guide · Evidence checked 2026-09-02
Govern the draft before the prompt
A practical, evidence-led decision guide. Product capabilities and limits are separated from anything that would require hands-on testing.
An AI writing trial often begins with the most sensitive material: an unreleased campaign, customer interview or internal strategy. That is backwards. The safe first step is to classify the content and establish what the product and its model providers may retain, reuse and expose.
The six questions that change approval
| Control | Evidence to obtain | Failure condition |
|---|---|---|
| Training use | Plan-specific statement and contract | Opt-out is assumed from a marketing headline |
| Retention | Input, output, logs and backup periods | “Not used for training” is treated as “not stored” |
| Model routing | Current subprocessors and regions | Sensitive text can reach an unapproved provider |
| Access | Roles, SSO, sharing and audit events | A public link or departed user retains access |
| Deletion | User, admin and contractual deletion path | No testable removal workflow exists |
| Source truth | Owner, approved corpus and review rule | Fluent output can introduce unsupported facts |
Start with synthetic documents. Then use one low-sensitivity real workflow and test access removal, export and deletion. Capture evidence for the exact plan: DeepL Write distinguishes free and Pro handling; Typeface is a contract-led enterprise decision; Lex and HyperWrite require their own review. One vendor's promise never transfers to another.
Procurement passes only when the team can name the data owner, allowed content classes, model route, retention period, deletion evidence and accountable human approver. A lower subscription price cannot compensate for an unbounded data path.
Begin with the document, not the vendor badge
Consider a hypothetical marketing team preparing a public product announcement from an approved fact sheet. The public facts, an unreleased price change and a named customer interview should not automatically follow the same route. Identify which material may be used, who owns it and which questions need the organization's responsible reviewer before any upload.
This is a procurement and workflow method, not a legal opinion or a certification of a product. A trial with synthetic text can establish whether the process is workable; it does not authorize using restricted business information in the production account.
A reusable one-page data-flow record
Copy this table into the team's approval record and fill every row for the exact product and plan. Leave unresolved fields visible rather than replacing them with a generic privacy link.
| Field | Example for an original hypothetical scenario | Evidence to retain |
|---|---|---|
| Decision job | Draft a public product announcement | Approved brief and permitted claim sheet |
| Material allowed | Public product facts and synthetic examples | Named owner and content classification |
| Material excluded | Private customer interview and unreleased price | Explicit exclusion in the brief |
| Product and plan | Exact account configuration, still to confirm | Order form, plan and relevant settings |
| Processing route | Vendor and model providers, still to establish | Applicable policy, subprocessors and contract |
| Retention and deletion | Separate answers for prompts, files, output and logs | Periods, exceptions and deletion process |
| Access and exit | Writer, reviewer, departing-user process | Permission test and export sample |
| Publication approval | A named human checks every factual claim | Final approved version and review record |
The example deliberately leaves vendor-dependent fields incomplete. That is the point of the record: a workflow should not pass because its unknowns were hidden in a checklist tick.
Training, storage and access are different questions
A statement that content is not used to train a model does not establish that it is never stored, never logged or never accessible to the service provider. Ask separately about file processing, chat history, generated output, diagnostic logs, backups and any third-party model route. Obtain the answer for the actual product and plan, not another product sharing the same brand.
The Notion pricing page captured on 19 October distinguishes LLM-provider retention for Enterprise from other plans; that is a narrower statement than “the whole workspace retains nothing.” Its AI FAQ also exposes settings controlling data sharing and models. Treat those as plan-and-setting questions, not a universal privacy guarantee.
Lex's pricing FAQ discusses its use of enterprise model APIs and states a training position, while its privacy policy discusses retention for several purposes without giving one universal duration for all writing content. The same pricing page says new customers are no longer accepted. Those documents do not support a new-customer rollout or an invented fixed deletion period.
Test a departure, not just a successful draft
In an authorized synthetic workspace, give a writer and a reviewer deliberately different access. Create a draft referencing an allowed source and an inaccessible source. Confirm what the writer can actually retrieve, then remove the synthetic user's access and check the normal sharing routes. Record the observed outcome; a vendor's promise and your test result are different evidence.
Next export a representative draft with headings, tables, links and reviewer comments. Identify which parts remain portable and which must be archived separately. Ask how deletion requests interact with active documents, generated files and retained logs. No access-removal, deletion or retention test was performed by BenPicks for this guide.
Make approval proportionate to the work
For a hypothetical pilot of six writers, assume one 25-minute briefing each and one 30-minute owner review of the workflow. That totals 180 minutes, or three hours. At an invented $30 hourly rate, the preparation cost is $90 before subscription or procurement time. A repeated ten-minute check on 12 monthly assets adds 120 minutes, or $60 under the same assumption.
These figures are original planning arithmetic, not a recommended legal-review budget or a claim about a vendor's costs. They help distinguish a one-time setup from recurring editorial work. Reuse the approved record for unchanged workflows, and review changes when the plan, data class, model route or sharing pattern materially changes.
For tool selection, use Notion AI vs Lex with its current availability caveat, or the marketing-team writing shortlist for other operating models. A purchase should solve a documented job while keeping unresolved data questions visible.
Primary action: build a one-page data-flow record before enabling real documents, then attach the current privacy notice, DPA and order-form exceptions to the approval.
Primary sources for verification
- OpenAI business data controls ↗
- DeepL privacy policy ↗
- Typeface security and governance ↗
- ProWritingAid privacy policy ↗
- Lex privacy policy ↗
- Notion AI FAQ ↗, Notion pricing ↗ and Lex pricing ↗ — captured 19 October 2026 for the scoped examples above.
Lex privacy and DeepL privacy were also recaptured on 19 October. The local OpenAI business-data request returned HTTP 403; that link remains a source for the buyer to verify, not a freshly confirmed policy in this update. Other retained sources are not all newly checked.