Where does AI fit in a business?

Two rooms, on the same morning

A friend of ours is a personal assistant to an executive at a solar company in London. He is enthusiastic about AI, in the way a lot of senior people currently are, and in a development meeting he asked the room the reasonable question: where can we use this to help us day to day?

She had no answer. Not because she is bad at her job — she runs his entire working life — but because the honest response is “I don’t know. Accounts? Chasing people for billing?” That is where the conversation dies, in thousands of companies, every week.

The same week, researchers were using Anthropic’s Claude Science — a workbench wired into more than 60 scientific databases, with toolkits for genomics, proteomics and cheminformatics — to work through literature and analysis at a scale no individual could read their way through.

Two rooms, same underlying tool, and the gap between them is not capability. It is that nobody has ever shown the person in the first room what the thing actually is.

What actually changed

Strip away the product launches and one sentence is left:

Think about what that constraint quietly decided for you. Which contracts get reviewed properly and which get skimmed. Whether anyone reads all the customer complaints or just the loud ones. Whether the invoices get checked against the delivery notes, or spot-checked, or neither. None of those were decisions anyone made on the merits. They were the shape that human attention forced, and they have been sitting in your operation so long they look like policy.

That constraint has moved. Almost nobody has redrawn anything around it — which is the actual opportunity, and it is much larger than a faster inbox.

Altitude one: it does your typing

Work somebody does today, mechanically. There is a simple test for it, and it takes ten seconds to apply to any job in your business: does a person read something, and then type something? If yes, it is here. If the hard part is a decision nobody in the building can explain, it is not.

ProLend is a private lending firm in South Africa. Its lenders finance body corporates and homeowners’ associations, it has deployed R2.7 billion, and it is currently rebuilding its brand and its distribution. It is a useful company to follow through this series because its product is essentially documents — which means if you run a practice, an agency, a brokerage or a freight company, you will recognise your own week in it.

Ten minutes of looking found this: one of its funding providers publishes its products as a single enormous PDF — six products, each with rates, minimums and maturities, laid out for a printer rather than for a person. Someone opens it, reads it, and types those figures into the systems that need them.

Read something, type something. It happens every time a provider moves its rates, and nobody had ever thought of it as a task worth naming. Work like this is invisible precisely because it is routine. Nobody writes a process document for it. It just eats an afternoon, quietly, forever.

This altitude is real and worth having. It is also the smallest of the three, and the mistake almost everyone makes is stopping here.

Altitude two: the work nobody could afford to do

Now the one almost nobody writes about, for a structural reason: you cannot automate a task that has never been performed. There is no incumbent doing it slowly, no process document, no line in a budget. It does not show up in an audit of how people spend their time, because nobody spends any time on it.

Examples, in any business:

• Read every contract the company has ever signed and list the ones that auto-renew, with dates.
• Read three years of support email and report what customers actually complain about — not what the management team believes they complain about.
• Check every invoice against every delivery note since 2023.
• Read every quote that was lost and find what the lost ones have in common.

None of that happens today. Not because it is difficult — it is tedious rather than difficult — but because at human reading rates the answer costs more than it is worth. Every one of those is now affordable, and each returns something the business has never had: not a saving, an answer.

So the question to walk around your own company asking is not “what can we automate?” That question can only find work you are already doing. It is:

At ProLend, that question surfaced something uncomfortable and completely ordinary. The prime lending rate — the number half the product line is priced from — existed in four separate places: the calculator, the proposal builder, the brochure and the marketing pages. Four copies, maintained by hand. They had drifted. One of them was quoting a rate that was nine months out of date.

Nobody was negligent. Checking four copies of a number across four systems is exactly the kind of task that never makes it onto anyone’s Monday, forever. It was found in minutes once someone thought to ask. Every business has a number living in four places.

Altitude three: work a person cannot do at all

The third altitude is not a faster version of anything. It is searching a space too large for a human being to hold in their head: the whole literature of a field, every known compound, every combination of a hundred variables.

This is where the scientific work sits, and it is worth being precise about what is and is not being claimed. The tools do not make discoveries. Researchers do. What has changed is the amount of ground one researcher can cover, and the fact that the work stays checkable while they cover it — Claude Science ships a reviewer that independently audits citations and calculations before anything reaches publication, and bundles the code, the environment and the full method with every figure it produces.

You are probably not doing drug discovery. This altitude belongs in a series for business owners anyway, for one reason: it tells you where the ceiling is. A person who believes the ceiling is their invoice admin makes small decisions for years.

The same tool at three altitudes

The continuum is the point of this entire series, so here it is plainly. The reason a workbench can help a researcher through work she could not read by hand is the same reason it can read a rate table in Johannesburg: it reads and reasons over more text than a person can, without getting bored on page four hundred.

And here is the detail that convinced me the continuum is real rather than rhetorical. Claude Science puts a separate reviewer over the work, auditing citations and calculations before publication. ProLend, at the other end of the scale, now refuses to publish a prime rate outside 5–20% — a gate that exists solely to catch a number that could not possibly be a real interest rate.

A research platform and a lender in Johannesburg arrived independently at the same idea: the check has to be a separate thing from the doing. Both got there because at this speed neither had a choice.

What the machine does, and what you still do

This gets misrepresented in both directions, so, using the rate table:

What the machine does: reads the PDF, pulls out six products with their rates, minimums and maturities, and hands back figures a system can use. Minutes rather than an afternoon, and it does not lose concentration on product five.

What a person still does, permanently:

• Decides which document is authoritative. No machine knows the PDF emailed in March was superseded in July.
• Checks the extracted figures against the source. Every time.
• Owns the number. When a rate reaches a client, a person is accountable, and that never moves.
• Decides what it all means commercially — which product to lead with, what to do when a provider moves.

That second list is longer than most people expect, and it is the honest answer to the question underneath the question. The work is not doing itself. The typing is doing itself. The judgement is staying exactly where it was.

Try this on your own week

Fifteen minutes, no software, nobody’s permission required. Walk through your own operation and ask three questions in every room:

1. What typing happens here? Somebody reads something and types something else. Write it down, with roughly how long and how often.
2. What has nobody ever been able to afford to check here? The review that gets sampled instead of done. The pile nobody has read end to end.
3. What would we ask, if reading were free? The question you would love the answer to and have never once considered asking, because it would have meant somebody reading for a month.

Most people are surprised twice: by how long the first list is, and by the fact that the third question — the one that felt like a fantasy — is the one worth the most money.

What if it is not your decision to make?

This is the common case and almost nobody writes for it. You can see the opportunity; you cannot authorise anything. The move is not to lobby — it is to hand the person who can authorise it two pages containing four things:

1. The job. Named, with who does it and roughly the hours a month it eats.
2. What would change. The typing goes; the checking stays; here is who still owns the outcome.
3. What it costs to find out. Not the full build — the cost of a real answer.
4. What it does not fix. The limits, stated by you, before anyone else raises them.

Two pages, because more will not be read. And there is an irony available to you: draft it with an LLM. Give it your three lists and ask it to make the case for the item at the top. If the tool cannot help you argue for itself, that is worth knowing before anyone spends money.

What this series is actually about

One claim, made now and earned over the next eight parts:

The speed is not the story. The discipline the speed forces is the story.

At ProLend a wrong interest rate reached a client-facing template in under a minute — a product quoted at prime plus three when it should have been prime plus two and a half. That is a different product. The fix was not “be more careful”. It was structural: one source of truth for every rate, versioned so it cannot be quietly edited, and the gate that refuses a number no real interest rate could occupy.

Those were always good practice. What changed is that the gap between a mistake and a client seeing it collapsed to nothing, so they stopped being optional. Part eight is about exactly that, and it does not get softened.

Have you got one of these?

The most useful thing you can send us is not a question — it is a job. One task from your own week: what you read, what you type, how long it takes, how often. Or better, the second kind — the thing you would check if checking were free. We will take it apart properly and publish the analysis, and the ones that teach something become future parts of this series.

No pitch attached, and you do not have to be anyone in particular. The person who best understands where a business loses its afternoons is almost never the person with the budget.

Next in this series

Part 2 — What comes in the door. Email, enquiries, forms and supplier documents: the highest-volume, lowest-glamour paperwork in any business, and the first place all three questions land.