Article

What is the agentic economy? And what does it mean for a small business

The agentic economy for growing businesses: what an AI agent takes over on an ordinary Tuesday, and where it goes from there.

BombosBombos team··12 min read

By Piet Baudoin · August 2026, updated September 2026

The agentic economy is an economy in which software does the work itself instead of just showing it to you. An AI agent reads what comes in, looks up the data in the software you already use, and prepares the work, after which a person approves it. At Anthropic, about 30,000 of these agents worked at once in August 2026 on the research and the build of the company itself.

That number is about an AI company, not about your business. It does show where this is heading: work you no longer do yourself but check. Below is what that looks like on an ordinary Tuesday morning at a small or midsize business, why it starts exactly there, and what I think remains once it keeps going.

What is the agentic economy, in plain English?

It is the shift from software that gives an answer to software that acts.

Microsoft puts the difference in one line in its own guide for developers: a language model goes from “Prompt to response,” an agent from “Goal to autonomous multi-step action” (Microsoft, 11 September 2026). So you no longer give it a question but a goal, and it chooses the steps toward it itself.

That sounds small, and it is not. With a chatbot, the work stays yours: you read the answer, you type it over, you click through your software. With an agent, that whole middle part disappears. What you get back is not an answer but prepared work, with the reason attached, that you only need to review.

That is also where the cost sits. A wrong answer you ignore. A wrong action has already happened. So this whole topic is not about intelligence, but about where the approval sits. How long an agent like this can keep working before it runs into something is in what an AI agent is and how long it can work alone.

Most business owners have never heard the word. In a British survey of small businesses from 16 September 2026, 44 percent knew the term agentic AI, and 5 percent used it regularly. That is not a gap to worry about. It means the term arrives later than the work does.

What does that look like on a Tuesday morning at a business of twenty?

Not in your strategy. In your inbox.

Take a shared inbox on a Tuesday morning. A supplier invoice comes in, as a PDF, with a description and an amount. It has no project number on it. Next to it is a phone note from a colleague who spoke to a customer yesterday about extra work. And in the software is the purchase order from two weeks ago.

The work between those pieces is work nobody enjoys. Someone has to work out which project that invoice belongs to, whether the amount matches what was ordered, and whether that extra work belongs with it. That is not bookkeeping. That is looking things up.

Here is how that goes with an agent alongside:

  1. It reads the email and the attachment, and sees that it is a purchase invoice: supplier, amount, description.
  2. It searches earlier messages for the purchase order with the same description, and finds the project it was ordered for.
  3. It compares the amount on the invoice with the amount on the purchase order, and sees that it is higher.
  4. It finds the phone note where the extra work was agreed, and attaches it.
  5. It prepares the entry against that project, with one line attached: on this project, because the purchase order from 2 September matches it and the difference is in yesterday’s note.
  6. A person reads that line, looks at the two documents underneath, and clicks Approve. Or corrects it, and that correction becomes the rule next time.

That is matching a purchase invoice to the right project, and at a business of twenty that job is split across three people. The same pattern sits in sorting a shared inbox and in the status question that comes back every day: read it, look it up, prepare it, get it approved.

Why does this start between your software and not inside it?

Because this work belongs to no one.

Your software is getting smarter too. Microsoft includes Copilot, the AI in its accounting and business software Business Central, “at no extra cost”, and adds that “fair-use policies, quotas, or pricing might be introduced later.” Business Central’s Payables Agent watches the company mailbox for vendor invoices, matches them to purchase orders and receipts, and “proposes posting accounts for unmatched lines”. That is real work that disappears. But both are built for the work inside Business Central, and the problem above was not inside the software. It was in the email, the note and the software together.

There is a consequence in that which is rarely said out loud. The better the AI in each piece of software gets, the scarcer the work between them becomes. Every vendor solves its own piece. None of them knows the route your work takes from one system to the next, because that route sits in your people’s heads. An AI agent is built to learn exactly that: how things go at your business, from which invoices always check out to which customer calls instead of emailing. Every time you correct it, it remembers. The difference between those two kinds of AI is worked out in the AI in your software or an AI agent alongside it.

There is also a door that can close. AFAS, a Dutch business software package, writes on its own customer portal that AI integrations are allowed provided they are certified, and that in June 2026 not a single one was. So what an agent can do is not the same question as what it is allowed to do. That is worked out separately in can ChatGPT operate your accounting software.

Who will be on the other end of your email soon?

Increasingly, an agent, and that is the part most business owners do not see coming.

Three measurements point the same way. Adobe measured that the traffic AI assistants sent to US online stores over the first three months of 2026 grew by 393 percent, and that those visitors converted 42 percent better than regular visitors in March 2026 (16 April 2026). On 15 September 2026, Google turned on the connections between Gemini and software such as HubSpot, QuickBooks and Salesforce by default for everyone with access to Gemini in Google Workspace. And on 10 September 2026, Microsoft described a campaign of more than a million emails that posed as well-known suppliers and tried to obtain payments of almost $50,000, with “several indicators consistent with AI-assisted template development”: signs that those emails had been put together with the help of AI.

Put those three side by side. Your inbox becomes machine-driven on both sides. The customer who finds you, finds you through their assistant. The question that comes in was prepared by software. And the scammer has the same toolbox you do.

That changes two things about how you work. Your payment checks need to be tighter, because a neat email no longer says anything about who sent it. And the information about your business needs to be findable and unambiguous, because a machine is reading along that will not just pick up the phone to check.

What happens if this pace continues?

Two measuring sticks point the same way.

METR does not measure how long a model is busy, but how big a job it finishes on its own, in human time: if it finishes work that takes an expert 12 hours, that is its measure. GPT-4 reached about 4 minutes in March 2023, Claude Opus 4.6 reached that 12 hours in February 2026, doubling every 129 days.

The second measuring stick is screen work: what share of the tasks on a computer does a model finish? OpenAI measures that on OSWorld and reached 38.1 percent there in January 2025. Since then the test has been made harder: on OSWorld 2.0, counting partly finished tasks, Sol reached 62.6 percent in July 2026 and Astra reached 72.6 percent on 3 September 2026. On the safety test in the same publication, Sol went wrong in 22.0 percent of cases, Astra in 2.4 percent.

So the work gets longer, screen work gets more reliable, and the outside world is working agentically too.

My hypothesis: by the end of 2028, the lookup work in a small or midsize business is largely gone, and the scarcity shifts to two other things. Who gets to decide, and how well your business can be read by the agent on the other end.

Two things break that hypothesis. METR itself writes that on tasks where it has to look at a screen, a model lasts 40 to 100 times shorter than on the programming tasks that 12 hours comes from, because it reads a screen poorly. And a vendor can keep its door shut, as AFAS shows.

So choose tasks you can check afterward, and hold off on anything that clicks around unseen in your software.

Further out, as a picture and not a prediction. I think we will stop making a fuss about data and numbers, because that calculation work gets done for you. What remains is the decision. And you will soon publish in a form that is convenient for an AI, because the reader on the other end is increasingly one itself.

What can this not do today?

An AI agent does not take over a role, and it knows nothing that lives only in someone’s head.

You see the first point in the numbers above: well-defined work with a trail you can check goes well, work that runs differently every week does not. Anthropic describes a contractor who caught the model calculating a quote with a square-footage multiplier instead of the real wall measurements, and who has asked ever since: “show me your work.” That is exactly the right attitude. Not trusting the outcome, but the trail underneath it.

The second is a real limit, not a disappointment. The agreement with that one customer who has been invoiced differently for twenty years is written down nowhere. We get it out by asking your people about it and recording the answer, so it is simply correct next time.

And then there is the approval. The makers are now building that in themselves: Anthropic writes about its package for small businesses, installed more than 900,000 times by September 2026, that Claude does the work and “waits for your approval before anything sends, posts, or pays.” At Bombos, no message goes to a customer and no payment goes out without someone at your business clicking Approve. That is technically enforced and always on. Why that should be a boundary and not a setting you can forget, is in should an AI assistant send email on its own. Who pays when it still goes wrong is in who is liable when an AI agent makes a mistake.

How do you start small?

With one task that comes back often and that you can check afterward.

Not with a project. Pick the job where something gets retyped the most, put something working alongside the people who do it now, and let them correct it. Within a few weeks you will know if it works. What a step like that costs per month is in what an AI agent costs per month, and whether you need several side by side right away is in do you need several AI agents at twenty people. What the big assistants already do today is in what Claude can do on its own for a small business.

Bombos reads what comes in, email, messages, receipts and invoices, looks up the data in the software you already use, and prepares the work with the reason attached. We start with one task, and after that your own team teaches Bombos the next one.

So this grows with the lines above. For us, the model is a part that can be swapped: we put the strongest and most cost-efficient model underneath right away, and switch per client to what that client wants. What Bombos learns about your business does not sit in the model but in what your team has approved and corrected, so a better model does more with that without you starting over. Everything runs on our own servers in Europe. We read the field every day and write a daily newsletter about it, with sources we open ourselves.

Our promise: after three months, the work we start with is ready every day, without anyone needing to think about it. Leave your number on the contact page, and we will call you back.

Start with the work that falls between your systems.

Sources

Every source was opened on 20 or 23 September 2026 and every quote appears in it word for word.

What an agent is and how many are working

What the software does itself

Who is on the other end

How fast it is moving

What goes wrong and what people notice