Stop prompting. Start handing over processes to AI agents
Since September 2026 the models can drive a computer end to end. The bottleneck is no longer the technology, it is knowing which work is delegable, how to describe it and where a human still has to approve. This programme builds that skill on one of your own real processes.
What is an AI agent, and why now?
An AI agent is a system that takes a goal and completes it across several steps without a human in between. A chatbot answers a question. An agent does the work: it researches, reads your files, drafts the output and stops at the approval point you defined.
This stopped being a demo in September 2026. OpenAI announced GPT-6 Astra on 3 September, the first widely available model that uses a computer on its own, scoring 0.86 on the OSWorld 2.0 computer use benchmark and cutting tasks that took 75 minutes down to about 40. Anthropic released Claude Fable 5.1 on 1 September, aimed at long running agents, with cache reads cut from 1 dollar to 0.25 dollars per million tokens.
The practical consequence for a company is uncomfortable: the gap between a team that hands processes over and a team that still writes prompts is no longer a productivity difference of a few percent. Details of both announcements are in my notes on GPT-6 Astra and the September 2026 AI roundup.
Chatbot
"How should I reply to this customer?"
You take the answer, edit it and send it. The work stays with you.
Agent
"Go through the pricing questions in the inbox, pull the customer history from the CRM, draft the proposal and put it in my approval queue."
It builds the steps. You review the result.
Which work is worth handing over?
Agents do not help everywhere. The honest version of this table is the first thing we go through in the discovery session.
| Good fit | Keep it with people | |
|---|---|---|
| Repetition | The same shape of work every week | One off, every case different |
| Sources | Known files and systems | Scattered, mostly in someone's head |
| Verification | A person can check it in minutes | Only provable months later |
| Cost of a mistake | A corrected draft | Legal, financial or reputational |
| Examples | Proposal drafts, monthly reports, recurring replies, content pipeline, screening, stock alerts | Legal decisions, performance reviews, pricing policy, crisis communication |
The four-week pilot
One process, one working agent, one measured result. Programmes that start with three processes usually finish none.
Pick the process
Repetitive work eating at least five hours a week, with a clear output. Chosen live in the discovery session.
Write it down
Steps, sources and the definition of a correct result. Most companies stall here, and it has nothing to do with AI.
Build the first version
Built together with the people who own the process, not handed to them finished. Ownership is what keeps it alive.
Place the approvals
Nothing leaves the company without a human. Plus a cost ceiling and a stop rule, which are technical necessities, not nice to haves.
Measure for four weeks
Hours before and after, error rate, and how often a human had to step in. An unmeasured pilot cannot be defended.
Where it pays off, by department
Six scenarios that come up most often with companies. Yours is built on your own workflow, not on this list.
Sales and proposals
Incoming requests classified, a draft proposal built from past pricing data, follow up emails queued for approval.
Customer service
Recurring questions answered from a source document, complex ones routed to the right person, conversation summaries logged.
Finance and reporting
Data pulled from several systems into a first draft of the monthly report. Usually the fastest payback in the building.
Marketing
Not content generation, a content pipeline: from brief to draft, draft to visual, then into the publishing calendar.
People and hiring
Application screening, interview notes turned into structured summaries, onboarding packs assembled.
Operations
Order, stock and supply thresholds watched, with a ready action draft attached to every alert.
The three mistakes that kill agent projects
1. Starting with the tool
When "which agent platform should we buy" is asked before "which process are we handing over", the project almost always ends as an unused subscription. The process decides the tool, never the other way round.
2. Building it without training the team
One person builds the agent, the team never uses it. If the people who do the work do not speak the same language about it, the system is abandoned within three weeks.
3. Ignoring shadow AI
Before the company decides anything, employees are already using AI on personal accounts and pasting company data into it. Not setting a rule does not remove the risk, it only makes it invisible.
Agent Discovery Session
The entry point. We pick the process together and decide honestly whether it is worth handing over at all.
4.900 TL / 90 minutes, online
- Your candidate processes scored against the delegability table
- A live first version built on the strongest candidate
- Data policy and shadow AI check
- Recording and a written next-step list
Four-Week Agent Pilot
For companies that want one real process actually running, with the team able to maintain it afterwards.
Tailored proposal by how many systems the process touches
- The process documented and the agent built with its owners
- Approval points, cost ceiling and stop rules in place
- Four weeks of real-condition running and measurement
- Team handover document and 30 days of support
Session price includes VAT. Pilot proposals are quoted plus VAT. If your team has not had a foundation day yet, start with corporate AI training.
Which of your processes could an agent take over?
Describe one repetitive process in a couple of sentences and I will tell you honestly whether it is worth handing over, free of charge.
Frequently asked questions
What is an AI agent?
An AI agent is a system that takes a goal and completes it across several steps without a human in between. The difference from a chatbot is simple: a chatbot answers, an agent does the work. It can research on the web, read your files, draft an output and stop at the approval point you defined.
Is this training different from your corporate AI training?
Yes. The corporate AI training is the foundation day: how the whole team uses AI in daily work. This programme is the next step and assumes that foundation. Here we pick one real process, write it down, and build the first working agent on it together with the people who own that process.
Does my team need to know how to code?
No. The work is describing a process precisely, defining what a correct result looks like and placing the approval points. Where code is genuinely needed I write it during the sessions and hand it over documented.
Which processes are worth handing over?
Repetitive work with a known source and an output that is cheap to verify. Proposal drafting, recurring customer questions, monthly report assembly, content production pipelines, application screening and stock alerts. Judgement heavy work with scattered sources and expensive mistakes stays with people.
How long does a pilot take?
Four weeks. Week one picks and documents the process, week two builds the first version, weeks three and four run it in real conditions and measure hours saved, error rate and how often a human had to step in.
Which models and tools do you use?
Whatever fits the job and your data policy. In practice that is mostly Claude (Anthropic) and ChatGPT, and since September 2026 the computer using models such as GPT-6 Astra. I do not sell a platform, so the recommendation follows the process rather than a licence.
What about our data security?
The first session covers this before anything is built: which data may leave the company, which must not, which plan gives you a no training guarantee and how to close the shadow AI gap where employees already paste company data into personal accounts.
What does it cost?
It starts with a 90-minute discovery session at 4.900 TL, where we pick the process and confirm whether it is worth handing over. The build programme is scoped per company and quoted as a tailored proposal, because the work depends on how many systems the process touches.