Skip to content
4 min read

AI Colleagues After Probation: What Really Works Today

Loose AI experiments have become fixed roles: narrow tasks, rules enforced in the system, and checks that never stop. What holds up today.

AI Colleagues After Probation: What Really Works Today

Our brilliant yet clueless AI colleagues are no longer experiments. Each one now has a fixed role with a clearly bounded area of work. This works because we describe the task precisely, enforce rules in the system, and check the results again and again.

What happened to the AI colleagues from back then?

Our series “Brilliant yet Clueless” followed the journey across four parts, from the new colleague through its superpowers and failures to the user manual. At the time, we focused on useful instructions, relevant context, and clear rules. The specialists were prompt files that we loaded as the task required.

Today, we divide the work more deliberately. Each AI colleague has an exact job description, receives only the context it needs, and may use only the tools required for that task. An assignment has a clear starting point, an expected result, and a point at which the colleague is either finished or needs a person.

The way we cut up the work is the real difference from back then. The areas of work are now small and clear enough for us to verify every result.

Where do the AI colleagues support us in practice?

In the Widget Builder, one colleague takes the first iterations off our hands. The AI Editor turns an instruction into a concrete widget proposal. The preview shows the result in its real usage context. During development, the AI colleague also looks for missing states, tests narrow and wide layouts, and confirms that the agreed tests pass. This gets us from an idea to something we can properly assess more quickly.

Another AI colleague partially automates our own accounting. It assigns incoming documents to a case, checks fields and relationships for contradictions, and prepares the next step. The workflow may handle clearly defined internal assignments within documented rules. For external payments, missing information, or unclear cases, it stops and waits for individual human approval.

Both follow the same principle: a small assignment, known data, permitted tools, and a result that we can check. What matters is the shared way of working, not the individual use case.

Why do narrow tasks work better?

“Take care of the accounting” is not a usable assignment. “Assign this document, check these details, prepare a proposal, and stop if anything conflicts” can be carried out and verified.

For each task, we define the permitted input, the data and tools that may be used, the expected result, and the point where approval is required. Consequential rules do not live only in the prompt. Permissions, fixed workflows, input checks, dry runs, and logs enforce them as well.

After an approved action, we read the new state back and compare it with the expected result. A technically successful request is not enough. What matters is whether it produced exactly the intended change.

Why is continual adjustment part of the job?

No AI workflow is finished after its first good result. New document formats, different input, and rare exceptions reveal boundaries that were not visible during setup. We therefore validate proposals, permitted decisions, and completed changes repeatedly with real cases.

When a result does not fit, we review the case, change the rule or the workflow, and run that exact case again. The earlier cases must continue to work too. A failure becomes a new check that runs automatically next time.

The AI colleagues are therefore never fully onboarded. Validation, adjustment, and optimization remain part of their job. The same applies to us: we maintain the job descriptions, recognize new boundaries, and decide when an area of work may be expanded.

What remains human work after probation?

People define what a good result means, which consequences an action may have, and which uncertainty requires approval. In accounting, external payments and unclear cases remain with us. In the Widget Builder, we review the proposal in the preview and through tests before it becomes a published change.

The AI colleagues handle preparation, repeatable checks, and clearly described steps. We keep the responsibility. Loose experiments became fixed roles because we separated the tasks more clearly. The job descriptions became narrower along the way, and the results easier to verify.

If you want to find tasks in your daily work that can be bounded just as clearly, talk to us. Together, we can examine the specific workflow, the rules it needs, and the decisions that must remain with a person.

What Is the Next Step in Your Company?

Our page AI Consulting and AI Adoption explains how an AI rollout works, what drives the cost and how data protection and the works council come in. Get in touch if you want to plan the next step concretely.

SH
Partner & Software Architect

Over 20 years of experience with development, SharePoint, and Microsoft 365. Sascha advises companies on implementing and optimizing digital workplaces – from strategy to technical implementation.

Learn more about the author →

Ready to Talk?

Let's talk about your project – no obligation, completely free.

Free Initial Consultation