1. Start with the evaluation problem, not the technology

A common mistake when building AI workflows is to begin with the available tool: ChatGPT, Claude, Make.com, Google Sheets or another platform.

For evaluation work, reverse the order.

The first question is therefore:

For this tutorial, use the following fictional evaluation question:

The purpose of the agent is not to decide whether the programme was successful. It is to support the evaluator by systematically examining the evidence relevant to the question.

2. Decide whether you actually need an agent

Not every AI-assisted task requires an agent.

Design principle: Do not call something an agent simply because an LLM appears in the workflow. The architecture should earn the label.

3. Define the agent’s actual analytical responsibility

Now specify what the system should do.

4. Build a realistic evidence base

This is where many simple AI workflows become methodologically weak. If the evaluator supplies an already-cleaned evidence table containing the final findings, the AI has little analytical work left to do.

For this exercise, therefore, create a small but deliberately imperfect evidence base.

The evidence is intentionally incomplete. There is no evidence proving why the reach target was missed.

This creates an opportunity for the agent to demonstrate whether it can distinguish:

5. Choose a privacy-safe environment

You do not need Google Sheets or Make.com to complete this tutorial.

Use synthetic data unless your organisation has explicitly approved the AI environment and the type of information being processed.

Option: Claude or ChatGPT

1. Create a dedicated conversation or project for the exercise.
2. Use only the fictional evidence package above.
3. Paste the evidence into the conversation or upload a synthetic document if permitted.
4. Provide the agent specification.
5. Run the initial analysis.
6. Introduce deliberately difficult cases.
7. Review the results against the source material.
8. Revise the system instructions.
9. Repeat the tests.

6. Give the AI a methodological task

Instead of asking the model to “analyse the data”, provide explicit analytical responsibilities.

7. Preserve the evidence chain

A reliable evaluation workflow should preserve the path from source material to judgement.

Ask the AI to keep these stages separate.

This distinction becomes particularly important in reporting, recommendations and evaluative synthesis.

8. Introduce triangulation

Now ask the agent to compare the sources.

The AI might identify that:

• Monitoring data indicates that numerical reach was below target.
• Staff accounts indicate outreach delays.
• The participant account indicates access difficulties.
• The survey indicates high satisfaction among respondents.

These findings should not automatically be treated as contradictory. They may describe different dimensions of programme performance.

This is an important methodological distinction: triangulation is not simply asking an AI to compare four columns. The evaluator must determine whether the sources address the same construct, population, time period and analytical question.

9. Ask the agent to identify what it cannot know

This is a critical test of reliability.

Give the system this statement:

Ask:

The evidence mentions access difficulties and outreach delays, but it does not establish that transport problems caused the shortfall.

If the AI accepts the statement as fact, that is not merely a “bad answer”. It is a documented agent failure that needs to be addressed in the design.

10. Add conditional routing

A more meaningful agent does not simply produce an answer and stop. It can classify the state of the evidence and determine what should happen next.

Example routing rules:

• PASS: Evidence is traceable and no major unresolved issue is detected.
• FLAG: Evidence is missing, contradictory, ambiguous or potentially overstated.
• HUMAN REVIEW: The evaluator decides whether the issue changes the finding or conclusion.

This is where the workflow begins to demonstrate genuinely useful agentic characteristics: the system is not merely generating text; it is assessing the state of the task and determining whether further action is required.

11. Test the agent with failure cases

Never evaluate an AI agent only with ideal examples.

12. Create an evidence-checking log

For every important AI output, compare the claim with the underlying evidence.

13. Introduce human methodological supervision

The evaluator should not simply approve or reject the AI’s final paragraph.

Human review should occur at the points where methodological judgement matters.

14. Connect the prototype to the evaluation cycle

The same architecture can support different stages of an evaluation, but the methodological task changes.

15. Only now consider Make.com

Once the analytical workflow has been tested manually, you can decide whether automation is useful.

Make.com can automate the movement of information between systems and trigger actions. It should not determine the evaluation methodology.

This distinction matters because an automation platform does not automatically make a workflow agentic.

A workflow that watches a spreadsheet row, sends a prompt to an LLM and writes the answer into another row can be useful. But its value comes from the quality of the underlying task design, not from the number of automation steps.

16. Understand the ChatGPT/API boundary

If you later connect an AI model through an automation platform, remember that this is a separate model interaction.

Therefore, if a workflow requires evaluation context, definitions, evidence, coding rules, methodological criteria or previous analytical decisions, those elements need to be deliberately designed into the system.

Do not assume that the model “remembers what you discussed earlier” simply because you are using the same AI provider.

17. Build an agent test set

Before applying the system to real evaluation work, create a small benchmark.

18. Record failures, not just successful outputs

An agent becomes more reliable through structured testing and revision.

19. Your final challenge

Now design your own privacy-safe AI evaluation agent.

Choose one substantive evaluation task:

• evaluation design review;
• qualitative evidence analysis;
• quantitative analysis support;
• data quality assessment;
• evidence synthesis;
• indicator monitoring and early warning;
• recommendation analysis;
• learning synthesis.

Do not start by selecting Make.com modules. Start by documenting:

1. The evaluation problem.
2. The evaluation question or analytical objective.
3. The existing human workflow.
4. The methodological decisions involved.
5. The evidence required.
6. The role AI can safely perform.
7. The decisions AI must not make.
8. The information that must remain traceable.
9. The validation rules.
10. The conditions requiring human intervention.
11. Three normal test cases.
12. Three difficult or adversarial test cases.
13. At least one expected failure.
14. The improvement you would make.
15. The retest result.

20. Final checklist