, not a query

The habit that changes this is small. Treat every prompt as a brief rather than a search string.

A brief says what the output is for and who will read it. It attaches the material you actually trust, which might be the policy document, the board pack, last quarter's numbers or the three research papers you have already vetted. It asks the model to show which source each point came from, and to say where those sources disagree with one another. It tells the model what good looks like, in the way you would tell a competent analyst who has never met your organisation.

You will notice the answers get longer, more hedged and less impressive. That is the point. You have put the disagreement back on the screen where you can see it.

// anatomy of a brief
Purpose and reader
What the output is for and who will read it.
Material you trust
The policy document, the board pack, last quarter's numbers, the papers you have vetted. Something trusted that you leverage within your Workspace or 365 account.
Sources and disagreement
Show which source each point came from, and where they disagree.
What good looks like
Said as you would say it to a competent analyst who has never met your organisation.
Move 02

Map one process before you subscribe for a license

Map one process before you deploy one AI tool or sign up for Copilot, Claude or ChatGPT Enterprise.

Work through the Four Levels of Leverage*, the model I use for how organisations actually adopt AI.

01

Assistance

A person uses AI on their own tasks and checks every output; the work still gets done without it, just more slowly.

02

Repeatability

Two different people, using the same input, produce materially the same quality.

03

Delegation

AI owns a bounded piece of work and people review the exceptions.

04

Transformation

The business does something it could not do before, and the org chart or the P&L would look different as a result.

// the attempted skip

The levels are cumulative, and most failures I see are an attempted skip. The common one is Assistance straight to Delegation, which in practice means automating a process nobody ever wrote down. The steps live in three people's heads, the data is patchy, and nobody is certain who decides when something unusual arrives. Put AI on top of that and the confusion arrives faster, with an audit trail that nobody can follow.

Delegation needs a documented process, data you trust, clear decision rights and an audit trail. No licence supplies those, and no vendor will tell you they are missing.

// the exercise

So the exercise is this. Pick one process your team runs every week. Write down each step, who makes each decision, what data it draws on, what happens to that data (manually) and where the exceptions go. One page is usually enough.

If you cannot finish that page, the process is not ready for AI to run, and you have just found the real first piece of work.

// the same discipline

Say what you want, what it is for, and what it should be built from

Both moves come down to the same discipline. The interface has become familiar enough that we forget to say what we want, what it is for and what it should be built from. When we do say those things, whether to a model or to our own teams, the work improves immediately.

Run the mapping exercise with your team

If you would like to run that mapping exercise properly with a team, that is a conversation I am always happy to have.

// faq

Questions leaders ask

No. It is an argument against using them the way we use a search bar. The two moves — write a brief, map one process — are about getting real work out of the same tools you already have.

A licence puts you at Assistance. Delegation needs a documented process, data you trust, clear decision rights and an audit trail. No licence supplies those, and no vendor will tell you they are missing.

One your team runs every week. Frequency matters more than importance for the first pass, because the steps, the decisions and the exceptions are all still fresh enough to write down honestly.

Reach out directly. We will pick one process, work through the page together, and see which of the four levels your organisation is actually ready for.