Among US Firms That Use AI at All, 52% Use It in Sales and Marketing, and Most Stop There

A Census Bureau working paper published in April 2026, The Microstructure of AI Diffusion, gives the clearest picture yet of where artificial intelligence actually sits inside American businesses. Among firms that had adopted AI in the reference period of November 2025 to January 2026, sales and marketing was the leading business function at 52%, ahead of strategy and business development at 45% and IT at 41%.

The finding underneath that is more useful. 57% of adopting firms use AI in three or fewer business functions, and 65% of firms with workers using AI limit it to three or fewer tasks. Adoption is real, and it is shallow. Most businesses have one or two things AI does and a large surface where nothing has changed.

For a small US business deciding what to do about AI, that is a more actionable picture than the headline adoption rate, because it says the question is not whether to start. It is where the second and third use case should go.

What did the Census paper actually measure?

It draws on the Business Trends and Outlook Survey and looks at three levels at once: whether a firm uses AI, which business functions it uses it in, and which worker tasks it appears in. The authors are Bonney, Breaux, Dinlersoz, Foster, Haltiwanger and Pande, and the paper is CES-26-25.

In the reference period, 18% of firms reported using AI in a business function, rising to 32% when weighted by employment, with firms expecting that to reach 22% within six months. Very large firms in information, professional services and finance reported the highest rates, at roughly 50% to 60%.

Why is sales and marketing the first function to adopt?

Because the output is text and images, the volume required is high, the quality bar is reachable, and a mistake is cheap. A poor draft caption is discarded. A poor automated journal entry is a problem. Businesses sensibly start where the downside of being wrong is smallest, and marketing is that place.

There is a less flattering reason too. Marketing output is easy to produce and hard to evaluate, so it is possible to increase volume substantially without anyone establishing whether the extra volume did anything. Some of the 52% is genuine gain. Some is faster production of material that was not working before.

What does shallow adoption cost a business?

Mostly it costs the compounding. One tool used in one function is a productivity improvement for one person. The returns that matter arrive when a task hands off to the next task without a human moving the information between them, and that requires more than one function to be involved.

The concrete example most small businesses recognise is the enquiry. A form is filled in, someone eventually sees it, someone replies, someone adds it to a spreadsheet, and someone remembers to follow up. AI in the marketing function makes the content that generates the enquiry. It does not touch the four steps after it, which is where most enquiries are actually lost.

Where should the second use case go?

Follow the delay. In most small businesses the largest recoverable loss is not production speed, it is the gap between a customer making contact and a human responding, and the gap between a first response and a follow-up that never happened.

That is unglamorous work: routing, acknowledgement, scheduling, reminders, and getting records updated without anyone retyping them. It is also where automation is most reliable, because the rules are clear and the tasks are repetitive. Our AI automation services for US businesses are built around that layer rather than around content generation, for exactly this reason.

Does this mean cutting staff?

The paper suggests not. It reports that only 2% of firms experienced AI-related employment decreases, with most organisations using AI to augment tasks rather than replace them. That is consistent with what shallow adoption implies: businesses are removing steps, not roles.

It is worth stating plainly because the fear distorts decisions. Owners either avoid automation because it sounds like a redundancy conversation, or they buy it expecting a headcount saving that does not appear and conclude it failed. The realistic outcome is that the same people handle more, and that fewer things fall through, which is harder to put in a spreadsheet but easier to notice.

How do you decide what to automate first?

Write down every task somebody does more than five times a week that involves moving information from one place to another. Copying an enquiry into a CRM, sending the same three follow-up messages, rekeying an invoice, chasing a document. That list is almost always longer than expected and almost always boring.

Then rank by frequency times annoyance, not by how advanced the solution sounds. The best first automation in most businesses is dull, saves twenty minutes a day, and works every day without supervision. The worst is ambitious, impresses people in a demonstration, and quietly stops being used within a month.

What should a firm not automate?

Anything where being wrong is expensive and the error is hard to spot. Final review of client-facing figures, judgement calls with regulatory consequences, and any communication where the tone carries the relationship. Those are the places to spend the time the automation gave back.

The same caution applies to published content. AI-assisted drafting is legitimate and widespread, and content that is generated, unedited and unchecked is a liability that search engines and readers both increasingly detect. We have written about where the line sits for professional practices in our guide to digital marketing for accounting firms, and about the specific tools that hold up in free AI tools US accountants are using.

What does implementation actually look like?

Smaller than most owners expect. One process, mapped as it currently happens rather than as it is supposed to happen, then the two or three steps in it that are pure information movement. Build those, run them alongside the manual version for a fortnight, then stop doing it manually once nothing has broken.

The failure mode is starting with a platform rather than a process. A business that buys a system and then looks for uses ends up with a subscription and a training problem. A business that fixes one irritating handoff, sees it hold for a month, and then fixes the next one, ends the year with something that works and that people trust.

How does this connect to being found in the first place?

Automation improves what happens after contact. It does nothing about whether contact happens. A business with an efficient enquiry process and no visibility has built a very responsive system for handling a small number of enquiries, which is a common and expensive place to arrive at.

The two pieces belong in the same plan. Our digital marketing services for US businesses and our AI automation work are usually sequenced so the visibility side runs first or alongside, because a faster response to nobody is not an improvement.

Why do so many firms stall after the first use case?

Because the first one is chosen by enthusiasm and the second has to be chosen by analysis. Somebody tries a writing tool, it works, and the obvious next step is not obvious, so nothing follows. The paper's finding that most adopters stop at three or fewer functions is the aggregate shape of that stall.

The other reason is that the second use case usually crosses a boundary between people. Content generation sits inside one person's job. Routing an enquiry from a form to a CRM to a follow-up involves marketing, sales and whoever owns the records, and now it needs a decision rather than an experiment. Businesses that get past the stall are usually the ones where someone was given the authority to make that decision.

What does a realistic first year look like?

Three or four automations, not thirty. Something that acknowledges enquiries immediately and routes them to the right person. Something that removes a rekeying step between two systems. Something that chases the document or the payment that people forget to chase. Possibly something that drafts a recurring report from data that already exists.

None of that is impressive to describe. All of it runs every day without anyone thinking about it, which is the actual measure. A business that ends the year with four dull automations that still work is meaningfully ahead of one that ran an ambitious pilot in March and quietly abandoned it in May.

How should a business measure whether it worked?

Pick the measure before building anything, and prefer counts to feelings. Enquiries answered within an hour. Follow-ups that happened on time. Records that were complete without anyone tidying them. These are countable before and after, and they move quickly enough to tell you something within a month.

Time saved is the measure everyone reaches for and the hardest one to trust, because the hours rarely reappear as anything visible. The more honest question is whether the specific failure you were trying to stop has stopped. If enquiries used to sit unanswered overnight and now they do not, that is the result, and it does not need converting into a currency figure to be real.

What we would not claim

We would not claim a number for hours saved, because it depends entirely on which processes a business has and how bad they currently are. Anyone quoting a universal figure has not looked at your operation. What can be said is that the Census data shows most adopting firms are using AI in three or fewer functions, which means the common situation is a business with one working use case and several obvious untouched ones.

We also would not claim automation replaces judgement about what to automate. That decision is the whole job, and it is made by looking at a real process rather than a product page.

We keep 82% of our clients, and the automation work is a large part of why, because it is the kind of change an owner can feel within weeks without needing a report to interpret it. Things that used to be forgotten stop being forgotten. That is a modest description, and it is the accurate one.

FAQs