Article

Why AI should work on the business, not just in it

AI can do more than perform existing work. It can help organisations examine how work actually happens before deciding what should be changed or automated.

Most organisations are approaching AI by asking what it can do for them. It is an obvious place to start. Generative AI can draft documents, summarise meetings, analyse spreadsheets, search information, produce reports, answer questions and automate an increasing range of tasks. As agentic AI develops, these capabilities will extend from assisting with individual activities towards taking responsibility for longer sequences of work.

There is considerable value in this, but it also tends to frame the problem in a particular way. We have some work that is already being done and we ask how AI might help us do it more quickly, cheaply or effectively. In doing so, we make an assumption about the work itself: that it should continue to be done broadly as it is now.

I think there is another use of AI that organisations should consider. Instead of only asking AI to perform the work, we can use it to help us examine the work.

This gives us a useful distinction: AI in the business performs work; AI on the business helps us understand the work.

Both are useful. At present, most of the attention is on the first.

Starting with the work rather than the technology

Consider a monthly management report. Someone has to gather the information, which may come from several systems. Some of it may arrive by email, and there may be a spreadsheet that somebody maintains because one of the systems does not quite provide what is required. Figures are checked, missing information is chased, and eventually the report is assembled, reviewed and distributed.

The obvious AI question is whether some of this can be automated. Can Copilot find the information? Can an AI system summarise it? Can it draft the report? Could an agent collect the inputs?

These are all reasonable questions, but they assume that the report, and the process that produces it, should continue broadly as they are. There is another set of questions we could ask first. Why is the report produced, who actually uses it, and what decisions does it support? Where does each piece of information originate, and why does somebody have to transfer information from one system to another? Why are the figures checked? What happens when information is late? Which parts of the report require judgement, which parts are actually read, and what would happen if we stopped producing some of it?

Once we start asking these questions, we are no longer looking for ways to apply AI to the process. We are trying to understand the process itself, and AI can help us do that.

There is also a less obvious reason for doing this before we automate. AI may make poorly designed work easier to tolerate.

Friction in a process is usually undesirable, but it has one useful characteristic: people notice it. If producing a report consumes a day every month, if employees repeatedly have to chase missing information, or if somebody spends hours copying data between systems, the effort itself provides evidence that something may need attention. People complain about it, managers see the cost, and eventually somebody may ask why the work operates in this way.

AI can remove some of that visible friction without addressing its cause. If a report that previously took a day to produce can be generated in a few minutes, that is an immediate productivity improvement. But if the report is largely unnecessary, we have also made an unnecessary activity cheap enough that nobody may question it again. The same applies to an awkward information transfer, an unnecessary reconciliation or a poorly designed approval process. AI may become very good at performing work that the organisation should have stopped doing.

This creates a slightly different risk from simply automating a bad process. As the cost of performing inefficient or unnecessary work falls, some of that work may become harder to see because we have removed the symptom that might otherwise have prompted us to investigate the underlying system. The ability to perform work more efficiently therefore increases, rather than reduces, the need to understand why the work exists in the first place.

How work actually happens

Organisations have formal processes, but people also develop ways of getting work done around them. A system does not provide quite the right information, so somebody creates a spreadsheet. Two systems do not communicate, so somebody copies information between them. A mistake occurred several years ago, so an additional check was introduced. A report has been produced every month for so long that nobody is quite sure who originally asked for it. An unusual case requires a workaround and, as that unusual case becomes more common, the workaround gradually becomes part of the normal process.

This is not necessarily evidence of poor management. People adapt to the organisations in which they work and, quite often, these adaptations are what keep the organisation functioning. The difficulty is that, over time, the process described in a procedure, process map or management meeting can become quite different from the process that people actually perform.

This creates a problem when we begin with automation. If we do not understand how the work operates, we cannot know whether automating part of it is a sensible intervention. We may automate a workaround or make an unnecessary activity faster. We may preserve a control that no longer serves a purpose, or remove one without understanding why it exists. We may optimise one part of a process only to move the problem somewhere else.

We therefore need sufficient understanding of the work before deciding how to intervene, while recognising that intervention itself will often reveal more about how the system operates.

Using AI to interrogate the work

Generative AI gives organisations another way of developing that understanding. An AI system can examine process descriptions, documents, spreadsheets, procedures, meeting notes and other information associated with a piece of work. It can compare different accounts of the same process, identify gaps in a description, ask questions, surface possible inconsistencies and generate hypotheses that can then be investigated.

This does not require the AI to know how the organisation works. In fact, it is better if we do not pretend that it does.

Suppose I describe a purchasing process to an AI system. Rather than asking it to redesign the process, I can ask it to question me about it. Where does the process begin and what information is required? Who supplies that information, and what happens if it is incomplete? Who makes the decision to purchase? What determines whether further approval is required? Where is the information recorded, and does anybody enter the same information elsewhere? What happens when the normal process cannot be followed? Which decisions require judgement, and where does work tend to wait?

I will be able to answer some of those questions. Others may require me to speak to somebody else, and some may expose things that I have assumed but never checked. Documents or data may provide further evidence, while the AI may identify contradictions between what I have described and what the available evidence suggests.

At this point AI is not performing the purchasing process, nor is it deciding how the process should be redesigned. It is helping us construct a better understanding of the work.

That is what I mean by using AI on the business.

The ideas behind this are not new

There is a substantial history of trying to understand organisations through their processes and systems. Systems thinking encourages us to examine relationships and the behaviour of the whole rather than optimising isolated activities. Lean places considerable emphasis on understanding work and the sources of waste and variation, while business process management and business process re-engineering provide ways of analysing and redesigning processes.

Process mining goes further by using operational data to reconstruct how processes actually behave rather than relying entirely on descriptions of how they are supposed to behave. Work on digital twins of organisations extends the idea again, creating representations through which organisational structures and processes can be analysed and, potentially, simulated.

Generative AI does not replace any of these approaches. What it changes is the ease with which people can begin asking questions of organisational information. The ability to examine a collection of documents, compare accounts, structure unstructured information, generate questions and challenge assumptions is increasingly available through general-purpose AI systems.

This lowers the threshold for organisational inquiry. An SME does not need to construct a digital twin of itself before it can begin examining how work moves through the business; it can start with the people who do the work and the evidence it already has.

This is a much more modest proposition than claiming that AI can diagnose an organisation, but I think it is also a much more useful one.

The objective is not to find somewhere to use AI

There is a potential trap in carrying out an AI opportunity assessment. If we begin by looking for opportunities to use AI, we are likely to find them. The technology has such broad capabilities that almost any sufficiently complicated process can be made to look like an AI opportunity.

An AI opportunity and a business improvement opportunity are not necessarily the same thing.

Return to the monthly report. Suppose we discover that an employee spends four hours each month transferring information between two spreadsheets. We could ask AI to perform the transfer, automate it using conventional software, integrate the systems, change where the information is captured, or redesign the report so that the information is no longer required. We might even discover that nobody uses that part of the report and stop producing it.

The problem should determine the intervention, rather than the available technology determining the problem.

AI on the business should therefore be intervention-neutral. Its purpose is not to discover where more AI can be deployed, but to understand where the organisation can work better. The eventual intervention may involve AI, conventional automation, process redesign, better data, clearer responsibility, standardisation or training. It may involve removing something entirely.

AI is working with a representation of the organisation

There are some obvious limitations. An AI system does not observe an organisation simply because we give it some documents. A procedure is a representation of work, as is a process map or a manager’s account. An employee’s account is also a representation, albeit from a different perspective. None should automatically be treated as the work itself.

This becomes particularly important with generative AI because a plausible explanation can easily be mistaken for an accurate one.

Suppose an AI system sees three approval stages in a process and concludes that they represent unnecessary bureaucracy. Perhaps they do, but perhaps one exists because of a regulatory requirement, another was introduced following a fraud, and the third compensates for poor information earlier in the process. Or perhaps nobody knows why they are there.

The AI cannot resolve this simply by reasoning about the process description. We need evidence.

This suggests a useful separation between evidence, interpretation, hypothesis and recommendation. If a system record shows that an activity took place, that is evidence. Our explanation of why it happened is an interpretation; a possible underlying cause is a hypothesis; and what we think should change is a recommendation. AI can help with each of these, but they are not interchangeable.

Used badly, AI could take an inaccurate description of an organisation and produce a very sophisticated analysis of something that does not actually exist. Used properly, it can help us identify what we do not yet know.

Organisations are more than processes

There is another limitation because organisations are social systems as well as operational ones. Authority, trust, incentives, professional identity, history and informal relationships all affect the way work is performed.

A process may appear inefficient when considered purely as a flow of information, but the behaviour may make sense once the organisational context is understood. People may maintain their own records because they do not trust the central system. A manager may insist on an approval because they remain accountable for an outcome over which they otherwise have little control. A team may resist standardisation because apparently similar cases require professional judgement that is not visible in the process description.

These things are difficult to understand from documents alone. AI can help us ask questions about them, compare different accounts and identify areas that require investigation, but it does not remove the need to talk to people, observe work and exercise judgement.

AI can accelerate organisational inquiry; it cannot substitute for it.

From work as imagined to work as understood

I find it useful to think about this as moving from work as imagined towards work as understood.

Work as imagined is our current representation of how the organisation operates. It is contained in procedures, systems, process maps, management assumptions and people’s individual understanding of their roles. We then gather evidence of the work itself, which might include documents, operational data, examples, observation and accounts from the people involved.

AI can help us interrogate that evidence by identifying gaps, comparing accounts, suggesting questions and surfacing patterns that require further investigation. We can check these with the people doing the work and against further evidence, gradually developing a better account of what is happening and why.

We will never create a perfect representation of a complex organisation, nor do we need one. We need enough understanding to make a better decision about what to do next, while recognising that any intervention gives us further evidence about the system. Understanding and improvement are therefore not separate phases so much as a continuing process in which each can inform the other.

The question nevertheless changes from where can we use AI? to what is preventing this work from performing better?

The answer gives us a much better basis for deciding what should change.

This can start with one piece of work

None of this requires an organisation-wide AI programme. We can start with one recurring piece of work: preparing a management report, onboarding a customer, processing an order, responding to an enquiry, approving expenditure or scheduling production.

Instead of immediately asking AI to automate it, describe the work and provide appropriate supporting material. Then ask the AI what it does not understand. Ask where information comes from and where it goes, where people wait, where information is entered more than once, where judgement is exercised and what happens when the normal process fails. Ask which activities appear to compensate for problems elsewhere, which controls have a clear purpose, and what evidence would be needed before concluding that an activity should be changed or removed.

Then investigate the questions by talking to the people doing the work, looking at examples, checking the data and correcting the assumptions. As the picture develops, AI can help test explanations and identify further questions rather than prematurely turning incomplete understanding into recommendations.

Try this with a piece of work in your organisation

Choose one recurring piece of work that you know reasonably well. It might be preparing a report, processing an order, onboarding a customer, responding to an enquiry or approving expenditure.

Describe how you believe the work currently operates and ask your AI system to help you understand it before suggesting how it might be improved.

Ask it to question you about the people, information, systems, handoffs, waiting, duplication, rework, workarounds, exceptions, controls and points at which judgement is required. Ask it to distinguish between what you know, what you believe to be true and what still needs to be checked against evidence or with the people doing the work.

Most importantly, tell it:

Do not recommend improvements or automation yet. Help me understand how the work actually operates first.

For a more structured version, try the full AI on the Business — Organisational Inquiry Prompt v1.0.

The AI is not the consultant, process owner or decision-maker. It is another capability available to them, and its value comes from the interaction between AI capability, organisational knowledge, evidence and human judgement.

What happens as AI becomes part of the work?

There is a further development to consider. At present, AI on the business might mean using a general-purpose AI system to interrogate a process or examine organisational information. More sophisticated approaches can incorporate operational data, process mining, organisational models and simulation.

At the same time, AI agents are beginning to participate more directly in organisational work. This creates an interesting relationship between AI in the business and AI on the business.

An AI agent may participate in a process and, in doing so, generate evidence about the work. That evidence can be examined alongside the activity of people and other systems, and the resulting analysis may suggest changes to the process, to the role of the agent, or to the relationship between human and artificial work.

This introduces a different set of management questions. We need to decide what should be delegated to an artificial agent, what authority it should have, how its performance should be observed and when it should escalate to a person. We also need to distinguish between an agent performing badly and an agent operating within a badly designed process, and determine who remains accountable for the outcome.

These are not simply questions about AI technology. They are questions about how organisations are designed and managed when artificial agents become participants in work, and we will need to address them increasingly as AI moves from being a tool used by employees towards becoming an active component of organisational systems.

Ask AI a different question

There is enormous scope for using AI within organisations. We should use it to remove repetitive work, increase capability, improve access to information and support people in making better decisions.

But we can also use it before we get to that point.

Before automating a report, we can ask why the report exists. Before accelerating an approval, we can ask why the approval is required. Before automating the transfer of information between two systems, we can ask why the information needs to be transferred at all. Before using AI to make a process faster, we can use AI to help us understand the process.

This is the distinction I want organisations to make.

AI in the business performs work. AI on the business helps us understand the work.

The first can improve how existing work is performed. The second can help us decide whether it is the right work in the first place.

Further reading and intellectual foundations

The distinction developed here sits within a much longer history of research and practice concerned with understanding and redesigning organisational work. Readers wishing to explore that background further may find the following useful:

Reveal how your work
really works.

A first conversation is free, and usually enough to tell you whether there's a leverage point worth pulling.

Start a conversation