Before putting customer work into an AI system, decide what information it needs, who may access the result and which actions require approval. OpenAI’s security controls matter, but your business also owns the application, the account configuration and the way people use it.
The useful question is specific: can this workflow handle this information, for this purpose, with these permissions? A blanket promise that “AI is secure” does not answer it. Start with one bounded task and make the boundaries visible before connecting live customer data.
Separate training from retention
OpenAI says API data is not used for training by default unless an organisation opts in. That does not mean nothing is stored. Default abuse-monitoring logs may be retained for up to 30 days, with exceptions; feature-specific application state has separate rules. Enhanced retention controls depend on eligibility and configuration.
Record which product and feature your system actually uses. Ask the account owner to verify the relevant settings. Do not assume that a statement about a personal ChatGPT account, a business workspace or one API endpoint applies to every other service.
Map the information before connecting it
Take one ordinary enquiry and follow it through the proposed workflow. Identify where the message starts, what the AI receives, where the result is saved and who gets a notification. Include the email service, customer database and any logging or monitoring system.
For an illustrative catering enquiry, drafting a request for a missing delivery time may only need the requested date, location and known order details. It does not need the customer’s payment credentials or every conversation the business has ever had. This is a proposed example, not a measured client result.
- Define the minimum information needed for the task.
- Use sample or appropriately redacted examples for initial testing.
- Name the person responsible for access and deletion requests.
- Record a retention purpose and review point for each saved record.
Give the workflow a clear limit
Reading an enquiry and drafting a reply are different from sending the reply, taking money or confirming a booking. Give the workflow only the access it needs. Keep an explicit approval step where a mistake could commit the business or affect a customer.
Test ordinary requests alongside missing details, contradictory instructions and failed connections. A customer’s message should be treated as information about their request, not permission to change the system’s rules. If the booking system is unavailable, the outcome should remain unresolved rather than appearing confirmed.
Keep the evidence for an external action: the responsible service’s confirmation, the record it belongs to and the person or rule that authorised it. An AI-generated sentence saying an action succeeded is not that evidence.
Know what the provider’s assurance covers
OpenAI publishes security information for its services, including its applicable independent assessments. Those assurances concern the provider’s stated scope. They do not certify your website, every supplier you connect or your team’s operating process.
Ask whoever is building your system to explain the remaining responsibilities in plain language: account access, secret management, approved data flows, incident reporting and what happens when someone leaves the team. If an answer depends on a configuration, ask to see that configuration verified.
How this applies at Klarnow
Our public Ask Klarnow service uses approved public company information and general business knowledge. It has no access to private client records and cannot book or send messages. It requests disabled response storage; that is not a promise that all provider logs are disabled.
Call is still being improved through early testing. Explore the pilot with sample information, then confirm live routing and activation with the team. Each business system needs an agreed scope for data, access and actions before it is put into customer use.
Start with the workflow that matters most. Our free diagnosis identifies the first priorities. A £250 Clarity Call can then map the system and agree what to build first; the full call fee is deducted from the agreed build fee when you go ahead with Klarnow.
Sources & further reading
Reviewed 13 November 2026 against OpenAI documentation and Klarnow’s current public website implementation. Provider policies and account settings can change. This guide explains practical operating decisions; it does not certify a system or replace a service-specific data review.