AI log retention: what to keep and explain to customers
Decide separately about payloads and usage, then tell customers how long records last and what they can establish.
Key points
- Decide on payloads and usage separately
- Fit inquiry deadlines within retention
- Explain scope for each log destination
Overview diagram in preparation
Contents
Decide separately about payloads and usage
A&A perspective
AI logs record exchanges with AI and information about its use. Decide separately whether to keep the text of customer documents and whether to retain usage information. Without the text, there are limits to what you can check later.
Hypothetical example
Suppose a bookkeeping company uses AI to extract dates and amounts from customers' invoices. A customer asks whether the documents remain in the company's systems. This is a hypothetical example, not a result achieved by an A&A customer.
From the sources
Cloudflare's AI Gateway, which intermediates calls to AI, has a setting that stops payload storage alone. It can record usage, model, cost and duration without storing request or response text.
From the sources
The setting to give your operator is cf-aig-collect-log-payload: false. In contrast, cf-aig-collect-log: false stops the entire log, including usage information. The latter takes precedence over the payload-only setting.
A&A perspective
Automate & Augment (A&A), the organization writing this article, recommends separating investigation purposes first. Finding which call incurred a cost requires different records from finding why an invoice amount was misread.
A&A perspective
Usage records let you examine the cost or duration of processing where records remain. They do not let you reread the original input or the incorrect response. Even a retained source invoice may not match the text actually passed to the AI.
A&A perspective
If reviewing incorrect responses later is important, keeping the necessary text in an approved location is another option. Specify the fields, who may access them and when they must be deleted. A policy of retaining no payloads is not right for every situation.
A&A perspective
This is a proposed application of Cloudflare's distinction to your own operations. The AI interface or integration you normally use may not offer the same controls. Ask your contractor to test the no-payload setting and verify that usage information still remains.
Accept inquiries while records still exist
A&A perspective
Set the inquiry deadline early enough for the operator to examine the necessary records before they disappear. Accepting inquiries until the final day of retention can leave no time to investigate after a customer contacts you.
From the sources
Cloudflare applies Workers Logs pricing and retention rules to new customers who created their first gateway on or after September 24, 2026. Workers Logs is its log storage system; customers who created a gateway earlier use the legacy logging system.
From the sources
As of November 1, 2026, Workers Logs retains records for three days on the Free plan and seven days on the Paid plan. The same page announces a move to a different pricing model from December 1, 2026.
A&A perspective
Three or seven days therefore cannot be presented as a retention period shared by all AI services.
A&A perspective
Check which logging system applies to your business and when its records actually expire before telling customers a period. Long-term commitments also need review when specifications change.
Hypothetical example
Our hypothetical bookkeeping company cannot rely solely on these seven-day records to investigate the previous month's work at month-end: the older records will be gone. If monthly investigations are necessary, keep the required fields in another approved location or change when inquiries are accepted.
From the sources
Retention does not guarantee complete text. Workers Logs says records exceeding 256 KB are truncated. After the daily limit of five billion logs is reached, it samples 1% of logs for the rest of that day.
A&A perspective
We do not expect five billion logs to be a routine issue for a small company. Even so, retention duration and completeness are different conditions. Check whether text has been truncated or settings record only a sample when explaining investigation capabilities to customers.
A&A perspective
Adding storage destinations increases storage costs and the work of managing access and deletion.
A&A perspective
Compare keeping usage alone, storing necessary text elsewhere and accepting inquiries earlier against the investigations you actually need. This article does not measure any cost savings.
| Storage choice | Records retained | What can be checked later |
|---|---|---|
| Keep payloads at the gateway | Requests, responses and usage | Stored text and usage; check gaps and expiry separately |
| Store no payloads at the gateway | Usage, model and other metadata | Usage; original text cannot be read from these records |
| Stop all gateway logging | Nothing in this gateway log | No investigation from this log; check external storage separately |
Explain investigation scope for each destination
A&A perspective
Tell customers what you retain and what you can investigate for each storage destination. A statement that you do not store AI payloads leaves unclear which location's records you mean.
From the sources
The AI Gateway documentation states that these logging headers control records stored by AI Gateway. They do not change the output of Logpush, which exports logs to another destination.
A&A perspective
Checking this setting does not establish that the AI provider or your application also stores no text. Ask the operator to check gateway records, exports and your own storage separately. Mark destinations whose behavior is unknown as awaiting confirmation.
A&A perspective
In a retention policy table, list each destination, whether it holds payloads and usage, its retention deadline, and what cannot be investigated. The purpose is to make the customer explanation cover the same scope as the operator's settings, rather than to list every technical field.
Hypothetical example
A hypothetical explanation could read: 'Our AI gateway logs retain usage information without storing document text or responses. We can examine usage in the remaining records, but cannot reread the original text there to investigate an incorrect response.' Add verified conditions for other destinations and the inquiry deadline.
A&A perspective
Also distinguish keeping original invoices for normal business purposes from retaining AI gateway logs. Enabling a no-payload setting is not evidence that original documents or past logs have been deleted. If older data must be removed, verify its scope and the deletion result separately.
A&A perspective
Ask the operator for a test input containing no personal information and the resulting records, as well as the settings screen. Use them to check that payloads are absent while usage remains, and whether the explanation also holds for external destinations. This is A&A's proposed verification method.
A&A perspective
Both sources are the provider's product documentation; neither evaluates the effectiveness of this policy table. Determining statutory retention duties or customer contract requirements is outside this article's scope. The table identifies what current settings support and which destinations still need confirmation.
Align retained records with investigation limits
A&A perspective
Decide separately whether to retain AI payloads and usage according to the investigations you need.
A&A perspective
Set the inquiry deadline with time for the operator to investigate, as well as the retention deadline itself.
A&A perspective
Verify what remains at each destination and explain to customers what can and cannot be established.
Decide separately whether to retain AI payloads and usage according to the investigations you need. Set the inquiry deadline with time for the operator to investigate, as well as the retention deadline itself. Verify what remains at each destination and explain to customers what can and cannot be established.
FAQ
Does disabling payload storage delete older logs too?
Do not assume older records have been deleted. The described headers control storage for individual calls. Verify deletion of older records separately, and distinguish future retention policy from the handling of existing data when explaining it to customers.
Can usage-only logs fully reconcile billed charges?
Not necessarily. In addition to retention limits and missing records, you need to check which processing corresponds to each cost. Having usage records does not replace reconciliation against the actual invoice.
Does retaining no payloads also prevent AI training use?
That conclusion does not follow. These controls govern storage of gateway records. Storage and training use by the AI provider require separate checks of the service and contract terms you use.
Sources & editorial note
Primary pages read for this article. Publication dates below belong to the sources; access dates record our research.
- Logging · Cloudflare AI Gateway docs
Cloudflare · Publication date not stated; updated 2026-10-26
Accessed 2026-11-01 - Workers Logs · Cloudflare Workers docs
Cloudflare · Publication date not stated; updated 2026-10-12
Accessed 2026-11-01