Enterprise SaaS assistant with governance controls

Architecture and data flow

Organisation size

50–20000 employees

Data classes

confidential, personal

Constraints

data may leave the company under contract; an identity provider already in place (Entra ID, Google Workspace, Okta); someone who can read a DPA and own the vendor relationship

Industries

Professional services, Technology, Financial services, Retail, Other

Jurisdictions

any

No hardware profile was sized for this answer.

API integrationapi-integration

development

Change managementchange-management

operations

Compliance and governancecompliance-governance

compliance

Identity and SSOidentity-sso

security

Security hardeningsecurity-hardening

security

Commands are copied from each project’s own documentation, and the page they came from is linked under the step. 0 of 7 steps currently open an evidence record. The rest are linked to their source; §10 says which of those documents were fetched and which were fetched without their anchor being found — two different states, named differently there.

1. 01

   Shortlist on terms and connectors, not on demos

   Assessment

   Write down the three or four things this has to do and the data it will touch, then ask each vendor the same questions in writing. Vendors publish enterprise privacy pages — OpenAI’s says its commitments give you "ownership and control over your business data" across its business products — but a published page is a starting point for the contract, not a substitute for it, and the wording on these pages changes.

   Source documentation

2. 02

   Get the DPA, the subprocessor list and the retention terms in writing

   Assessment

   Ask for: the data processing agreement; the current subprocessor list and how you are notified of changes; where processing happens; the retention period and the deletion path; and a written no-training commitment covering uploads and connected sources. Note the boundaries the vendor documents: Microsoft states that for EU users "EU traffic stays within the EU Data Boundary while worldwide traffic can be sent to the EU and other countries or regions for LLM processing".

   Source documentation

3. 03

   Connect SSO and automatic provisioning first

   Assessment

   Configure SSO before the pilot and SCIM provisioning before the rollout. Decide which groups get access, and make membership of those groups the only way in — a tool with its own user list becomes a shadow directory within a month.

4. 04

   Fix the source permissions before connecting anything

   Assessment

   Connectors respect the permissions they find. Run an over-sharing review on the systems you intend to connect — the shared drive everyone can read, the site with inherited permissions — because retrieval will surface what browsing never did. Vendors sell controls for this (Glean markets "enforced data permissions"), but the underlying permissions are yours to correct.

   Source documentation

5. 05

   Write the acceptable-use policy before the licences arrive

   Assessment

   One page: what may be put in, what may not, that output is a draft to be checked, that usage is logged, and who to ask. Publish it with the licences, not after the first incident. Name the sanctioned tool explicitly — most shadow AI use is people trying to do their job with the only tool they have.

6. 06

   Pilot with one department and measure something

   Assessment

   Thirty people, six weeks, two or three named tasks, and a before-and-after measure you agreed in advance. Collect the answers that were wrong — they tell you which content is stale or over-shared, which is worth more than the satisfaction survey.

7. 07

   Put the vendor on a review cycle

   Assessment

   Diary a quarterly check of the subprocessor list, the retention terms and the admin audit export, and re-read the terms at renewal. These pages change without notice; that is precisely why they are worth watching.

   Source documentation

Structured issue-spotting to support your own review — not legal advice. Verify against the cited primary sources and your counsel.

• Self-hosted, if the answer to "may documents leave?" is no

  When confidentiality or professional secrecy makes a processor relationship unattractive, running the stack yourself removes the question rather than governing it.

  • — No vendor in the data path
  • — You take on the operational work and the hardware
  • — Fewer connectors and a less polished product

  open-webuivllmpgvector

• Private cloud, as the middle option

  One tenancy, one region, open weights: a processor relationship with an infrastructure provider rather than an application vendor.

  • — Region and retention under your control
  • — You build the application rather than buying it

  vllmopen-webui

• Private cloudPrivate-cloud RAG in a single regionA GPU instance in one region running vLLM and Open WebUI, a managed PostgreSQL with pgvector for chats and embeddings, object storage for the original files, and your existing identity provider for sign-in. Same software as the on-premise recipe; different trust boundary.
• Self-hostedPrivate company knowledge base (self-hosted RAG)Open WebUI as the employee interface, vLLM serving a Qwen2.5-14B-Instruct model on a single 24 GB GPU, PostgreSQL with pgvector for chats and embeddings, and OIDC single sign-on — all in Docker on one server in your office or colocation rack. Ollama replaces vLLM for teams under about 20 users; a 48 GB GPU lets you run a 32B model for better answers.

Not yet in the ledger

1. OpenAI — Enterprise privacy

   Vendor page

   not yet fetched

2. Microsoft — Data, privacy, and security for Microsoft 365 Copilot

   Vendor page

   not yet fetched

3. Glean — Security and data protection

   Vendor page

   not yet fetched

Deployed this stack, or hit something this page does not cover? Corrections, sources and implementation reports are what keep a recipe worth reading.

If you would rather not build it, we can introduce a forward-deployed engineer who has deployed this stack before. The enquiry form starts from this recipe.