Chatbot Privacy and What Your Bot Should Not Store
I read fifty transcripts from a small physiotherapy clinic's website bot last spring, as a favour.

I read fifty transcripts from a small physiotherapy clinic's website bot last spring, as a favour. Seven of them contained something the clinic had no business holding: two full dates of birth, one description of a back injury detailed enough to identify the person, a mobile number typed twice because the first message had a typo, and one sixteen digit card number that a visitor had pasted in because the bot asked "how can I help with your payment?" All of it sat in a vendor dashboard with no retention setting switched on, going back fourteen months.
Nobody at the clinic had made a decision to collect any of that. They had installed a widget, accepted the defaults, and forgotten about it. That is how almost every chatbot privacy problem starts, and it is why a privacy policy page is the least useful part of the fix. The useful part is deciding what your bot writes down, and then checking that it is writing down only that.
Why the transcript fills up with things you never asked for
Visitors treat a chat box like a person. Someone who would never put their medical history into a web form will type it into a chat window, because the format implies a conversation with a human who needs the context. The clinic bot's opening line was friendly and open ended, and open ended invitations get open ended answers.
The second cause is the question itself. A bot that asks "what is your order number?" gets an order number. A bot that asks "tell me a bit about your problem" gets everything. The wording of a prompt determines the shape of the data you end up storing, which makes prompt copy a privacy decision rather than a marketing one.
Third, defaults. Most widget platforms log the full conversation, the visitor's IP address, the pages they visited before opening the chat, their browser fingerprint, and often a geolocation estimate. All of that is on because it helps the vendor's analytics, and none of it has a retention limit until you set one. Several platforms also email a copy of every transcript to the account owner, which quietly duplicates the whole archive into a mailbox that will never be cleaned.
A field by field view of what to keep
This is the table I now hand to anyone setting up a bot. The retention column assumes a small business with no regulated obligation to keep records longer.
| What the bot sees | Keep it? | Retention | Reason |
|---|---|---|---|
| Message text, redacted | Yes | 60 to 90 days | You cannot improve answers without reading real questions |
| Intent, tags, outcome | Yes | 13 months | Aggregated counts, no personal content |
| Email or phone, given on purpose | Yes, in your CRM | Normal customer rules | The visitor chose to hand it over for a reply |
| IP address | No | Not stored | Fraud screening is the only honest use, and you are not doing it |
| Card numbers, passwords, ID numbers | Never | Redact before write | You have no safe place to put these |
| Health, financial or household detail | Only if you are the service | Shortest possible | The category that turns a leak into a serious incident |
The redaction job is less work than it sounds. A pattern that catches runs of thirteen to nineteen digits, with or without spaces and hyphens, replaces almost every card number anyone will ever paste. A second pattern for long numeric strings beginning with a national identifier format covers most ID numbers in your country. Run both before the message is written to storage, not after, and keep the replacement visible so you can see it happened: a line reading "card number removed" tells you the pattern is working.
The advice to just add a privacy policy link is backwards
Every setup guide I have read puts the same item at the top of the list: link your privacy policy in the chat window and add a consent checkbox. Do that, certainly. It takes four minutes. But treating it as the main protection has the logic inverted, and I think it does genuine harm by making people feel finished.
Consent is not a container. A notice telling visitors you process their messages does not reduce what you are holding, does not shorten how long you hold it, and does not help at all on the day someone gains access to the dashboard. Meanwhile the clinic's card number was sitting in plain text behind a password that three staff members shared, and no policy wording changes that fact.
There is a second version of the same mistake, which is to overcorrect and switch logging off entirely. I have seen two businesses do this after a scare, and both ended up with bots that slowly got worse, because the only reliable way to find out why a bot is failing is to read what people actually typed. The metrics that matter are built from those transcripts, as the piece on chatbot analytics and the metrics worth tracking sets out in more detail. Redaction plus a short retention window gives you the diagnosis without the liability. Blanket deletion gives you neither.
A fifteen minute audit you can run this week
Open your widget dashboard and work through five things in order. The whole exercise took me eleven minutes on the clinic account and removed roughly 90 percent of the exposure.
- Find the retention setting and set it. Sixty days for message content is enough for almost everyone. If the platform has no retention control, that is a reason to change platform.
- Turn off IP and geolocation storage. Then check whether your analytics integration is capturing it separately, because two of the three platforms I have audited were doing exactly that.
- Switch off transcript emails. Replace them with a daily digest of counts, or nothing at all.
- Read the last fifty conversations and count how many hold something sensitive. That number tells you whether your opening prompt is too open ended.
- Check where the vendor stores data and who their sub-processors are. If you are in Europe, look for a region setting, because several platforms default to storage outside the area and then mention it only in a sub-processor appendix.
Then fix the prompt. Changing the clinic's opener from an open invitation to "which of these do you need?" with four buttons cut sensitive detail in transcripts from seven in fifty to one in fifty across the following month. That is a single line of copy doing more than any setting.
One last thing worth planning for. When a conversation genuinely needs a person, the handoff is the moment sensitive detail appears, because the visitor repeats their whole situation to the human. Decide in advance what the handoff summary carries and what it leaves behind, which is the part most setups skip in the chatbot handoff to a human flow. There are more build notes of this kind in the chatbots section.
FAQ
How long should I keep chat transcripts?
Sixty to ninety days for the message content, which covers a full cycle of reviewing and fixing answers. Keep the aggregated numbers, such as intent counts and resolution rates, for a year or more. Those carry no personal content and are what you need for a year on year comparison.
Do I need a consent checkbox before the chat starts?
A short notice, yes. A blocking checkbox usually costs you more conversations than it gains you in protection. One sentence above the input box, naming what you store and for how long, is clearer and does not add a click.
What if a visitor asks me to delete their conversation?
You need to be able to find it and remove it, which means knowing how your platform searches by email or session. Try it once before anyone asks. On two platforms I have used, deletion from the inbox view leaves the record in the analytics export.
Is a self hosted bot safer?
It removes the vendor from the picture and hands you every other responsibility, including patching, access control and backups. For a small team without someone who enjoys that work, a reputable hosted platform with retention configured is the more realistic choice.
More in Chatbots
Chatbots
Chatbot Buttons vs Free Text and When to Use Each
A friend who runs a two person accountancy practice replaced her old button menu with a shiny new text box last spring, because...
Order Tracking Chatbot for a Small Online Shop
If you run a small online shop, a large share of your support inbox is probably one question asked fifty ways: where is my...
Chatbots
Chatbot Analytics and the Metrics Worth Tracking Each Week
Your website chatbot has been live for a month, the dashboard says it handled 1,400 conversations, and you still cannot tell...