AI is moving fast. Why has adoption been slow in the everyday economy?

Roughly one business in five uses AI at all. Working with more than 250,000 businesses gives us a view of what holds adoption back. The gap is not just about what models can do or what owners want. It is a distribution problem, and closing it starts with making AI work inside the technology these businesses already rely on. 

Every major survey of business AI adoption now lands in roughly the same place. The European Commission's State of the Digital Decade report, published in August, puts AI use at just under 20% of EU enterprises. The US Census Bureau's survey of 1.2 million businesses found AI adoption at between 17% and 20%. 

Two patterns sit underneath those headlines, and they matter more than the headlines do.

The first is size. In the US, 37% of firms with 250 or more employees use AI, compared with fewer than 20% of those with four or fewer. Recent increases in adoption have been concentrated among firms with at least 20 employees, with little change among smaller businesses. A separate Census analysis reinforces the point: only 18% of firms used AI, but those firms accounted for 32% of employment. The businesses that make up the everyday economy, with a dozen staff and a system of record they have used for years, are the ones standing still.

The second is depth. The ONS asked adopters how extensively they use AI. One in ten said extensively. The rest have it somewhere in the business, most often a language model drafting text, while the work that actually runs the business carries on as it did. Put those figures together and only around three or four businesses in a hundred are using AI extensively. 

Our own view is that the headline numbers flatter the picture. Circeus companies work with, and talk to, more than 250,000 businesses regularly, and what we see from inside them is that genuine implementation, meaning AI that has changed how a workflow runs rather than how an email gets written, is closer to four or five per cent, roughly one in twenty. Re-engineering a complex workflow around it is rarer still.

That is the gap this piece is about, and it is enormous. Two explanations are usually offered for it. Neither explains the gap on its own. 

The models are ready

One useful measure of what a model can do is the complexity of the work it can complete on its own. METR, an independent evaluator, measures this through task length: how long a human expert would take to complete work that a model can finish successfully half the time. Its tests focus mainly on software and technical tasks. In early 2023, that meant a few minutes of human work. More recent models can handle tasks requiring hours, with the measured horizon doubling roughly every four months since 2023.

The practical applications are increasingly concrete. With access to the right systems, a model can read a supplier invoice, match it to the purchase order, flag a price variance, and draft a query. It can take a support ticket, pull the order history, diagnose the likely cause, and write a reply. It can help reconcile a bank feed, classify exceptions, and prepare a journal entry for someone to approve. Given the data and a clearly defined job, these are practical uses of AI today.

METR adds a caveat that matters when bringing this capability into a business. A task a model completes successfully half the time is not one you can simply hand over unsupervised. Requiring greater reliability shortens the task horizon, and where mistakes are expensive, such as issuing a refund, submitting a filing, or making a payment, the checks must reflect the consequences. Reliability depends on the system around the model as well as the model itself. The data has to be right, the actions have to be bounded, and the checks have to be built in. Which is why how AI is built into a business's everyday software matters as much as what the model can do.

The owners are willing

The second explanation is that small business owners are slow to change. In reality, the opposite is often true. Owners adopt anything that saves them an hour, immediately, provided it costs them nothing to adopt. What they will not do is spend a week becoming systems integrators, and a great deal of AI, as it is currently sold to them, asks for exactly that.

The Commission's own list of what holds SMEs back is skills, data access, infrastructure and resources. Each of those is an integration problem rather than a question of appetite.

A 40-person distributor has no data team, no integration budget, and no spare afternoon to wire a model into the order system. A chat window sitting beside the workflow does not remove that work. Someone still has to find the data, paste it in, check what comes back, and paste it into the system where it belongs.

For the owner, that is a new job, not a saved one. So the pilot stalls, and the numbers stay where they are.

Adoption is a distribution problem

The useful lesson from earlier waves of technology is not that adoption takes a fixed number of years. It is that access to a capability and changes in how people work are different things. AI can reach businesses through the computers, cloud services and subscriptions they already use. That can make access fast. Redesigning the work around it is a different process.

Cloud offers a useful example of this route. Small businesses did not adopt cloud services because they were persuaded of its merits. They adopted it because the accounting package, the till and the booking system moved to the cloud and took them along. The software they already paid for, and already trusted with their data, changed underneath them. Nearly half of EU enterprises now run on cloud, more than double the share actively using AI. For most, the shift to cloud came through software they already used.

AI can reach the everyday economy through the same route, but with more active user involvement to build trust, evaluate its work and retain control. That does not mean asking millions of owners to become integrators. 

Consider what the system of record a small business already runs on can provide: operational data; permissions that determine who can see and change what; and workflows that define where decisions get made. Those are three of the hardest problems in any AI deployment, and inside the system of record part of that foundation is often already in place, even if it needs further work. Starting there lets us build on what already exists, rather than asking every business to assemble those foundations from scratch. 

This is also why the horizontal AI assistants now being sold to small businesses run into the same wall. They still have to be bought, installed, connected and trusted, one business at a time. Our focus is on software that has already been bought, installed, connected and trusted, often for years. That is a significant head start.

Three layers

Building this properly means being precise about who does what. We think about it as three layers.

The software knows. It holds the orders, the invoices, and the history, with the same access controls the human user has.

The model does. It works within a bounded set of actions designed before deployment rather than discovered afterwards. An agent that triages support can read the ticket, the account, and the knowledge base, and it can propose a reply. It cannot issue a refund unless a person approves it.

The person stays in control. Where human judgement is needed, a person approves rather than performs. Otherwise, the agent acts independently. That balance evolves with ongoing evaluation. It is a different job from the one they had, and in our experience a better one. 

Underneath all three sits evaluation. Nothing reaches production without it, and nothing stays there without it: task completion, error rates, how often a person has to step in. If we cannot measure whether it is doing the job, it does not ship. Cost is visible per business from day one, because AI that improves a product while damaging its unit economics has not improved it.

The task is not simply to put a new tool in someone’s hands. An agent takes on part of a job, so it must complete the work reliably, with a clear route back to a person when it cannot. That responsibility does not end at launch. As models, data, and workflows change, the system needs retesting, with new failures informing new checks. That is what turns a demonstration into something people can depend on every morning. 

The advantage of building AI across businesses

Circeus acquires and evolves mission-critical software for the everyday economy, the businesses whose software has to work every day, and who sit on the wrong side of every adoption chart above. Across the group, that software serves more than 250,000 businesses, and many of those relationships go back years.

The reason a holding company is the right vehicle for this, rather than each software business working alone, is the same reason adoption is stuck at one in five. Building the capability properly is expensive the first time. The evaluation tooling, the guardrails, the safe way to give a model access to data, the patterns for keeping a person in the loop: a vertical software company with a 20-person team would struggle to build and maintain all of that alone, and should not have to.

So we build it once, on a centralised AI platform, and reuse it across the group. 

An agent that triages support in one product becomes a capability the next product pulls in, configured for its own customers, its own data and its own workflow. The portfolio companies stay decentralised and keep building for their own markets. What they share is the plumbing and the lessons. Every redeployment is designed to be cheaper and faster than the last, which is what makes the model compound rather than repeat.

Building from the inside

There is a deeper reason we do this as owners rather than as a vendor. The knowledge of how a business actually runs, which exceptions matter, which supplier can be chased and which cannot, does not live in any general model. It lives in the people who run these businesses, most of it unwritten, and none of it visible to a lab training on the open internet. For AI to be useful to the everyday economy, that knowledge has to shape how the systems behave. The only way we know to make that happen is to build alongside the people who hold it.

Long-term ownership makes that possible. We can work inside a business, understand its particulars, and shape AI to the way its operators actually work rather than to an abstraction of their industry. A shared stake in the business's continued success gives everyone a reason to share what they know, correct mistakes early, and measure honestly whether the technology is helping. It also gives us the responsibility to act on what we learn.

For an engineer, this is a rare place to build. We work with people who have spent their careers in an industry and understand its problems in detail. The hard, interesting part is making AI reliable enough for them to depend on every day.

The businesses in the Circeus portfolio, and the quarter of a million businesses they serve, provide things people will keep needing however automated the world becomes. They deserve a fair chance to compete through this shift, with tools as good as anyone's. We believe helping them do so is among the most important things people working in applied AI can do. It will not be done without the people who understand the work. 

When Circeus acquires a software business, we commit to making investments it would struggle to make alone, with the intention of owning it for decades. We want the founders who sell to us to become more ambitious about their products, not less, and to have the capital and engineering talent to act on that ambition.

From systems of record to systems of action

The software the everyday economy runs on today records what happened. Orders placed, tickets raised, invoices paid. The next version does more of the work, with a person guiding and approving.

That shift will not come from asking every business owner to become an AI expert. It will come through the software they already trust, evolved by owners who intend to hold it for decades.

The everyday economy should not have to work out AI on its own. AI will reach these businesses through the software they already rely on, embedded into the workflows they already use and trusted to run every day.