What Is AI Operations?
AI operations is the work of making artificial intelligence part of how a company actually runs: choosing the functions it touches, rebuilding the workflows around it, owning its data, and keeping the result alive in production. A function, not a tool.
I can be more precise than my own opinion, because somebody has now measured it. In April 2026 the United States Census Bureau published a working paper, CES 26-25, on the 2026 artificial intelligence supplement to its Business Trends and Outlook Survey. It is the only nationally representative read I know of that separates three things most people run together: whether a firm uses AI at all, which business functions it uses it in, and what the firm changed about itself in order to use it. That third layer is AI operations. I read the paper this morning. Every number below comes out of it or out of the Census Bureau's own reporting on the same survey.
First, the two things this phrase means
Search for AI operations and you land in a collision of two different subjects, which is most of why the answers you get feel unsatisfying.
The first is AIOps, a specific discipline inside information technology. IBM defines it as the application of artificial intelligence capabilities, such as natural language processing and machine learning models, to automate, streamline and optimize IT service management and operational workflows. Datadog, which sells it, calls it a discipline that leverages machine learning algorithms to identify the root cause of incidents, helping teams wrangle incoming alerts, remove duplicate alerts, identify false positives, and provide early anomaly, fault and failure detection and analysis. Both definitions describe AI pointed at your infrastructure. If you run a platform with a pager rota, that is a real and useful thing to buy.
The second is what an owner usually means: AI pointed at the running of the company. Quotes, invoices, scheduling, onboarding, claims, chasing, reporting. The daily business of the house. Nobody sells that as a product, because it is not one. It is a function, and the rest of this piece is about what that function owns.
Where AI actually sits in American companies right now
During the supplement's reference period, November 2025 to January 2026, 18 percent of American firms used AI in a business function. On an employment-weighted basis, which is to say counted by how many people work at those firms, the figure is 32 percent. Adoption is expected to reach 22 percent within six months. In the Census Bureau's separate reporting on the same survey through to May 2026, overall usage hovered between 17 and 20 percent, and the size gradient is steep: 37 percent for firms with at least 250 employees, 32 percent for firms with 100 to 249, and under 20 percent for firms with four or fewer.
Among the firms that do use AI in at least one business function, it concentrates in a handful of places. Sales and marketing, 52 percent. Strategy and business development, 45 percent. Information technology, 41 percent. Research and development, 40 percent. Production, sourcing and supply chains, quality management and distribution trail well behind. The survey covers fifteen functions in all, from production of goods through finance and accounting, customer service, human resources and legal and compliance.
The number I would put on a wall is this one: 57 percent of the firms using AI use it in three or fewer of those fifteen functions, 24 percent in exactly one, and roughly 1 percent in all fifteen. The paper's own latent class analysis names the largest group the minimalist adopters, 37 percent of functional users, with low probability of use across every function. Comprehensive adopters are 4 percent.
So the common picture is not a company transformed by AI. It is a company with AI in the marketing corner.
The gap is the job
Here is the finding that turns a survey into a brief. Of businesses already using AI, 64 percent report no institutional adjustment whatsoever. The most common adjustments, training staff and developing new workflows, are each reported by about 15 percent of AI-using firms. Changes to data management and storage practices, and complementary capital investment, sit in the 7 to 8 percent range. The least common adjustment of all is hiring staff trained in AI.
Two thirds of the companies using this technology have bought the tool and changed nothing about themselves. The paper's own phrase for what that implies is a reliance on off-the-shelf tools, or a significant lag in organisational restructuring.
And the same paper, in its regressions, finds that breadth of business function use, worker task use and operational investment all carry positive and statistically significant associations with the likelihood of above-average current and future performance and with current sales increases, with functional breadth the strongest of the three. The authors are careful to call these associations rather than causes, and so am I. But the shape is hard to miss. The dimension most strongly linked to performance is the dimension most companies have skipped.
That skipped dimension has a name, and the name is AI operations.
What the function actually owns
Five things. I have never seen a deployment survive production without all five having an owner, and I have watched several die because two of them belonged to nobody.
The list. Which processes get automated, in what order, and which ones must not be. This is a judgement about the business, not about the model, and it is the decision that determines everything downstream. I have written separately on which processes to automate first.
The access. The data the system reads, the systems it is allowed to write to, and the credentials that make that real. Seven to eight percent of AI-using firms changed how they manage and store data. Everybody else is asking a model to run their operations through a keyhole.
The build. The actual integration into the tools the company already pays for, rather than a parallel universe of new tabs. Sixteen percent of AI-using firms have replaced existing software and equipment with AI-integrated solutions, and that is the healthy pattern: the new thing takes over a job the old thing was doing, instead of joining the queue beside it.
The production seat. Monitoring, failure handling, the human in the loop, the rules about what the system may do unsupervised. This is the part that gets skipped and the part that decides whether the pilot becomes infrastructure. It is why I wrote about making agents safe enough to run real operations and about why pilots never reach production.
The measure. One number per process, taken before and after. Hours, error rate, cycle time, cost per unit of work. Without it you cannot tell an expensive habit from an operating advantage, and neither can your board.
What AI operations is not
It is not a headcount reduction plan, whatever you have been told. In the Census data, 95.7 percent of AI-using firms report no employment change attributable to AI, 2.3 percent an increase and 2.0 percent a decrease. Task augmentation is the dominant effect at 44 percent of AI-using firms, task substitution about 10 percent, task creation about 11 percent, and 52 percent report none of the three. Among firms that do substitute, 71 percent replace only a small number of tasks, though the share replacing a large number has risen from 2.5 percent in the first supplement in early 2024 to 7 percent now. The direction is real. The magnitude is not what the headlines say.
It is not an IT project either. Of the firms not planning to adopt AI, 65 percent say the reason is that AI is not applicable to their business, well ahead of lack of knowledge at 22 percent and privacy or security concerns at 20 percent. A company that believes AI does not apply to it has usually only ever been shown the chat window. Somebody has to walk the actual processes before that belief can change, and that is an operating job, not a software purchase.
Who does this work
Somebody has to sit on that raised platform. The Census evidence says the least common thing American firms do is hire a person trained in AI, so in practice the function is either carried by an existing operator who now has two jobs, or it is brought in.
Bringing it in is the model the largest technology companies picked for themselves, and they gave it a name: the forward deployed engineer, a senior engineer who works inside the customer's systems and stays accountable through to production. Reporting by The Pragmatic Engineer, The New Stack, Perspective AI and Paraform puts growth in forward deployed job listings at roughly 800 percent between January and September 2025 and above 1,000 percent year on year into 2026. OpenAI launched a deployment subsidiary in May 2026 and Anthropic launched a services arm in July 2026, both built on the same idea. I counted the open roles myself yesterday and wrote up what a forward deployed engineer actually is, from 240 live postings at ten companies.
That is the seat I sit in. I spent my career as a chief financial officer before I did this, which is the reason I care about the measure more than the model, and at Beyond Elevation I run AI operations inside small and mid sized companies as a forward deployed engineer rather than as an adviser: your systems, your credentials, my accountability until the process runs without me. If that is the shape you need, the work is described at meethayat.com/services/fde.
How to tell whether you have an AI operations function
Five questions, and you can answer them before lunch. Name the processes AI touches in your company today. Name the person accountable when one of them produces a wrong answer at two in the morning. Say what changed about your workflows, your data or your systems in order to make room for it. Say what number moved, and by how much. Say which process is next and why that one.
If you can answer all five, you have the function, whatever you call it. If you can answer one or two, you have tools. On the national evidence, that puts you with the majority, which is not a comfort so much as a description of the opportunity.
If you want a second pair of eyes on what to automate first, I do a free audit call: one call, then a written list of what to automate first, what it saves and what it costs. Book it at beyondelevation.com/call/hayat.
Questions people actually ask
What is AI operations?
The function that makes AI part of how a company runs rather than a tool sitting on top of it. It owns which business functions AI is allowed to touch, the workflows rebuilt around it, the data and system access it needs, and the accountability for the result once it is live. The United States Census Bureau measures those layers separately in its 2026 AI supplement, and the separation is the whole point: 18 percent of American firms used AI in a business function in the reference period, while 64 percent of AI-using businesses had changed nothing about themselves to accommodate it.
Is AI operations the same as AIOps?
No. AIOps is AI pointed at your infrastructure, defined by IBM as the application of artificial intelligence capabilities to automate, streamline and optimize IT service management and operational workflows, and by Datadog as a discipline for finding the root cause of incidents and cutting alert noise. AI operations in the sense an owner means is AI pointed at the running of the company: finance, sales, service, administration. Same two words, different department, different buyer.
How is AI implemented in a business?
Today, mostly without any structural change, which is why so little of it lands. Sixty four percent of AI-using American firms made no organisational adjustment at all. Training staff and developing new workflows were each reported by about 15 percent, and changes to data management and storage by 7 to 8 percent. Implementation that survives runs the other way round: one business function, the workflow rebuilt around what the model can genuinely do, real data access, a named owner for the output, one number measured before and after. The tool is the last decision, not the first.
Which business functions use AI the most?
Among American firms using AI in at least one function: sales and marketing 52 percent, strategy and business development 45 percent, information technology 41 percent, research and development 40 percent. Production, sourcing and supply chains, quality control and distribution trail. Breadth is rare. Fifty seven percent use it in three or fewer of the fifteen surveyed functions and about 1 percent use it in all fifteen.
Does AI operations mean job cuts?
Not on the current American evidence. Of firms using AI, 95.7 percent report no employment change attributable to it, with increases and decreases each around 2 percent. Augmentation leads at 44 percent of AI-using firms against roughly 10 percent substituting a task a person used to do. Meanwhile 16 percent have replaced existing software and equipment with AI-integrated solutions. In this phase, AI is a substitute for legacy capital more often than for labour, which is a different conversation to have with your team and a better one.
Do I need to hire someone for AI operations?
You need the work owned, not necessarily a new employee. Hiring staff trained in AI is the least common organisational adjustment American firms make, while functional breadth and operational investment are the dimensions most strongly associated with above-average performance. So the work matters and the headcount is optional. A fractional or forward deployed operator covers it for a small or mid sized company, and the test for either is simple: do they have write access to your systems and do they stay past go-live.
Where these numbers come from
The firm, function, task, adjustment, employment and capital figures are from The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks, by Kathryn Bonney, Cory Breaux, Emin Dinlersoz, Lucia Foster, John Haltiwanger and Aditya Pande, Center for Economic Studies working paper CES 26-25, April 2026, drawn from the 2026 AI supplement to the United States Census Bureau's Business Trends and Outlook Survey with a reference period of November 2025 to January 2026. The overall usage range and the figures by firm size are from the Census Bureau story Large Firms With at Least 20 Employees Biggest AI Users, May 2026. The AIOps definitions are quoted from IBM and Datadog, both read on 12 September 2026. The forward deployed hiring growth figures and the OpenAI and Anthropic launches are from reporting by The Pragmatic Engineer, The New Stack, Perspective AI and Paraform, and are flagged as reporting rather than as something I counted. The working paper is a Census research paper and carries its own caveat: it has not undergone the review accorded Census Bureau publications, and its regression results are described by the authors as associations, not causes. I have kept them that way here.