Who runs AI in your company: why businesses need a Claude Code operator

Most companies already have AI. Some have paid for employee subscriptions, some have commissioned a chatbot or voice assistant from an integrator, and some have connected AI to their email and documents. Yet a few months later, management candidly admits that the impact is smaller than promised. The bot quotes outdated prices, the assistant is "doing something" but nobody can say how many hours it has saved, and half the staff have gone back to their familiar spreadsheets.

The cause is rarely the model itself. What is almost always missing is a person who assigns tasks to the AI every day, signs off on the results and is accountable for them in the same way a chief accountant is accountable for the books. We call this role the operator, and in this article we use Claude Code as an example to explain why the role is needed even after the very best implementation.

What the numbers say

The most honest picture comes from the quarterly tracker of the Federal Reserve Bank of St. Louis. According to its data, in the third quarter of 2026 generative AI saved US workers an average of about 2.1% of all working hours. That is an average across all employed people aged 18–64: those who do not use AI are counted as zero, which is why the overall figure is so modest.

The result is very different where AI has been built into specific work. The study "Generative AI at Work" (Brynjolfsson, Li, Raymond, NBER) followed the rollout of an AI assistant to 5,179 customer support agents: the number of issues resolved per hour rose by 14% on average, and by 34% among newcomers and the least experienced staff. The gap between "simply giving people AI" and "building AI into the process" is several-fold.

The MIT NANDA report "The GenAI Divide" (2025) explains why not everyone manages to build it in: according to the report, only about 5% of corporate AI pilots deliver a tangible financial result. The authors attribute this not to the quality of the models but to the fact that the tools are not embedded into the company's workflows and do not adapt to them. In other words, the problem is not technological but managerial.

The conclusion is simple: the impact comes not from access to AI but from a person who turns it into a process.

Claude Code is not a chat but an agent in your working environment

Most people know AI as a chat window: question, answer. Claude Code works differently. It is an agent that operates inside the company's working environment: it reads the project and documents, proposes a plan, makes the changes itself and shows exactly what it has changed. Through connectors (MCP) it can be linked to email, calendars, accounting systems and the website, and then it works not with a single file but with the company's real data.

That is precisely why it is more powerful than a chat, and precisely why it needs to be managed. An agent that can change a web page, a price in the database or the wording of a contract must have an owner: someone who decides what it is allowed to do and who checks the work before it reaches the customer.

Tasks you can already hand over to an agent inside the company

We work with Claude Code every day on our own products — Smart CMS, Smart 360, Smart LMS, BOT-panel — and on client websites. Here are the tasks an agent already handles reliably for us under an operator's supervision:

• Websites and content. Migrate a website from a legacy system, check every page against the original on desktop and mobile, translate the site into several languages, update the news section.
• Documents. Draw up a contract or a certificate of completion using the company's template, reconcile payments with the bank statement, prepare a document register.
• Email and calendar. Sort the inbox, prepare draft replies, find agreements buried in correspondence. A human does the sending.
• Chatbots and voice assistants. Update the bot's knowledge base when new products appear or prices change, and test its answers with control questions.
• Infrastructure. Scan a server for vulnerabilities, set up backups, monitor failures in automation scenarios.
• Reports. Pull data from several systems into a single report for management.

None of these tasks requires the operator to be a programmer. What they do require is an understanding of one's own domain and the ability to give the agent a task in a way that produces a result that can be signed off.

Where the agent fails silently

The most dangerous AI mistakes are not the ones where it refuses to work, but the ones where it confidently reports "done". A few situations from our own practice:

• "Done" is not the same as "finished". The agent completed the layout of a page and showed green checks. But on a phone, one block slid off the edge of the screen. The checks cannot see this; a person who opens the page at 390 pixels can.
• Literal execution. The agent will flawlessly carry out the instruction "replace the phone number everywhere", even if the new number in the instruction is wrong. The operator first states the old and new values digit by digit and only then launches the bulk change.
• A bulk change with no way back. A single inaccuracy in a replacement pattern can corrupt dozens of database records in a second. The operator's rule: before any bulk action, make a backup. Then a mistake costs minutes, not a day.
• Outdated knowledge. For several weeks a voice assistant kept giving customers the old names of pricing plans, because nobody updated its instructions after the plans were renamed. The model was not to blame: it was faithfully repeating what it had been given.

In every case, what saves the day is not a "smarter model" but a person with a clear working cycle.

Why you need an operator after the integrator

An integrator is a project: they install the system, configure it, train people and hand it over. But the system lives on. Prices change, new products appear, new people join, and data and security requirements evolve. Without an owner, even a perfectly implemented AI starts to lag behind the company's reality within a few months.

	Integrator	In-house operator
Format	Fixed-term project	Day-to-day work
Responsible for	Launching the system	The system's results every day
Knows	The technology	The technology and their department's processes
When prices or the product change	A new order	Updates it the same day
Access and client data	Hands over after launch	Holds and controls

This is not competition. A good integrator actually wants the client to have someone to hand the system over to; otherwise, six months later, they get called back to "fix the AI that broke".

The operator's cycle: four steps for every task

1. Task. "Make it look nice" is turned into a scope of changes and a definition of done: what exactly we are changing, what we are leaving alone, and how we will know it is finished.
2. Plan. The agent lays out its steps before changing the first line. A wrong direction is stopped here, not afterwards.
3. Diff. The operator reads the changes themselves, line by line, rather than the agent's summary.
4. Acceptance. "Done" is said after the checks — on the live page, on a phone, on real data — not instead of them.

Until this cycle becomes second nature, the agent remains an interesting toy. Once it does, it becomes a tool that takes hours of routine work off the team every week.

The "Claude Code Operator" course

We teach this role at SmartGT Academy. The course is not about the tool itself but about managing it: six weeks of practice on each participant's own tasks.

• Cohort 03 starts on 1 December 2026. 12 sessions of 90 minutes, two a week, online in Ukrainian, with recordings.
• Groups of up to 8 people: everyone shares their screen and works on their own real task rather than a training example.
• Eight modules: the operator's workstation, project context, task setting, accepting the work, skills and subagents, connectors to Microsoft 365 and Google, security and access boundaries, rolling it out across the team.
• Who it is for: developers and team leads, analysts and system administrators, managers introducing AI in their teams, lawyers and anyone who works with documents. Programming skills are not required; you need basic command-line skills and a Claude subscription, which we will help you set up.
• After the course — a support chat for another month.
• Price: $500 · UAH 18,000 per participant, $400 · UAH 15,000 for existing SmartGT clients, $3,200 · UAH 120,000 for a corporate group of up to 8 people. Payment under a contract; you can withdraw after the first session.

The cohort is led by Mykhailo Vasianovych, founder of SmartGT, who works with Claude Code every day on the very products that serve as material for the practice sessions.

Ranks: a qualification that grows with practice

A "course completed" certificate says nothing about what a person will actually do on Monday. That is why we maintain a rank scale for operators: ten levels, each confirmed by hours of work with the agent, the number of working sessions, tokens used and a practical assignment.

Rank	Threshold	What the person does unsupervised
R0 · Connected	first conversation	Tool installed, permissions configured
R1 · Intern	6 h · 3 sessions	Supervised tasks; stops the agent when it heads in the wrong direction
R2 · Operator	14 h · 7 sessions	Runs a task alone: from setting it to an accepted diff
R3 · Reviewer	26 h · 13 sessions	Reviews the agent's work done by others
R4 · Skill Master	42 h · 21 sessions	Captures the department's routines in the team's tools
R5 · Integrator	60 h · 30 sessions	Connects the agent to email, calendars and accounting systems
R6 · Warden	80 h · 40 sessions	Responsible for access, secrets and isolation of client data
R7 · Conductor	95 h · 48 sessions	Runs several agents in parallel and can train operators
R8 · Architect	180 h · 90 sessions	Owns the process in the department: branches, checks, procedures
R9 · Mentor	350 h · 175 sessions	Trains operators and upholds the standard across the company

Participants earn the first two or three levels during the course itself, through practice on their own tasks. After that, the rank grows with the work: a short monthly survey on what they have managed to build into their processes, and a quarterly assessment for the next level. The rank is not included in the course fee: it cannot be bought, only earned.

Each rank comes with a pass bearing a permanent code of the form OCC-XXXXX and its own page that an employer can open and verify: the rank, hours of practice, sessions and the history of levels. Here, for example, is the cohort trainer's badge — rank 8, "Architect". Operators can be trained from rank 7 ("Conductor") upwards: a person who confidently runs several agents themselves is able to teach colleagues to do the same. For a company, this is a clear criterion: a job posting can specify the required level, access to secrets and client data should be granted no earlier than level six, and training new operators can be assigned from level seven.

Where to start

Not with a big project. Pick one task that takes someone on your team several hours every week, and a person who knows that area well. That is your future operator. On the course, they will complete that task with us, and after that they will handle the next ones on their own.

Course programme and registration for cohort 03 · Message us on Telegram — if you want to train a whole team at once, we will set up a corporate group for your department.

Sources

• Federal Reserve Bank of St. Louis, FRED: Generative Artificial Intelligence, Time Savings: Employed Adults (RPSGENAITSALL), data for the third quarter of 2026.
• Erik Brynjolfsson, Danielle Li, Lindsey R. Raymond. Generative AI at Work, NBER Working Paper 31161.
• MIT NANDA, "The GenAI Divide: State of AI in Business 2025", as reported by Virtualization Review.

SmartGT is an IT team providing comprehensive business solutions. SmartGT Academy is an independent educational project. Claude and Claude Code are trademarks of Anthropic; we are not an official partner of the company.