How Much of My Business Can Actually Be Automated?

How much of my business can actually be automated?

Roughly half the activities can be automated, almost none of the jobs, and far less than either number in the first year.

The research everyone quotes is the McKinsey Global Institute's A Future That Works, published in January 2017, which analysed more than 2,000 work activities across 800 occupations and found that almost half the activities people are paid to do in the global economy could be automated by adapting technology that had already been demonstrated. The same study found that fewer than 5% of occupations could be fully automated, while about 60% of occupations had at least 30% of their constituent activities in scope.

Those two findings together answer the question as most owners mean it. Your bookkeeper's job is not half automatable in the sense that you need half a bookkeeper. A third of what your bookkeeper does is mechanical, and that third can move.

Actual adoption runs far behind the technical ceiling. The US Census Bureau's Business Trends and Outlook Survey, which asks hundreds of thousands of firms directly, found the share using AI to produce goods or services rose from 4.6% at the start of 2024 to 10% by September 2025. When the survey reworded the question in November 2025 to ask about AI use in any business function, the national rate came out at 17.3%, and the Economic Innovation Group's tracking of that series puts it between 17% and 20% from December 2025 through May 2026.

The distance between almost half technically possible and a fifth actually doing anything is not a gap in the technology. It is the cost of choosing, building, integrating and operating, which is the work we do.

Why is the answer about activities rather than jobs?

The answer is about activities because automation takes tasks, and tasks do not come in job-shaped packages.

A dispatcher's day is a sequence: answer the phone, qualify the job, check the calendar, assign the right technician, send the confirmation, chase the one that did not confirm, handle the reschedule, update the notes. Some of those are rules with a clear right answer. Some need a person who knows that this customer's building has no lift and that technician cannot park there. Automating the sequence means taking the first kind and leaving the second, which changes what the dispatcher's day looks like without removing the dispatcher.

Task level research on language models lands in the same place by a different route. Tyna Eloundou, Sam Manning, Pamela Mishkin and Daniel Rock, in GPTs are GPTs, published in Science in 2024, estimated that around 80% of the US workforce has at least 10% of their tasks exposed to large language models, while about 19% have 50% or more exposed. Exposure there means a model could cut the time to do the task by at least half without losing quality, which measures reach rather than replacement.

What that means for a business planning work: stop asking which roles AI affects and start asking which activities repeat. Repetition is the signal. An activity that happens 400 times a month in roughly the same shape is a candidate. The same activity happening four times a month is not, however annoying it is, because the fixed cost of automating it never comes back. Where repetition tends to cluster by trade is mapped on our industry boards, with a maturity label on each capability.

Does "can be automated" mean technically possible or worth doing?

Technically possible and worth doing are two different questions, and conflating them is how companies end up with systems nobody uses.

Technically possible means a working version can be built with today's tools. Worth doing means the volume, the error cost and the operating cost make the arithmetic work after you account for review and exceptions. Most of the gap between the two is volume.

A rough filter that holds up in practice. An activity is worth automating when it happens at least weekly, follows a pattern you could describe to a new hire in a page, has an input that arrives in a form a system can read, and has a cost of being wrong you can live with while a person reviews the output. Miss two of those and it belongs on next year's list.

Business use of AI already skews towards whole-task delegation rather than assistance, which is what you would expect once a process is worth automating at all. Anthropic's Economic Index report of 15 September 2025 found that 77% of business API use showed automation patterns, against about 50% of consumer conversations on Claude.ai, and that fully directive use, where the whole task is handed over, rose from 27% to 39% within eight months.

There is also a hard ceiling on the judgment end that marketing tends to skip. TaxCalcBench, a benchmark built from Column Tax's own test cases and published on 22 July 2025, gave frontier models complete 2024 federal tax returns to compute. The best models produced fewer than a third of the returns correctly even on a simplified sample set, and the study reports that they consistently misused the tax tables, miscalculated, and got eligibility wrong. Work that has one correct answer defined by a rulebook is not automatically work a model gets right.

Which parts of a process can be automated today, and which cannot?

Four tiers decide which parts of a process can be automated today, and knowing which tier a candidate sits in is most of the decision.

Two rules we apply to that table on every engagement. The first is that a review step is part of the design rather than an admission of weakness, and the error rate gets measured rather than assumed, for the reasons set out in what happens when AI gets it wrong. The second is that the tier a capability sits in gets published rather than implied: the capabilities on our industry boards are marked proven, working or early, and nothing gets promoted a tier to make a sale.

The honest summary for most businesses between 10 and 200 people: the first tier is where the first year's money is, the second is where the second year's money is, and the third is where the sales pitches you should distrust come from.

How do I work out the number for my own business?

You work out the number for your own business with four columns and an afternoon.

List every repeating process. Not projects, not one-offs. The things that happen on a schedule or in response to an event: enquiries arriving, quotes going out, jobs getting scheduled, invoices arriving, invoices going out, reports getting produced, follow-ups getting chased, records getting updated.

For each one write down monthly volume, minutes of human touch time per unit, and the tier from the table above. Multiply volume by minutes to get monthly hours, then total the hours in the first two tiers. That total, over the total hours of the people involved, is your real automation share. In the businesses we have mapped it concentrates in two or three processes rather than spreading evenly across the list, which is why the ranked order matters more than the overall percentage.

Then sort by hours and look only at the top five. The arithmetic that matters next is per process: hours a month, times a fully loaded hourly cost, against what the automated version costs to run and review. The method for that side of it is in how to measure whether AI automation is saving money, and the order to attack them in is the subject of where AI creates leverage in a business.

One thing to check before you commit to any of it: whether the inputs are in a state a system can work with. A process that depends on information living in three inboxes and one person's memory needs that fixed first, and is our data good enough for AI covers how to tell.

How much can be automated in my industry?

How much can be automated in your industry depends on where the repetition sits, and it sits in different places by trade.

In home services the volume is in the phone and the schedule: enquiries, booking, confirmations, reminders, follow-up on quotes that went quiet. Dispatch and pricing stay with people who know the jobs. The capability by capability picture is on our home services board.

In manufacturing the repetition is in documents and data moving between systems: quotes, work orders, certificates, supplier paperwork, production reporting. The machines are already automated; the office is not. See the manufacturing board.

In legal and accounting work the volume is in intake, document handling and the administrative layer around matters and engagements, while the professional judgment and the sign-off stay where the liability is. See the legal and accounting board.

In dental and medical practices it is the front office: calls, scheduling, recalls, forms, insurance paperwork, clinical documentation with a clinician reviewing it. Patient data makes the compliance work part of the build rather than an afterthought. See the dental and medical board.

In coaching, consulting and training firms the repetition is in enrolment, scheduling, materials and the reporting clients expect. See the professional firms board.

Each board carries the specific capabilities with their maturity labels, so the answer to "how much" in your trade reads as a list rather than a percentage.

What happens to the work that cannot be automated?

The work that cannot be automated gets more of your people's attention, which is the point of doing any of this.

The arithmetic of a service business is that the work only a person can do is the work that earns. A coordinator who spends fourteen hours a week confirming appointments and keying data is not spending those hours on the client who is about to leave or the quote that needs a phone call. Moving the mechanical third of a role is how a team of nine takes the volume that would have needed a team of twelve.

That is why we write about capacity rather than savings. A business that automates its way to the same output with fewer people has cut a cost once. A business that automates its way to more output with the same people has changed what it can take on, and can do it again next quarter. The work left behind is harder, more varied and closer to the customer, which is also what people would rather be doing.

Where this goes wrong is when nobody decides what the freed hours are for. The hours get absorbed, the team feels slightly less pressed, and the capacity shows up nowhere. Deciding in advance what the capacity is for belongs in the plan, next to the measurement, and both are part of the roadmap rather than an afterthought.

Frequently asked questions

Related resources

• How do I measure whether AI automation is actually saving us money?
• Where AI creates the most leverage in a business
• How much time and money can AI automation actually save?
• Is our data good enough for AI automation?
• What is an AI roadmap, and what should be in one?
• AI for small business: what actually works
• What AI can do, by industry