Most AI privacy assessments are not wrong. They are disconnected. A framework is selected, a questionnaire is circulated, a document is produced, and six months later someone pastes a customer list into a new browser extension that nobody has assessed.
The failure is rarely a missing policy. It is a missing link between the policy and the way staff actually use AI tools. This guide lists the symptoms that show an assessment has come loose from reality, then pairs each one with a cause and a fix you can implement without a large programme.
Symptom 1: nobody can list the AI tools in use
Cause: procurement records cover paid enterprise tools but not browser extensions, free tiers and tools embedded in existing software. Staff adopt them because they save time, and nobody asks.
Fix: build an AI register with one row per tool: purpose, business owner, data classes processed, vendor, contract owner, renewal date and approved use. Populate it from expense reports, single sign-on logs, browser extension inventories and a short anonymous survey. Review it quarterly, and add a rule that any new tool processing customer data needs a register entry before use.
Symptom 2: the assessment is a one-time questionnaire
Cause: the assessment is treated as a procurement gate rather than a control that has to survive configuration changes, model updates and new use cases.
Fix: tie each identified risk to a control, an owner and a test. Use the NIST AI Risk Management Framework functions, Govern, Map, Measure and Manage, as the organising structure, and record the residual risk the business has accepted. Re-assess when the use case, data source, vendor or model changes, not only at renewal.
Symptom 3: customer data appears in prompts
Cause: the tool is convenient, the prompt box is unrestricted, and nobody defined what may be pasted into it. Call-centre notes, order histories and identity documents are the usual offenders.
Fix: classify data first, then restrict the tool. Use a private tenant or enterprise deployment where the vendor contract excludes training on your data. Add redaction at the point of use: customer ID instead of name, masked phone number, no identity documents. Where the workflow needs real records, run it through an approved system with logging and access control rather than a public chat window.
Symptom 4: the vendor terms were never read
Cause: a free trial was approved by a team lead, and the terms were accepted with a click.
Fix: require five answers in writing before approval: where data is processed, whether it is used to train models, how long it is retained, how deletion is requested, and how breaches are notified. Add subprocessor disclosure and audit rights where your risk level justifies them. If the vendor cannot answer, that is the answer.
Symptom 5: retention is unknown and deletion is untested
Cause: retention settings live in a vendor dashboard that nobody opens, and the internal copy of prompts or outputs is kept forever because storage is cheap.
Fix: set retention explicitly for both the vendor and your own systems. Write down the deletion process, including backups and derived datasets. Test it once with a sample record and keep the evidence. Unstructured AI outputs stored in shared drives are the most commonly forgotten copy.
Symptom 6: access is broader than the use case
Cause: the pilot used a shared account, and the shared account survived into production.
Fix: named accounts, single sign-on where available, role-based access to the AI tool and to the data it retrieves. Log queries and administrative changes. Remove access when a project ends, and review the list quarterly. A shared login makes every other control unprovable.
Symptom 7: there is no incident path for AI
Cause: the incident response plan was written for servers and email, so a leaked prompt or a poisoned knowledge base has no owner or severity level.
Fix: add AI scenarios to the existing plan: sensitive data entered into a public tool, a model producing harmful advice, a retrieval source containing malware or misinformation, and a vendor breach affecting your prompts. Define who decides to disable the tool, who contacts customers, and what evidence is preserved.
Symptom 8: the data was called anonymous without a test
Cause: names were removed, but combinations of district, age, order value and date still identify people.
Fix: treat anonymisation as a testable claim. Attempt re-identification on a sample, document the method, and use aggregation or synthetic data where the workflow does not need individual records. If the data cannot be made non-identifying, assess it as personal data.
Board questions
- Which AI tools process customer data, and who owns each one?
- What is the worst plausible privacy incident, and what would it cost us?
- Which controls have been tested rather than documented?
- What data have we decided not to put into AI systems, and how is that enforced?
- When did we last remove access for a finished project?
Metrics
- Percentage of AI tools with a register entry and a named owner.
- Percentage of tools with a documented data processing position.
- Number of redaction tests passed in the last quarter.
- Median time to revoke access after a project closes.
- Number of AI scenarios covered in the incident response plan.
- Open high-risk findings and their age.
A note on local context
Bangladesh's data protection framework has been developing through draft and ordinance stages, and the commencement position should be confirmed with legal counsel rather than assumed. Whatever the current text says, the practical expectations are stable: tell customers what you collect, collect less, keep it only as long as needed, and be able to delete it. Cross-border processing is common because most AI vendors host abroad, so record where data goes and who is responsible for it.
Where this fits in our work
We help teams design AI workflows with data limits and human review built in through Business Automation. Engagement scope is described on the pricing page. For the policy side, read the AI governance policy guide.
Sources
- NIST, AI Risk Management Framework, accessed 17 November 2026.
- NIST, Privacy Framework, accessed 17 November 2026.
- OWASP, Top 10 for Large Language Model Applications, accessed 17 November 2026.
- DataGuidance, Bangladesh jurisdiction overview, accessed 17 November 2026.
- ISO/IEC 42001, AI management systems, accessed 17 November 2026.
Responses (0)
No comments yet. Add the first one below.
Leave a response
Your email address will not be published. Required fields are marked *.