Three quotes, two colleagues' opinions, and six marketing pages – and no one can say which source to trust. Imagine someone hangs a comparison chart titled 'truth' on the break room wall. It's a fictional example. In this video, you'll build an agent that sources AI customer service tools for you – and an inspector that rejects the work if the evidence is insufficient.
Here's the difference. When you ask a chat 'what is the best AI customer service tool', you get a smooth answer – without knowing where it came from. The agent gets a goal, tools, boundaries, and acceptance criteria. It plans the steps, searches, opens, compares, and records sources, and stops if the evidence is insufficient. The workflow is: search, compare sources, compile evidence. The example company is fictional: a 30-person online store, about 400 customer messages per week.
Goal, tools, boundaries, definition of done
Start with the task. You can copy this as is: 'Investigate three AI customer service tools suitable for a 30-person online store's customer service, about 400 messages per week. Compare Finnish language support, integrations with email, e-commerce platform, and CRM, data security and GDPR description, pricing model, and the work required for implementation. Use only public sources. Mark each claim with a source and retrieval date. Separate fact and estimate. Do not guess prices. Return a comparison table, source list, and open questions. If a source is missing, stop and report.' Note: no passwords, no customer personal data – they are not needed here.
Search, open, and save sources with dates
The agent's tools are search, page reading, table, and notes. It first plans which claims need verification. Then it searches and opens primary sources: the tool's own documentation, pricing page, privacy page, and release notes. It reads long documents in one go and extracts sections discussing Finnish language, integrations, and data processing. Each line gets a source and date. It can also extract information from a scanned price list or price list screenshot – but marks the image as the source, not its memory.
Unknown is marked as unknown
The comparison table has one row per tool and one column per criterion. The most important rule: unknown is marked 'unknown', not as a guess. The vendor's own marketing claim is separated from verified information – 'vendor promises' is not the same as 'found in documentation'. Estimates are marked as estimates and their assumptions are documented.
Evidence package and open questions
The result is an evidence package, not a sales pitch: claim, source, retrieval date, and a note on whether it's a fact or an estimate. It includes a list of open questions to ask vendors – for example, data storage location, deletion logic, and implementation responsibilities. The safe next step is not a contract, but a limited trial with the same questions for three vendors.
Inspector agent rejects if evidence is missing
This requires a separate inspector agent. It does not see the original marketing material, only the report and acceptance criteria. A typical mistake: the agent has marked the price as a fact, but the source is a comparison blog, not the vendor's own price list, and an estimated savings percentage has slipped in without a source. The inspector rejects: correct the sourcing, change the percentage to an estimate or remove it, and add the retrieval date. If even one point remains open, the status is 'not accepted' and the report is not published. The inspector never fills gaps with guesses – it stops.
Take away three things. One: the task defines – goal, tools, boundaries, and definition of done are documented. Two: source and date for every claim, and unknown is marked as unknown. Three: a separate inspector with the power to reject the work.
Exercise
Select three public AI customer service tools. Give the agent the task from the video, but replace company information with fictional example data (e.g., 'Example Ltd, 20 people, 300 customer messages per week'). Do not provide passwords or customer personal data. Then run the inspector agent with the acceptance checklist and correct at least one incomplete source citation or one fact-marked estimate.
Checklist
- Every factual claim has a source and retrieval date, and the source is primary (vendor's own documentation or price list), not a comparison blog.
- Unknown information is marked 'unknown' – it is not guessed – and estimates are separated from facts with assumptions.
- The material contains no passwords, credentials, or unnecessary personal data.
- The inspector agent has run through the acceptance checklist and the work status is either 'accepted' or 'not accepted' – nothing in between.
- If even one point remains open, the report is not published and the gaps are reported as open questions.
— Aino Pilvi, AI-asiantuntija, Cloud Computing Oy
Kysyttävää? aino@cloudcomputing.fi | WhatsApp: +358 33 119 3196
