There are plenty of opinions about AI in business and very few facts. But one source lies less than the rest: the job ad. A company writes it not to make a point, but because it is ready to pay a salary. We collected 98 of them and read every one in full.
Developer hiring wasn’t what interested us — that market is fairly predictable. We were after the layer above it: the people companies hire to get AI into their actual processes. Implementation specialists, heads of AI, project managers, product managers, analysts, consultants.
The ads are collected by our agent: it queries the hh.ru API against job titles and saves each one in full, description and requirements included. Out of what it brings back, we kept the adoption and management roles and dropped pure development. So this isn’t market statistics — it’s a slice through one particular filter: what companies wrote about AI at the moment they opened a role and a budget for it. A full description of the work and requirements exists in 91 of the 98, and every percentage about content below is calculated on those.
This snapshot is dated 10 September 2026. The agent keeps collecting, so we’ll refresh both the numbers and the conclusions. Every update is stamped with its date in the article header.
Here’s what it shows.
Business wants an implementer, not an engineer
The first thing that stands out: pure development is a minority in this sample.
Add up everything that isn’t an engineering role and you get 65 against 22. Companies aren’t looking for someone to build a model. They’re looking for someone to carry an existing technology into a working process and keep it running there.
And the hiring isn’t coming from the tech industry. The sample runs across manufacturing, retail, construction, mining, agriculture, banking, breweries, furniture plants, clinic chains and dentistry. Companies for which AI isn’t a product but a cheaper way to do what they were already doing. A third of the roles sit outside Moscow and St. Petersburg — Ufa, Kazan, Tomsk, Arkhangelsk, Lipetsk, Ozersk.
That’s a real shift. A year ago this conversation lived mostly inside technology companies. Today a plant in a company town is opening a role for it.
The most honest thing in these ads is what’s missing
Read the requirements back to back and the surprise isn’t the list of technologies. It’s the gaps in it.
Almost nobody intends to build their own AI. Research, model training, custom weights, expensive hardware — practically none of it appears in the requirements.
What they ask for instead is something else entirely.
The reading is unambiguous. The intelligence is bought from outside; the value is created where it connects. The model belongs to someone else — the data, the processes, the systems of record and the accountability are yours.
Which explains the requirement sitting at the top of the list: reach. Forty-two percent name CRM or ERP outright. Useful AI isn’t the kind that answers nicely in a chat window. It’s the kind that can see the orders, contracts, stock levels and tickets, and write results back into the system where the business actually lives.
Agents went mainstream faster than the habit of checking them
Half the ads talk about agents — not a chatbot, but a system that takes steps on its own, calls tools and does something with data. One in six mentions MCP, a protocol less than a year old. Phrases like “a digital employee that independently handles routine operations” show up as casually as “set up a data export” once did.
Now put the second number next to it.
Hallucinations, output validation, quality metrics, testing responses — all of that appears in one ad out of seven. Confidentiality, personal data and access control show up in one out of six, and this is for a role that by definition gets access to internal systems.
We don’t read this as carelessness. It looks more like the age of the market. Verification tends to enter the requirements after the first time something wrong reached a customer. Most of the companies here haven’t had that moment yet.
For a candidate, incidentally, this is the most valuable ground to stand on in an interview. If you arrive with an answer to “how will you know it’s working correctly,” you’re answering a question the employer hasn’t yet learned to ask.
The top is accountable for the economics, the bottom builds the thing
A third of the ads require measuring the impact of adoption: economics, ROI, payback period, hours saved. But the requirement is distributed very unevenly.
So you get a familiar arrangement. The person defending the budget is obliged to justify the money, while the person actually choosing the solution barely sees money mentioned in their own job description. Between the two sits a gap — and that gap is usually where you find the projects that work technically and make no economic sense at all.
This is the case where a job ad accidentally describes the company’s future problem.
Adoption is work with people, not with a model
Here’s the part that surprises anyone who pictures AI adoption as a technical project.
More than half the implementation roles require training employees: running short sessions, writing instructions, checklists and internal rules for using AI. Forty-one percent of all ads mention operations — monitoring, logging, support, keeping the thing stable after launch. For project managers it’s the single biggest item.
Companies have worked out something important: building the workflow is the smaller part of the job. What follows is the unglamorous part. Someone has to stop doing it the old way. Someone has to notice the agent broke. Someone has to update the instructions when the process changes.
Fifteen percent describe the role as the first AI person in the company: a function built from scratch, backed by the owner, with a tools budget and a short decision cycle. That’s an honest and difficult brief. That person will spend less time programming than explaining, negotiating and defending results.
The market doesn’t know what this costs yet
Three quarters of the ads don’t name a salary. Among the twenty-three that do, here’s the picture.
Midpoint of the stated range, monthly, normalised to take-home pay. 23 ads out of 98 — the rest disclose nothing.
A four-and-a-half-fold spread across requirements that read almost identically. Two ads can ask for the same things — agents, RAG, ERP integration, measurable results — and differ threefold in pay.
That isn’t stinginess or generosity. It’s the absence of a reference point. The market hides the price because it hasn’t decided what it’s buying: an engineer, a consultant, a manager, or one person who covers all of it.
The same uncertainty has another side. Thirty-six percent don’t want a CV — they want proof: a case study, a portfolio, links to your own projects, a test assignment. The wording gets blunt, up to “applications without a specific case will not be considered.” The line “experience with AI” has stopped meaning anything, and employers have noticed.
Who they’re really looking for
Collapse the requirements into a single portrait and you get someone who sits right on the seam between the two.
They should be able to take a process apart, talk to the department that owns it, assemble a solution from ready-made pieces, connect it to the accounting system, calculate the impact, write the instructions and train the people. The median ad lists three tools; one in five lists eight or more, from Python and SQL to n8n, vector databases, BI dashboards and CRM.
Experience is a story of its own. Thirty-six ads out of ninety-eight want three to six years, seven want more than six. In a field that has existed in its current form for about two. What they mean is clear enough — general experience in IT, automation or project management. But formally the market is asking for experience nobody can physically have, which is one more sign that these requirements are being written on autopilot from older IT templates.
What to do with this map
Three different conclusions from the same numbers.
If you own or run the business. Opening an AI role before you’ve named the process is the most reliable way to lose a year. These ads show clearly which roles the market has learned to describe and which it hasn’t — the vaguest ones are precisely those that name neither a process nor a way to measure the result. Before hiring, answer four questions: which flow of work changes, who inside is accountable for the outcome, who checks the quality, how you count the savings. With those answers the job ad writes itself. Without them you’ll hire someone who has to go looking for a purpose, which rarely ends well.
If you’re hiring. “Experience with AI” is no longer a filter. What works is asking for one adoption described end to end: the original problem, what the person did themselves, what they built it with, what the numbers were, what broke afterwards. And check your own ad for the two gaps typical of this market: does it say who is responsible for verifying quality, and does it say what happens to the solution after launch? Strong candidates read exactly those lines.
If you’re the candidate. Demand has moved away from models and toward assembly and adoption: integrations, agents, RAG, systems of record, measurable impact. One case carried all the way into daily operation, with numbers, outweighs any list of technologies. And close to half of these roles expect you to hold a conversation with the business, not just build — that isn’t a bonus, it’s the main part of the job.
The overall picture strikes us as encouraging. The market has stopped buying AI as a miracle and started buying it as work: with a process, an owner, written instructions and a number at the end. Quality control and economics are lagging behind, but that’s the lag of a young market rather than a dead end. Requirements like those tend to appear right after the first painful experience — and at this pace, they’ll appear soon.
The next update comes once a meaningful new batch has accumulated. If the picture shifts, we’ll rewrite the conclusions, not just the figures.