The independent record of what your AI actually told your customers.
Most online shops now have an AI assistant answering customer questions about returns, refunds, shipping and promotions. It was set up once, and in most cases nobody has looked at it since. It is talking to your customers, in your name, and you probably could not say what it tells them.
Companies write their own accounts, and investors still want someone independent to check them. Your vendor built the bot and your team runs it, so neither of them is the outside check. That is the job.
Free. No access to anything but your public chat widget. One working day.
Check your own, right now
You do not have to take my word for any of this. Here are three questions. Paste them into your own chat widget and see what comes back.
What is your return window on sale items?
If something arrives faulty, who pays the return shipping?
Can I use two discount codes on the same order?
Now open your own returns page and compare. If the answers match, that is genuinely good, and it is worth knowing rather than assuming.
If they do not match, you have just done the thing I do, on your own bot, in about two minutes. That was three questions out of the hundreds a customer might ask. I do not know what the rest return, and I suspect you do not either.
What this is
I ask your chatbot the questions your customers ask. What your return window is. Whether sale items can be sent back. Who pays return shipping when something arrives faulty. Then I compare every answer against your own published pages, and I report anything that does not match, quoted both ways with a link.
There are things I deliberately do not do. I do not attack your bot, try to trick it, or attempt to make it misbehave. I do not use anything except the public chat widget any customer can use. I hold no credentials, no keys and no access to your systems, and I do not want any.
Before a finding goes in a report it has to appear in at least three separate sessions, and be quotable word for word on both sides. Anything I cannot show you verbatim is left out. The list you get is shorter than it could be, and every line on it holds up.
A report with no findings is a complete deliverable. You are buying coverage and a dated record, not a guaranteed number of problems. If your bot comes back clean, I will tell you so and you will have that in writing.
Who I am
I am Shrikant. I have a degree in information technology and I work independently on how deployed AI systems behave once they are live.
I built VoxCart, a full ecommerce chatbot, as a system to test against. The code is public, so rather than take my word for what I can build, you can look at it.
I test from inside your market, so the answers I see are the answers your customers see. Every finding is your bot's words next to your own page, so none of it depends on what you make of me. I am one person, based in India, and I would rather you knew that now than found it out later.
What it costs
A first check
Free
Six to ten of the questions your customers ask, sent to your live chat widget, and a short document showing anything it said that contradicts your own published pages. Quoted both ways, with links.
The free check is how I find clients, and it is still true whether or not you hire me, because you can check every line of it against your own pages.
Risk spot-check
EUR 1,200 to 1,500
A bounded check of your three highest-risk areas: returns and refunds, shipping, and promotions. Findings only, cited to your own pages.
Chatbot health check
Most people start here
EUR 2,500 to 3,000
The full check across every published policy area, plus how the bot holds up when a customer pushes or asks awkwardly. Findings ordered by what they actually cost you, the exact change needed for each, and a re-check after you make them.
Done for you
Health check plus EUR 2,000 to 3,000
The health check, plus I write the exact change for each finding and verify it after your team applies it. I never touch your live system; you keep the keyboard.
Ongoing
EUR 600 to 900 per quarter
A quarterly re-check with a short report on what changed. Policies drift and bots go out of sync; this catches it.
Final scope and price are confirmed on a short call. Payment is half upfront, or in full upfront for the first two engagements, which is normal for a small independent practice. All work runs under a written agreement.
How I work, and who I do not work for
Crestseal has no commercial, financial or referral relationship with any chatbot platform vendor, agency or implementation partner. Crestseal does not resell, license or receive commission on any product or tool. Crestseal does not implement the systems it tests. Testing is performed against sources the client designates, and findings are reported regardless of which party is responsible for them.
- One person. No employees, no contractors, no sub-processors handling your data.
- No client data in any cloud service. Local encrypted storage, sole access.
- No credentials, keys or system access requested or held.
- Testing uses invented customer details, never real ones.
- Raw records deleted ninety days after an engagement ends, with written confirmation.
- Liability capped at fees paid.
Ask me to check yours
Send me your chat widget URL and I will send back anything it says that contradicts your own pages. Free, one working day, no obligation either way.
- If the form does not work for you, email me directly at shrikant@crestseal.com.
- I send what I find to you and to nobody else. I will not publish it or refer to it anywhere without your written permission. If you do not reply, I delete everything within thirty days.
- A free check is six to ten sessions. I never ask to be passed to a human agent, and I give you the exact timestamps so your team can exclude them from your reporting.
- I usually reply within one working day.
- I test only the public chat widget. Nothing else is touched.
- I will tell you if I find nothing.