Article Details
AI Features That Touch Customer Data: The Compliance Baseline
The fastest way to kill an otherwise good AI feature is to build it first and ask about data handling second. Enterprise buyers will ask, and their security questionnaire is not negotiable. Retrofitting these controls is expensive; building them alongside the feature adds days.
The seven controls
- A written data flow. Which fields leave your infrastructure, to which vendor, in which region, and what comes back.
- Retention terms. Confirm in writing that your provider does not train on your data and how long prompts are retained. Configure zero-retention endpoints where they exist.
- Tenant isolation. Retrieval must be scoped by customer at the query level, not filtered after the fact. Cross-tenant leakage through a shared vector index is the most common serious defect we find.
- Audit logging. Every AI-assisted decision that affects a customer needs a record: input, output, model version, timestamp, and who acted on it.
- PII minimisation. Redact or tokenise what the model does not need. The cheapest compliance story is the one where sensitive fields never left.
- Human review where it matters. Decisions with legal or financial consequence need a documented person in the loop, not just a confidence score.
- Disclosure. Your privacy policy, DPA and subprocessor list must name the AI vendors before a customer discovers them in a network trace.
Why this sells rather than slows
Every one of these controls turns into an answer on a procurement questionnaire. Teams that have them close enterprise deals in weeks; teams that do not spend a quarter negotiating exceptions or lose the deal to someone who prepared. The baseline is not overhead — it is the thing that lets you sell upmarket.
You will be asked where the data goes. Having the diagram ready is worth more than any feature you could have built in the same three days.
If an AI feature is already live and none of this exists, treat it as a two-week remediation rather than a rebuild. In most systems the controls are additive, and the isolation fix is the only one that usually touches architecture.
Our News
Choosing Your First AI Workflow: A Scorecard
Six criteria, scored one to five. Anything under twenty is a second project, not a first one.
The Two-Week Paid Discovery That De-Risks a Six-Figure Build
Free proposals are priced to win work, not to be accurate. A paid discovery is the cheapest insurance a founder can buy.
Fixed Price, Time & Materials or Dedicated Team: Which Protects You
Every contract model moves risk somewhere. Here is where each one puts it, and the clauses that matter more than the model itself.
Red Flags in a Development Proposal: A Line-by-Line Teardown
Nine phrases that appear in ordinary-looking proposals, what each one is actually protecting, and the question that exposes it.