Assessment and objectives
We map the steps, exceptions and time spent on manual tasks. We define an acceptable error rate and the time savings needed to make the trial worthwhile.
02 / WORKING TOGETHER
How we test automation
Building a prototype
I build a limited proof of concept (PoC) using agreed samples. Depending on the task, I choose API integration, rule-based automation or AI processing. Connecting systems may eliminate repeated data entry.
Testing and measuring together
We use representative test cases (evals) to measure quality, time and cost. For AI processing, I use structured outputs with a defined schema, while checking content accuracy separately.
Deciding how to proceed
We decide how to proceed based on the results. We define human review points (human-in-the-loop), error handling and the conditions for production deployment.
The colleagues doing the work provide examples of routine and exceptional cases. I agree the checking rules with them, and they test the solution before we evaluate it.
03 / DELIVERABLES
What do you receive, and how can you use it?
- Process description: distinguishes steps that can be automated from those requiring professional judgment.
- A prototype that runs on sample cases: colleagues can evaluate it against their own tasks.
- Comparison: total time, correction effort and usage costs become visible.
- Deployment recommendation: decide whether to proceed, revise the solution or end the trial.
The prototype helps you decide whether automation is worth deploying. Production deployment tasks, responsibilities and operating conditions are defined in a separate agreement.
04 / GETTING STARTED
What is needed to get started?
- One selected process and a contact familiar with its daily operation
- Lawfully usable, preferably anonymized samples and examples of exceptions
- Agreed tools, access and data handling. Confidential data may only be sent to external AI services with explicit approval
- Predefined success criteria, human review points and budget
05 / FREQUENTLY ASKED QUESTIONS
Frequently asked questions
How long does it take to build a prototype?
I estimate the work and time for one selected process. We tailor the trial scope to the data and systems available.
Does every automation need AI?
No. With predictable rules and structured data, system integration or conventional automation may be sufficient. I recommend AI when, for example, interpreting varied text is a substantial part of the task.
What if the trial does not deliver the expected results?
The evaluation identifies what limits the solution, such as data quality, too many exceptions or correction costs. We can then decide whether to revise it or end the trial. The trial does not commit you to deployment.
Could an AI agent be part of the solution?
An AI agent may be appropriate when several steps and tools need coordinating. Tool calling is restricted to operations approved in advance. Sending external messages or changing business data requires human approval during the trial.
Is production deployment included?
The prototype runs in an agreed test environment. Changes to production systems, deployment and ongoing operation require a separate agreement.