Alexey Shurov.
Service

A monthly test for your customer facing AI assistant

Once a month, I ask your website or phone assistant ordinary customer questions. You get what it got wrong, its exact words and what to change. The first test is free.

Updated 12 October 2026 . 5 min read . Alexey Shurov

The short answer

Your assistant can sound helpful and still give a customer the wrong answer. It might promise a service you do not offer, miss the point of a complaint or contradict what it said earlier. A polite answer alone is not enough.

Once a month, I test your website or phone assistant as six kinds of customer. I ask ordinary questions, record the exact words it uses and explain what needs changing. The first test is free. You get a clear record that you can give to whoever looks after the assistant.

1. Agree what a correct answer looks like

Start with your current service details and the questions customers normally ask. I set down what each answer must contain before testing. If your own information is unclear, that needs resolving before an answer can fairly pass or fail.

2. Ask as six kinds of customer

I act as a first time buyer, a customer with a complaint and someone in a hurry. I also act as someone who writes badly, someone asking about something your firm does not do and a returning customer who refers to an earlier answer.

The questions use everyday language. A spelling mistake or a short, rushed message is ordinary customer behaviour. Each kind of customer tests whether the answer remains useful when the question arrives in a different form.

3. Keep the exact answer

For each question, I record what a correct answer must contain and what the assistant actually says. A pass means it meets that written requirement. A catch means there is something to fix. The report includes the exact words, so your team can check the judgement.

4. Say what to change

Each catch comes with a specific change to consider. That might mean correcting a service detail, asking a missing question or sending an unclear request to a person. The report points to the answer that needs attention, rather than leaving you with a general score.

5. Repeat the check each month

The next monthly test checks earlier catches and ordinary questions against your current information. This helps distinguish an unresolved issue from a new one. A pass applies to the question asked at that time. It does not establish that every possible answer is right.

The blank test sheet

Write the expected answer before asking the question. Use one row per question and keep the assistant response exactly as given.

Kind of customerQuestion askedWhat a correct answer must containWhat the assistant saidPass or catchWhat to change
      
      
      
      

Download this sheet as a file for Excel

A made up report for Northfield Heating

Made up example. All rows are made up. Northfield Heating is an invented firm. Its service details and every assistant answer here are fictional examples of the method.

Kind of customerQuestion askedWhat a correct answer must containWhat the assistant saidPass or catchWhat to change
First time buyerCan I ask for a boiler service?Say a boiler service enquiry is accepted and ask for the postcode.Yes. What is the postcode for the property?Pass 
Customer with a complaintNobody has replied to my complaint. Who can help?Offer a route to a person without saying the complaint is resolved.Please choose the contact option to ask a person to review your complaint.Pass 
Someone in a hurryCan you promise an engineer today?Explain that a person must confirm availability.Yes, an engineer will come today.CatchRemove the promise and ask a person to confirm availability.
Someone who writes badlyboiler servce can u do itRecognise the request for a boiler service and ask for the postcode.Do you mean a boiler service? What is the property postcode?Pass 
Someone asking about something the firm does not doCan you repair my washing machine?Explain that washing machine repairs are outside the service.Yes, we repair washing machines.CatchCorrect the service information and state the limit.
Returning customer referring to an earlier answerYou asked for my postcode earlier. Why do you need it?Explain that the postcode is used to check the service area.It helps check whether your property is in the service area.Pass 
First time buyerCan I book before anyone checks my address?Explain that service area eligibility must be checked first.Your address needs checking before a booking can be confirmed.Pass 
Returning customer referring to an earlier answerEarlier you said someone must confirm the visit. Is it booked now?Keep the earlier limit and explain that confirmation is still needed.Yes, your visit is booked.CatchKeep the booking status consistent and require confirmation from a person.

Download this sheet as a file for Excel

Ordinary questions only

I never try to break, trick or attack an assistant. I use only ordinary customer questions. If a member of staff joins the chat, I stop at once. The test is about the answers the assistant gives, and ends when a person takes over.

When this is the wrong fit

An assistant that only shows fixed menu buttons is the wrong fit. There is no open question for it to answer. An internal tool with no customers also needs a different test based on the work its users do.

If you want a security test, that is a different job done under a written agreement. This monthly service checks normal customer conversations. It does not assess whether someone can gain access or misuse the system.

Where to start

Choose the assistant that customers already use. Begin with questions your team can answer clearly from current information. Use the blank sheet to agree the expected answers. The made up report shows how a useful catch connects an exact answer to a proposed change.

Common questions

Is the first test free?

Yes. The first test is free. It gives you a first report to review before deciding whether a monthly check fits your assistant.

Can you test a phone assistant?

Yes. The service covers website and phone assistants. The questions stay within ordinary customer enquiries, with the same need to record the exact answer and assess it against your information.

Does a catch mean the whole assistant is bad?

No. It identifies one answer that needs attention. Read the question, the expected answer and the exact response together before deciding what to change.

Can my team use the sheet itself?

Yes. Agree the expected answers first and keep the wording of each response. The blank sheet gives your team the same structure for its own checks between monthly tests.

Want this in your operation

I build and run production AI agents that take repetitive work off operational teams. Tell me what your team spends too long on.

shurco.aiGuidesFree toolsSolutionsInsightsWork with meRSSllms.txt