AI Adoption at Work: What U.S. Business Data Really Shows

The headline number needs context

The Census Bureau’s BTOS supplemental AI questions offer a grounded view of workplace adoption across U.S. firms. The topline finding is modest: only 18% of firms reported using AI in the prior two weeks in any business function.

That number sounds low until you look at how the question is framed. It appears to capture more formal, recognized business use rather than casual employee experimentation. In other words, “the company uses AI” is not the same as “someone on the team pasted something into ChatGPT.”

That distinction matters because formal adoption stats can miss a lot of real behavior.

Workers are ahead of management

One of the clearest signals in the data: employees are often using AI before their employers officially are.

In 36% of firms where workers use AI for job tasks, the firm itself reports no formal AI adoption. That’s bottom-up diffusion in plain English. The tool shows up at the desk before it shows up in the strategy deck.

This creates a familiar workplace gap:

• Workers want speed
• Managers want control
• IT wants visibility
• Legal wants fewer surprises

Everyone is technically correct, which is annoying.

The practical issue is governance. If employees adopt AI on their own, firms may have little visibility into what data is being shared, which outputs are being trusted, and whether anyone has set review standards. The result is not just faster work. It’s shadow AI.

Most AI use is still narrow, not enterprise-wide

Even among firms that do use AI, adoption tends to stay confined to a small number of functions.

More than half of AI-using firms deploy it in only one to three of the 15 business functions tracked in the survey. That suggests most companies are not redesigning the whole business around AI. They’re plugging it into a few obvious places and seeing what happens.

That’s a very normal adoption pattern. Companies rarely transform all at once. They start where friction is highest and risk is lowest.

Where AI is actually showing up

The most common business functions for AI use are the ones you’d expect: areas heavy on text, analysis, and coordination.

Sales and marketing lead, followed by strategy and business development, IT, and R&D. These are functions where AI can help draft, summarize, research, organize, or accelerate existing workflows without touching the physical core of operations.

At the other end, production-linked functions lag behind. Distribution, quality management, and goods production see much lower rates of use, though some early activity is emerging in areas like logistics and warehousing.

That split tells you something simple but important: current AI adoption follows workflow shape. The more digital and language-based the work, the easier the fit.

AI is mostly helping with tasks, not replacing jobs

A lot of AI discussion collapses into one question: “Is it taking jobs?” The BTOS data gives a more nuanced answer.

The survey separates AI’s effects into three buckets:

• Substitution: AI performs a task previously done by an employee
• Augmentation: AI supplements or enhances a task
• Creation: AI introduces a new task that wasn’t previously done

Among AI-using firms, augmentation is the dominant pattern. About 44% report using AI to enhance employee tasks. Substitution and task creation are much less common, each around the 10–11% range.

That means the most common current use case is not “AI replaces the worker.” It’s “the worker now does the task differently.”

This matches what many teams are experiencing in practice. The job still exists, but parts of it get compressed, sped up, or reshaped.

The substitution story is still worth watching

Limited does not mean irrelevant.

While task substitution remains less common than augmentation, the intensity appears to be rising. The share of firms reporting a large number of substituted tasks increased notably between the first and second BTOS AI supplements.

So the current picture is calm, but not static. Substitution is not yet the dominant mode, but it may be deepening where firms have already started using AI more seriously.

That’s the kind of shift worth tracking carefully. Not because it proves a labor shock is here, but because it can signal where job design and skill requirements are beginning to change under the surface.

Employment effects are modest, at least for now

When firms are asked directly about net employment effects from AI, the results are strikingly uneventful.

Most AI-using firms report no change in total employment attributable to AI. Small shares report employment increases or decreases.

That does not mean AI has no labor impact. It means the impact is not yet showing up as widespread net headcount change in this data.

This is an important correction to both extremes:

• “AI is already replacing everyone” does not fit the survey
• “AI changes nothing” also misses the task-level shifts already underway

Tasks can change before payroll does. Often, they do.

AI seems to be replacing software more than people

One of the more revealing findings is that AI appears to be displacing older technology more visibly than labor.

A meaningful share of AI-using firms report replacing existing software or equipment with AI, while only a small share report AI-related employment decline.

That’s a useful framing for buyers and operators. In many workplaces, the first casualty of AI is not the employee. It’s the older tool stack.

This helps explain why AI can spread inside firms without immediately producing dramatic staffing changes. It may first act as a layer that absorbs functions previously handled by software, workflows, or fragmented systems.

Why most firms still aren’t using AI

The biggest barrier is not cost. It’s not regulation either. It’s a more ordinary problem: many firms do not think AI is relevant to what they do.

Among firms not planning to use AI in the next six months, the most common response is that AI does not seem applicable to their business. After that come lack of knowledge about AI’s capabilities and privacy or security concerns.

That mix matters.

It suggests the adoption problem is often not “we can’t afford AI” but “we don’t see a sensible use case.” For many businesses, especially outside knowledge-heavy sectors, AI is still being interpreted as something adjacent rather than operational.

The real bottleneck is understanding, not just access

When firms say AI does not apply to them, that can mean two different things.

Sometimes they’re right. Not every workflow benefits from AI, and forcing use cases is how teams end up with expensive demos and no actual value.

Sometimes, though, “not applicable” is really “we haven’t mapped this to a business problem yet.”

That’s why AI literacy matters so much. Not literacy in the abstract, but practical literacy:

• What kinds of tasks AI can handle well
• Where human review is still essential
• Which workflows are low-risk places to start
• What policies are needed before employees improvise at scale

Without that, firms either avoid AI entirely or adopt it through unmanaged worker behavior. Neither is ideal.

What this means for teams choosing tools

If you’re evaluating AI tools for real work, the data points to a few grounded takeaways.

First, don’t confuse employee enthusiasm with organizational readiness. A team may already be using AI informally while the company has no policy, no controls, and no shared standards.

Second, narrow adoption is normal. You do not need an “AI everywhere” strategy to get value. One clear use case in marketing, support, research, documentation, or internal ops is a more realistic starting point than a grand transformation plan.

Third, measure task impact before job impact. Ask:

• What gets faster?
• What gets easier?
• What gets reviewed differently?
• What gets dropped from the workflow entirely?

Those answers are usually more informative than broad debates about replacement.

The useful takeaway

The U.S. business data paints a picture that is less dramatic than the hype and more interesting than the backlash. AI at work is real, but uneven. Workers are often leading, firms are adopting selectively, and the clearest effects are showing up in tasks and software stacks more than in headcount.

If you’re trying to choose what actually works, the lesson is simple: look for concrete workflow fit, not AI theater. The companies getting value appear to be using AI in specific places, with specific purposes, long before they can honestly say they’ve “transformed.”