What a CI platform touches in your stack: CRM, call recordings, data and security FAQ

What the OAuth grant lets a CI tool do

A CI tool reads your CRM through an OAuth app. Its reach is set by the scopes the app requests, and the consent screen is the first document to read. Ask for the exact scope list, which of them are write scopes, whether the app receives a refresh token so it can work while nobody is logged in, and which objects and fields it writes to. Vendors rarely publish the strings, so ask in the review.

HubSpot's OAuth documentation (read October 7, 2026) states that each scope opens a set of API endpoints, and an app's access token does not reflect the permissions or limitations of the user who connected it. An app approved by an admin can therefore write to records that the connecting user's own role could not edit. HubSpot also drops optional scopes automatically when the account lacks the tool they cover, so the list the vendor requested and the list you granted can differ. Salesforce's developer pages could not be opened for this article, so ask the vendor the same questions there.

Gong is different again. Gong's own app documentation says it does not support user-level OAuth; authentication happens once, globally, and the setup must be done by a tech admin. Whoever approves the connection is granting access to every call in the instance that the requested scopes cover. The page shows scopes such as api:calls:read:basic and api:calls:create without enumerating all of them, so ask which call and transcript scopes the CI tool requests and whether any of them can create or change records.

Slack is the easiest to read, because Slack publishes every scope. channels:history lets an app view messages in public channels it has been added to; chat:write lets it post; users:read.email lets it read email addresses across the workspace. Crayon's Slack Marketplace listing, for example, declares that it can view content and information about channels, conversations and the workspace. A tool that only needs one channel should only be in one channel.

What is copied out, and for how long

Separate three retention periods: your data while the contract runs, prompts and completions at the LLM provider, and what remains after termination. Most vendors state the third and skip the first two.

Klue's terms (version dated October 17, 2025) commit to destroying, deleting or de-identifying personal data on termination; a retention period during the contract is not stated. Crayon's Slack listing says it retains customer data for the duration of the engagement, deletes it on termination or request, and stores LLM prompts and completions for 30 days for abuse monitoring. For Kompyte we found no retention statement on Semrush's security page. Ask each vendor to put all three periods in the data processing agreement.

Training, LLM providers and sub-processors

Read the training sentence to the end, because it can carry a carve-out. Klue's terms say customer data will not be used to train, validate or enhance AI features except solely for the benefit of the customer, and never to train publicly available AI systems or public large language models. Crayon's listing says it uses Microsoft Azure OpenAI, that the data is tenanted to Crayon, and that Crayon's data is not used for model training. Semrush's security page, which covers Kompyte, says nothing about AI training.

Then ask for the sub-processor list as a URL and the notice period for changes. Klue's terms point to a published list and commit to incident notice within 72 hours. Semrush commits to data processing addenda and to making a sub-processor list available to customers. The LLM provider should appear on that list by name, with its retention term beside it.

Why "SOC 2 compliant" is not an answer

SOC 2 is an attestation: an auditor reports on controls over a stated period. The useful questions are which type (Type I is a point in time, Type II covers a period), which period, which Trust Services Criteria, and whether you can read the report.

Beyond the table, Klue's security page adds AES-256 encryption at rest and annual third-party penetration tests, with no report period and no mention of ISO. Crayon's only public statement is the news post of February 24, 2022; the notice is dated 2022, so ask for the current period. Kompyte sits under Semrush, whose security page covers PCI DSS, GDPR and data processing addenda and names neither SOC 2 nor ISO 27001. "Not public" in the table means we found nothing; it is not a No.

DPA, transfers and Japan

For a European or UK controller the questions are familiar: a data processing agreement that names the vendor as processor, the transfer mechanism, and the incident notice clock. Klue's terms appoint Klue as processor and cite the June 2021 Standard Contractual Clauses for its sub-processors. Semrush states GDPR adherence and data processing addenda. Check each vendor's own clause rather than the summary on its security page.

Japan needs a separate line in the review. The Act on the Protection of Personal Information has rules on providing personal data to a third party in a foreign country, and the Personal Information Protection Commission publishes guidelines and Q&A on them, including on foreign cloud services. Whether they apply to a given vendor depends on the facts and the contract, so take the question to your legal team rather than settling it from a vendor page. Only the Japanese text of the law is binding, and this is not legal advice. The practical questions for a Japanese buyer are whether the vendor handles the personal data in the transcripts and CRM records it ingests, where its servers are, and which mechanism the vendor relies on.

The reviewer's worksheet

Send these to the vendor as written questions and file the answers with the contract. What was said in the demo cannot be filed.

• The exact OAuth scope list per integration, marked read or write, with the objects and fields written to.
• Whose identity each integration uses: a named user, a dedicated integration user, or a global app approved by an admin.
• Retention during the contract, at the LLM provider, and after termination, each as a number of days.
• The training sentence in full, including any carve-out, and the named LLM provider with its retention term.
• The sub-processor list as a URL, and the notice period for additions.
• SOC 2 type, period covered, Trust Services Criteria, and when the report and bridge letter will arrive under NDA.
• The DPA, the transfer mechanism, and the personal-data incident notice clock in hours.
• For Japan: where the servers are, whether the vendor handles the personal data, and which APPI mechanism applies.
• Offboarding: export window, deletion date, written confirmation, and the destruction standard.

How Steve answers these, from its SLA

Steve's SLA page (read October 7, 2026) covers some of these questions. It states that Trissino does not train AI models, and it lists sub-processors, including AI providers, so ask for the named list and the retention terms that apply to each. It states a SOC 2 Type 2 attestation for 1 October to 31 December 2025; ask how and when the report is released.

On scopes, Steve's Slack page says Steve can see messages shared only on the channel it is added to, and by default the person who posted an insight approves it before it enters Steve's knowledge base. The Gong page says Steve imports the call transcripts Gong makes available and scans each one for competitor mentions, with review before a battlecard changes.

The site does not state several things, so ask for them in the review: the exact OAuth scope strings for Salesforce, HubSpot and Gong; which identity each integration runs under; SSO, SCIM, ISO 27001 or data residency; and any terms specific to customers in Japan beyond SLA section 11.5. If those four matter to your reviewer, the integrations page lists what connects, and the SLA is where the written answers will have to land.