Perspective
What AI can actually do in a business today
Public debate about artificial intelligence revolves around models and vendors. Inside a business the more interesting question is a different one: what part of this can work independently today, and what does that mean for a routine a person has run by hand until now.
Last updated: August 2026
01
Three stages of build-out
Almost every discussion of AI in a company conflates three things that have little to do with each other technically or commercially. It is worth separating them - not least because most businesses sit on the first stage and take it for the end of the road.
- Stage 1 - the assistant at the desk
- A person asks, the model answers. It sees what is in the window and is finished as soon as the window closes. This lifts the productivity of individuals and is, through Microsoft and team subscriptions, already paid for almost everywhere.
- Stage 2 - the process that runs by itself
- An event sets it off: an email, a document, a date. It reads from your systems and writes back into them, logging as it goes, within a defined scope of action. A person releases whatever is binding. The difference from the first stage is not model quality but the question of who starts the process.
- Stage 3 - the system that spans departments
- Several routines interlock. Agents work on a task over hours and days, fetch what they need along the way, and come back when a decision is due. People lead and supervise instead of typing.
The test for which stage you are on is simple: does anything happen when nobody is in the office?
02
Why the models made the jump
Until recently the honest answer to "can this run without supervision" was usually no. A model could write a text but not hold a task across several steps: it lost the thread, invented intermediate results, or broke off at the first unexpected turn.
That has shifted. Today's models hold a task across many steps, use tools while doing so, check their own intermediate results and work over periods measured in hours rather than seconds. Anthropic has described multi-hour to multi-day autonomous work on a single task for the current models - and that stamina is exactly the difference between an assistant and a process.
In practice this means the limit rarely runs where a business assumes it does. It does not run at "can the model do this" but at "can we reach the data" and "does the effort pay".
03
What separates an agent from a tool
The right-hand column describes something closer to a role in the business than to a piece of software. That is not a metaphor but the practical consequence: an agent gets an assignment, a scope of action and a release threshold - the same three things a new colleague gets.
And it has an uncomfortable flip side that belongs to an honest account: whatever acts independently can also be independently wrong. Which is why every project of ours sets out in writing what the system does alone and what a person releases, and why anything binding and anything irreversible requires a release as a matter of principle.
04
The mid-market fallacy
A sentence that comes up regularly in first conversations: "That would not work here, we are set up too traditionally." What is usually meant is a system landscape that grew over time, a lot of paper, a lot of experience held in people's heads and few documented processes.
But that is precisely the starting position with the greatest leverage. A business where everything is already digitised and cast into clean interfaces has long since taken the easy efficiency gains. A business that grew organically still has them ahead of it - and the new models are strong at exactly what classic automation failed at there: unstructured records, inconsistent formats and rules nobody ever wrote down.
The second part of the fallacy is a matter of timing. Anyone founding a company today builds with these tools as a matter of course. Sooner or later the incumbents therefore compete with businesses whose cost structure looked different from day one. That is no reason to rush, but a good reason to start now rather than in three years.
It is not the digitised business that has the greatest potential but the one that grew. That is where the manual work sits.
05
Where the limit actually sits
- Not at the model
- What a routine demands in substance - read, compare, classify, formulate, write into a system - sits within what is possible today for the overwhelming majority of processes.
- At data access
- Whether we can reach the necessary data, in sufficient quality and with permissions that can be reproduced cleanly. That is the point at which projects tip over, and it can only be settled by trying.
- At repetition
- A routine that occurs four times a year does not carry the effort - however irritating it is. The benefit comes from repetition.
- At unsettled responsibilities
- If it has not been decided internally who is responsible for what, or how a case should be handled, a system cannot apply a rule that does not exist. That question then comes before the project.
Which is why every project here begins with an audit and not with a quotation. Not because the technology is uncertain, but because the last three points sit differently in every business and cannot be answered from a desk.
Frequently asked
What can AI take on in a company today?
Considerably more than boilerplate text. A system today can pursue a task across many steps, reading from the ERP, the mailbox and the file store while doing so, write results back and come back when a decision is due - over hours to days. The limit rarely sits at the model but at data access and at whether the effort pays at your rate of repetition.
What is the difference between an AI assistant and an agent?
An assistant is operated by a person and sees only what is in the window. An agent is set off by an event, pursues a goal across several steps, fetches what it needs along the way and writes back into systems. The difference is not model quality but who starts the routine.
Is our business too traditional for AI?
Rather the opposite. Businesses that grew organically still have the easy efficiency gains ahead of them, and the new models are strong at exactly what classic automation failed at there: unstructured records, inconsistent formats, unwritten rules.
Do we have to tidy up our processes first?
No. A model copes with filing that grew over time considerably better than a classic interface does. What does have to be settled is who is responsible for what - a system cannot apply a rule the business does not have.
Does a system like that then decide on its own?
Within a defined scope of action yes, outside it never. Anything binding and anything irreversible goes past a person; that is in the specification, not in the conversation.
Tell us where your time goes.
In the audit we record the process, work out what is worth automating, and you get a plan with fixed prices - credited against the build if you go ahead.