# AI doesn't need a better model. It needs onboarding like every new colleague.

Jaro Rais · 8 min read · 2026-10-19 · Working with AI
Source: https://itashu.co/en/know-how/ai-does-not-need-a-better-model

Companies debate whether ChatGPT or Claude is better. But the difference between AI that helps now and then and AI that really works lies elsewhere: in how you onboard it in your company. We explain it with something every manager knows: a new colleague's first weeks.

1. Brief
2. Context
3. Rules
4. Triggers
5. Team
6. Department

AI as an assistant → AI as an employee → AI as a department

Developers call this **[agentic engineering](https://itashu.co/en/know-how/six-ai-terms)**. It has six steps that build on each other like stairs. On the first, AI answers questions. On the last, it handles a whole process from brief to result.

**Every step is a management decision, not a technical one.** You make the same decision with every new hire: what you ask them to do, what you give them, what you allow them and when you let them work on their own.

## 01 Prompt engineering: a good brief does more than a bigger model.

A prompt is the brief AI receives. Prompt engineering means writing it so that AI knows exactly what you want, for whom and at what quality. It is the only step that costs nothing extra: same model, better brief.

![A person hands a robot at a desk a precise brief with an example](https://itashu.co/img/art-00-k1-1280.webp)

- **With a new colleague:** You don't tell them "make a quote". You tell them who it is for and how big it is, and show them a past quote that worked.
- **With AI:** The same applies. A better brief comes from someone who knows the field, not someone who knows AI. Briefing by voice helps: a few minutes of talking gives AI more information than two lines of text.

> **In practice:** “A law firm gives AI two sample replies to clients and one rule: never promise a date. Suddenly the replies sound like the firm, not like a textbook.”

**Ask yourself:** Are your best briefs written down somewhere, or does everyone keep them in their head?

## 02 Context engineering: AI works only with what you provide.

Context is everything AI sees during a task: documents, data and history. Context engineering means choosing these materials so that AI has [what matters and nothing more](https://itashu.co/en/know-how/more-documents-do-not-mean-better-answers).

![A robot works with one client's folder while a helper robot brings a summary of a long document](https://itashu.co/img/art-00-k2-1280.webp)

- **With a new colleague:** You don't hand them the whole company archive. You give them the folder for their client. When 200 pages need reading, someone prepares a one-page summary.
- **With AI:** AI has a limited workspace, and the quality of its answers drops even when that space is only partly full. So it gets only what it needs: the client's history from the CRM, the current price list, the latest emails. A helper agent reads long documents for it, and it summarizes long work as it goes, like a shift handover.

> **In practice:** “Before a call, a salesperson asks: ‘What do we know about this company?’ AI answers from your CRM and your communication, not from the internet.”

**Ask yourself:** Does your AI know where your data is, or does someone keep copying it into the chat window?

## 03 Harness engineering: an agent without rules is an intern with keys to the nuclear button.

The harness is the environment in which the agent works: tools, access and checks. It sets what the agent may do on its own, what it may only propose and how it verifies that it worked correctly.

![Robots at a control panel with keys and locks: what they may do alone and what they may only prepare](https://itashu.co/img/art-00-k3-1280.webp)

- **With a new colleague:** They get access only where they need it. During probation someone checks their work. They don't sign contracts on their own.
- **With AI:** We set up the agent's environment: which tools it may use, what it may only prepare for approval and how it checks its own result. Because it verifies the result itself, you don't have to approve every step. It tries new procedures first in a separate environment where it can't break anything.

> **In practice:** “An agent at a dental clinic drafts a reply to a patient and proposes an appointment. It adds it to the calendar only after checking the practice's rules. It never changes prices.”

**Ask yourself:** What may AI do on its own in your company, and what may it only propose?

## 04 Loop engineering: AI you have to start is just an assistant.

A loop is a cycle in which the agent works without a person: something triggers it, it takes a step, checks the result and continues. An assistant you have to call becomes a colleague who reacts on their own.

![A robot at a conveyor handles emails and tasks that arrive on their own, a clock beside it](https://itashu.co/img/art-00-k4-1280.webp)

- **With a new colleague:** A good colleague doesn't need instructions every morning. They know that when an order comes in, they process it. And that the report goes out on Monday.
- **With AI:** The agent has triggers: an email arrived, a payment came in, it is eight in the morning. It acts on its own, looks at the result and decides whether to continue. A person approves only where it matters.

> **Our own example:** “Our sales system finds and contacts companies on its own. We join the conversation only when someone replies. Nobody starts it in the morning.”

[Lead generation](https://itashu.co/en/case-studies/lead-generation)

**Ask yourself:** Which tasks does someone in your company do every day only because they need to be ‘started’?

## 05 Graph engineering: three agents work on three parts of a task at once, not one after another.

A graph is a map that sets which agent does which part of the work and how they hand it over. Instead of one agent for everything, a team works, each on its own task.

![Three robots work at once on three parts of a task and their results come together](https://itashu.co/img/art-00-k5-1280.webp)

- **With people:** A bigger project isn't done by one person. One gathers facts, another runs the numbers, a third writes. The project manager puts it together.
- **With AI:** Agents work in parallel, each on its own part, and a coordinator combines the result. Splitting the work is the same decision as with people. But a team of agents costs more than one agent, so we use it only where it pays off.

> **In practice:** “Preparing for a sales meeting: one agent researches the company, a second its finances and hiring, a third drafts an email. The brief is ready before a person could even start putting it together.”

**Ask yourself:** Which process in your company waits only because one person does all of it?

## 06 Software factory: the result is not a tool, but a working department.

A software factory is the last step: specialized agents handle the whole process from brief to result. The term comes from software development, where agents take over the path from brief through code and tests to deployment. The same principle applies to any business process. People set priorities, check quality and handle exceptions.

![A department of robots at desks and people who set priorities and check quality](https://itashu.co/img/art-00-k6-1280.webp)

- **With people:** A well-run department has a process from brief to delivery, and everyone has a role in it. Everyone has their own desk, their own badge, and their work leaves a record. The manager sets priorities and checks quality.
- **With AI:** Instead of one agent for everything, specialized agents work, each on its own task, for example from inquiry through quote to invoice. Each needs the same as a person: a separate workspace, its own login with clear access and a record of its work. So you know what it did at three in the morning.

> **What it means for people:** “They won't stop working. They will move to where AI isn't enough: talking to clients, making decisions and taking responsibility for the result.”

**Ask yourself:** If you were building the department from scratch today, which roles would people fill?

## Most companies we talk to are on the first two steps.

Everyone uses AI their own way, nothing is connected and someone always has to start it. That is not failure, it is a beginning. What matters is knowing which step comes next.

1. **Prompt engineering:** AI is a smart newcomer who needs to be told everything. You decide what briefs are standard in the company.
2. **Context engineering:** AI is a colleague with access to the right folders. You decide which data AI can access.
3. **Harness engineering:** AI is a colleague on probation. You decide what it may do on its own and what it may only propose.
4. **Loop engineering:** AI is an independent colleague. You decide which tasks run without a person.
5. **Graph engineering:** AI is a project team. You decide how work is split and who puts the result together.
6. **Software factory:** AI is a whole department. You decide where people stay.

## Your competitors already have AI. The difference is who puts it to work properly.

Anyone can buy a Claude or ChatGPT license today for about 20 euros per person per month, including your competitors. **So your advantage doesn't come from the tool itself, but from how you put it to work.**

The six steps are not a technical list. They are a sequence of decisions made by management: what you give AI, what you allow it and how you check the result. None of this is new. It is good old management of people and processes, just with a new kind of colleague.

**How to assess it: go through the questions at each step.** Where you answer "no" or "I don't know", that is your next step. Not every company needs the sixth step. But the first three apply to everyone who works with AI, from the assistant to the owner.

---

Tell us which step you are on. We'll show you what comes next. https://itashu.co/en/consultation