Most companies now use AI somewhere, and most operations look much as they did before. That isn't because the tools are weak. It's because an operation moves at the speed of its slowest step, and AI usually gets pointed at whichever step is easiest to automate, not the one that holds everything up. Speed up the wrong step and you just build a bigger queue in front of the right one.
What is AI for business operations?
AI for business operations means using AI to move work through your company faster, cheaper or with fewer errors, by taking on the reading, sorting, drafting and checking that sit inside processes like order handling, invoicing, scheduling, customer service and reporting. It pays off when it relieves the step that limits the whole process, and it rarely pays off anywhere else.
This guide is for COOs, operations leads and the leaders who own an operation in a company of 20 to 1,000 people. It covers where AI is used today, why task-level gains so often vanish, how to find the bottleneck, which operations work to hand AI first, and how to measure it. It doesn't cover AI for IT operations (often called AIOps), which is a separate field about monitoring systems.
How is AI used in business operations today?
Less than the headlines suggest, and mostly away from operations. The best data comes from the US Census Bureau, which added detailed AI questions to its Business Trends and Outlook Survey. In a working paper revised in October 2026, Census researchers report that 18% of firms used AI in a business function between November 2025 and January 2026, or 32% when weighted by employment. They expected that to reach 22% within six months.
Among firms that use AI, it's concentrated in a few functions. Sales and marketing comes first, used by 52% of AI-using firms, then strategy and business development at 45%, IT at 41% and R&D at 40%. The researchers note that "use in production and supply-related functions, quality control, and distribution are relatively less common." And 57% of AI-using firms use it in three or fewer functions.
The pattern inside the work is the same. Where staff use generative AI, writing and editing is the most common task, reported by 85% of those firms, with information search far behind at 50%. In other words, most AI use today is people writing faster. That's useful, but an operation doesn't run on writing. It runs on work moving from one step to the next.
Why doesn't AI speed up the whole operation?
Because the operation can only go as fast as its slowest step, and faster work anywhere else just waits in front of it. If invoices take a day to key in and three days to approve, cutting data entry to an hour saves almost nothing. The invoices reach the approver sooner and sit there longer.
The surveys show the gap this creates. In a Richmond Fed survey of more than 700 executives, run between November 2025 and January 2026, firms reported that AI raised output per worker by 1.8% in 2025. But when the researchers worked out the gain implied by the same firms' revenue and headcount changes, "the implied gains are much smaller across all major industries." They call it a productivity paradox.
A larger study points the same way. The October 2026 revision of Firm Data on AI, an NBER working paper based on nearly 9,000 executives in seven countries surveyed in summer 2026, found that 78% of firms use AI. Yet executives say AI added only 0.2% a year to labor productivity growth over the past three years. They expect 0.6% a year over the next three.
Neither study names bottlenecks as the cause. But it's what you'd expect when AI makes individual tasks quicker in a process nobody has redesigned. Our framework puts finding the bottleneck before building anything for exactly this reason.
Where does AI pay off in business operations?
Where the work at or just before the bottleneck is reading, sorting, drafting or checking information, at volume, with a person who can catch mistakes. That describes a lot of operations work in a mid-size company. Here's where to look, by the shape of the work rather than the department:
| Kind of work | What AI does well | What to watch |
|---|---|---|
| Intake and triage (shared inboxes, request forms, tickets) | Reads each request, tags it, routes it, flags what's missing | Misrouted urgent items. Keep a person on the exceptions queue |
| Documents in, data out (invoices, purchase orders, forms, contracts) | Pulls fields into your system and flags mismatches | Wrong figures that look right. Check against a total or a second source |
| Standard replies and documents (status updates, quotes, letters) | Drafts from your templates and the record | Promises nobody approved. A person sends, not the AI |
| Status and reporting (weekly reports, handover notes, exception lists) | Summarizes records into a short, consistent format | Summaries that skip the one exception that mattered |
| Answering staff questions (procedures, policies, "how do I") | Answers from your own documents, with the source | Out-of-date procedures. Name an owner for the source |
| Scheduling and coordination | Proposes slots and sequences against stated rules | Rules nobody wrote down. Write them first |
The last column matters as much as the middle one. Every row is a place where AI's output feeds the next step, so an error doesn't stay put: it travels downstream to whoever acts on it. Set the check before you switch anything on. The output standard lesson in our course covers what "checked" should mean.
There's also hard evidence that AI helps most where people are newest. In a study of 5,179 customer support agents published in the Quarterly Journal of Economics in 2025, an AI assistant raised the number of issues resolved per hour by 14% on average and by 34% for novice and low-skilled agents, with minimal gains for the most experienced. If your bottleneck is a team that turns over often, or new staff who can't yet work without help, an assistant that answers from your own procedures is one of the better-evidenced places to start.
How do you find the bottleneck in an operation?
Follow one piece of work from the customer backwards, and look for the step where it waits longest or piles up highest. Don't ask teams which step is slowest. Everyone thinks it's someone else's. Look at the work instead.
- Pick one value stream. For example, order to cash, request to resolution, or hire to first day. One stream at a time.
- Trace it back from the customer. Start at what the customer receives and walk upstream, step by step, noting who does what and where the work sits between steps. The lesson on tracing work back from the customer shows how.
- Look for the pile. The bottleneck usually has the biggest queue in front of it: the oldest items in a shared inbox, the longest approvals list, the backlog everyone works around.
- Look for rework loops. Work that comes back (missing data, wrong format, a second approval) often points to the real constraint upstream of the step that complains.
- Check the people. Sometimes the bottleneck is one manager who signs everything off, or a leader who decides everything. That's still the bottleneck.
- Name one, and put a value on it. What would an extra day of capacity at this step be worth in revenue, cash or cost? The lesson on naming one bottleneck shows how to value it.
Then write down what you expect to change and by how much, before you touch anything. That written prediction is what lets you tell later whether AI helped or just felt helpful.
Which operations tasks should you hand to AI first?
A task that sits at or just before the bottleneck, repeats often, uses data you're allowed to send to the tool, and produces output a person can check quickly. Run every candidate through five questions:
- Does it touch the bottleneck? If speeding it up only feeds a longer queue, skip it, however tempting it is.
- Does it repeat? Daily, at volume. A task done twice a month won't repay the setup.
- Can a person check it faster than doing it? If checking takes as long as the work, you've added a step, not removed one.
- Is the data allowed? Customer, employee and financial data each have rules. Your data map should say which tools may see what.
- Can you measure it? You need a before figure for cycle time, queue or errors, taken before you start.
Before any of this, redesign the step. Most bottlenecks have slack that costs nothing to recover: approvals nobody needs, batches that could flow, handoffs that could merge. Redesign first, then automate what's left. The redesign-first lesson explains why the first result should come from improving the work, not from a tool.
If a task needs AI to act in your systems on its own (update the order, send the email, change the record), that's an agent, not an assistant, and it needs more care. Our guide to AI agents for business covers what to decide before an agent gets access.
What AI tools do operations teams use?
Mostly the ones they already have. For a company of 20 to 1,000 people, AI for operations usually comes from three places:
- The general assistant you already pay for. ChatGPT Business, Claude, Microsoft 365 Copilot or Gemini in Google Workspace handle reading, drafting, summarizing and answering from your documents. Our shortlist of the best AI tools for business compares their prices and data terms.
- AI features inside your existing software. Helpdesk, accounting, ERP, scheduling and document tools increasingly ship AI that vendors switch on by default. Find out what's already running before you buy anything new. The free Shadow AI Audit helps you list it.
- An automation tool, once the process is redesigned. When work has to move between systems, a workflow or automation platform connects them. Buy it for a specific redesigned step, not as a platform in search of a use.
The best AI tool for business operations is the one that relieves your bottleneck. That's why the tool decision comes after the bottleneck, not before.
How do you measure AI in operations?
Measure the operation before and after, against the prediction you wrote down, and ignore usage counts. Four measures cover most value streams:
| Measure | What it tells you |
|---|---|
| Throughput | Units finished per week at the end of the stream: orders shipped, tickets closed, invoices paid |
| Cycle time | How long one unit takes from start to finish, including waiting |
| Queue at the bottleneck | How much work is waiting in front of the limiting step, and how old the oldest item is |
| Rework rate | The share of units that come back to an earlier step |
If throughput and cycle time don't move, the change didn't help the operation, whatever the time-saving survey says. If the queue just moves to the next step, good: you've found your next bottleneck. Fixing one always exposes the next, which is why this is a loop and not a project.
Executives' own expectations are a useful reality check. In the Richmond Fed survey, firms said they expect AI to replace "administrative work, data entry, customer service, and other routine operational roles." In the Census data, though, 66% of AI-using firms used it only to help people with tasks, and AI-related employment decreases appeared in only 2% of firms. Plan for jobs to change before headcount does, and decide in advance where freed time goes. Time saved only counts once it has gone somewhere.
How to start with AI in your operations
01 · Pick one value stream and its owner
Choose the stream that matters most to customers or cash this year, and name one leader who owns it end to end. This gives you the scope for everything that follows.
02 · Find the bottleneck and write down the prediction
Trace the stream back from the customer, find the step with the biggest pile, value it, and write down what you expect to change by when. This gives you the baseline and the target.
03 · Redesign the step before adding AI
Remove approvals, batches and handoffs that don't need to exist, and measure again. This gives you the first result at no cost, and shows what's left for AI to do.
04 · Give AI one task at or before the bottleneck
Choose a task that passes the five questions above, set the check a person will do, and use a tool you already have where you can. Write a one-page case with a cost cap if it needs anything new. This gives you a small, reversible change you can measure.
05 · Measure, then pick the next bottleneck
Compare throughput, cycle time, queue and rework with your prediction. Keep what moved the numbers, drop what didn't, and move on to the next constraint. This gives you a loop that keeps paying, which is what AI deployment looks like when it works.
If you'd like help running the first loop, our AI implementation service does it alongside your team.
What's still open
The evidence on AI and operations is thin and moving. The Census data on business functions covers November 2025 to January 2026, before agents were common, and it spreads operations across several functions (production, supply chains and purchasing, quality, distribution) rather than measuring it as one. The executive surveys measure what leaders report, not audited results, and the NBER and Census papers are working papers that haven't been peer-reviewed. The best controlled evidence, the customer support study, comes from one company's contact center. We'll update this article when the Census Bureau publishes its next AI supplement or new controlled studies of AI in operational work appear. The date at the top of the page shows when we last checked.
Sources
- The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks (CES 26-25R), US Census Bureau Center for Economic Studies, October 2026: working paper on the 2026 AI supplement to the Business Trends and Outlook Survey. We used its adoption rates, the use rates by business function and task, and the figures on augmentation and employment.
- How Might AI Change the Workplace? Evidence From Corporate Executives, Federal Reserve Bank of Richmond, March 2026: survey of more than 700 executives. We used the reported 1.8% gain in output per worker, the smaller revenue-implied gains and the tasks firms expect AI to replace.
- Firm Data on AI (NBER Working Paper 34836), National Bureau of Economic Research, February 2026, revised October 2026: survey of nearly 9,000 executives in seven countries. We used its adoption rate and its estimates of AI's effect on productivity growth.
- Generative AI at Work (NBER Working Paper 31161), Brynjolfsson, Li and Raymond, published in the Quarterly Journal of Economics, 2025: field study of 5,179 customer support agents. We used its productivity results for all agents and for novice agents.