← All news and guides
AI governance

Enterprise Readiness for Your First Big Customer: SSO, Audit Logs and AI Controls

What a first enterprise customer expects from your SaaS: SSO and SCIM, roles, audit logs and the tenant-level AI controls admins ask for, with proof to attach.

Mono Team ·

An empty glass-walled office lobby in early morning light with a reception desk and a row of turnstile gates.

Short answer: for a first big customer, expect to ship five things: single sign-on, automatic user provisioning, role-based access, audit logs the customer can export, and, if you sell AI features, tenant-level AI controls that an admin can set. None of these is a legal requirement. They are what an enterprise security team checks before it lets a vendor near its staff and data. Build them in the order your deal needs them, make each one enforce on the server, and keep a screenshot, a config or a test as proof for each.

This guide covers each control, what to log, which AI settings buyers ask for, and where contracts and frameworks fit. It is an engineering view, not legal advice. If your deal is already waiting on a review, read why enterprise deals stall in AI security review alongside this.

What do enterprise buyers expect from a SaaS vendor before they sign?

Requirements differ, but the same six areas come up again and again. Each maps to something you can build and show.

AreaWhat the buyer's security team asksWhat you buildProof to hand over
IdentityWhat the buyer's security team asks: Can our staff sign in with our identity provider? Is access removed when someone leaves?What you build: SAML or OpenID Connect SSO per tenant; SCIM provisioning and deprovisioningProof to hand over: SSO setup guide; a test run showing a deactivated user loses access
AccessWhat the buyer's security team asks: Who can see and change what? Can we limit admins?What you build: Role-based access with least privilege; scoped API keysProof to hand over: Role and permission matrix; API key settings screen
Audit logsWhat the buyer's security team asks: Can we see who did what, and export it?What you build: Tenant-scoped, append-only event log with exportProof to hand over: Sample log entry; export screenshot; retention setting
Data controlsWhat the buyer's security team asks: Where does our data live, how long, and can we delete it?What you build: Retention settings, deletion and export tools, region choice where offeredProof to hand over: Data flow diagram; deletion job; contract clause
AI controlsWhat the buyer's security team asks: Can we switch AI off, and does our data train your models?What you build: Per-tenant AI settings enforced on the serverProof to hand over: Admin screen; test showing the switch blocks provider calls
Contract and evidenceWhat the buyer's security team asks: What do you promise in writing, and who are your subprocessors?What you build: DPA, subprocessor list, trust page, answer libraryProof to hand over: The documents themselves

Treat the last column as part of the build. A control you cannot show is, to a reviewer, a control you do not have.

How do you add enterprise SSO and provisioning?

A neat row of identical blank lanyard badges laid out in a tray on a reception counter, soft morning light, shallow depth of field.

SAML 2.0 and OpenID Connect are the two common standards for enterprise SSO. Support at least one, and expect the buyer's identity team to ask which. Provisioning is a separate protocol: SCIM, defined in RFC 7643 (schema) and RFC 7644 (protocol), lets the customer's identity provider create, update and deactivate users in your product.

  • Configure SSO per tenant. Each customer stores its own identity provider metadata, certificates and attribute mapping. One global setting will not work once you have two enterprise customers.
  • Verify the customer's domains. Let a tenant enforce SSO only for email domains it has proven it controls, for example with a DNS record.
  • Enforce SSO, do not just offer it. Give admins a setting that turns off password login for their users. A buyer that cannot enforce SSO will count it as a gap.
  • Make deprovisioning real. When the identity provider deactivates a user, revoke their sessions and API tokens, and reassign what they own, such as scheduled jobs. Removing the login but leaving a live API key is the classic miss.
  • Map identity provider groups to roles. Customers want to manage access in one place.
  • Keep a break-glass path. If the customer's identity provider is down or misconfigured, someone needs a controlled way back in. Document it, restrict it to named admins and log every use.

Test against the identity providers your buyers use, such as Microsoft Entra ID, Okta or Google Workspace. Whether you build this or use an off-the-shelf identity layer, the per-tenant settings and deprovisioning behaviour are yours to get right.

What roles and permissions do enterprise customers expect?

A workable start is owner, admin, member and read-only roles, with finer control later and security settings kept apart from day-to-day admin rights.

  • Check permissions on the server. Hiding a button is not access control. Every request should be authorised against the user's role and tenant.
  • Scope API keys. Tie each key to a user or a service account, give it only the permissions it needs, show when it was last used and let admins revoke it.
  • Restrict security settings. SSO configuration, AI settings and data export are high-impact. Many buyers want them limited to a named admin role.

What should a customer-facing audit log contain?

An audit log answers one question: who did what, to what, and when. The customer's security team uses it for investigations. It is not your debugging log: it needs a stable format, tenant scoping and a way out of your system.

FieldWhy it matters
Timestamp in UTCWhy it matters: Lets the customer line your events up with their own systems
Actor (user, service account or system) and tenantWhy it matters: Establishes who acted and on whose data
Action and targetWhy it matters: For example role changed, API key created, export started
ResultWhy it matters: Success or failure, including failed sign-ins
Source IPWhy it matters: Helps spot access from unexpected places
Request or event IDWhy it matters: Lets your support team trace one event end to end

Log these events at a minimum: sign-ins and failures, SSO configuration changes, user creation and deactivation, role changes, API key creation and revocation, data exports, deletions, and changes to AI settings. Then add the rules that make a log trustworthy:

  • Write it on the server and append only. Users and admins should not be able to edit or delete entries.
  • Keep secrets and content out. Log that a document was exported, not the document. Log that an AI request happened, with its feature and size, not the prompt text.
  • Let admins export it. A download or an API is the baseline. Streaming to the customer's SIEM is a common later request.
  • State the retention period. Say how long entries are kept and whether the customer can extend it.

For AI features, ISO/IEC 42001 has a control on AI system recording of event logs (Annex A, A.6.2.8), so an AI usage trail also helps any later certification work. If you sell to Indian customers, remember that the DPDP Rules set a one-year minimum for keeping processing logs, and CERT-In directions call for logs to be kept for 180 days in India. Ask counsel how those apply to your retention settings.

Which tenant-level AI controls do enterprise buyers expect?

A long row of identical plain rocker light switches on a pale office corridor wall, some up and some down, softly lit, shallow depth of field.

AI features are where enterprise reviews get slow. The buyer's security and privacy teams want to decide, for their own organisation, whether AI runs, what it touches and what is kept. If the only answer is "it is on for everyone", you are asking them to accept your risk appetite. Tenant-level settings let them set their own.

SettingWhat it doesA sensible default to consider
AI on or off, per featureWhat it does: An admin can turn each AI feature on or off for the whole organisationA sensible default to consider: Off until an admin enables it, or on only where the feature is low risk and clearly labelled
Use of data for model trainingWhat it does: Controls whether customer content is ever used to train or improve modelsA sensible default to consider: No. Make any exception opt-in per tenant and put it in the contract
Prompt and output retentionWhat it does: Sets how long AI history is stored, including zero where the feature allows itA sensible default to consider: Shortest period the feature can work with
Data sources the AI can readWhat it does: Lets admins choose which connected apps, folders or workspaces feed retrievalA sensible default to consider: Nothing connected until the admin picks it
Model and provider disclosureWhat it does: Shows which provider handles requests, linked to your subprocessor listA sensible default to consider: Visible to admins
Human approval for actionsWhat it does: Requires a person to confirm before an AI agent changes data or sends anythingA sensible default to consider: Required for any write action
AI usage and audit exportWhat it does: Lets admins see AI use per user and per featureA sensible default to consider: Metadata only, written to the audit log

Whether a customer's content trains your models is a question that often slows deals, so spell it out. Our post on answering "do you train on customer data?" covers the wording.

How do you enforce AI settings so they actually work?

A setting that only hides a button will fail the first time a reviewer asks you to prove it. Enforce each one where the data moves:

  • Check settings in one place. Route model calls through a single gateway or service layer that reads the tenant's settings before it calls any provider.
  • Cover every path. Background jobs, embedding pipelines, scheduled summaries, integrations and the public API all need the same check. Most gaps are in the paths nobody thought of as "the AI feature".
  • Define what off means. Decide what happens to existing embeddings, AI history and cached outputs when a tenant switches a feature off, and write it down.
  • Log every change. Record who changed which AI setting, from what to what, and when, in the same audit log.
  • Keep a kill switch. You need to be able to disable an AI feature for one tenant, or for everyone, quickly if a model provider has an incident or a feature misbehaves. If your product has AI agents, buyers ask more, which we cover in new questionnaire questions about AI agents.
  • Write a test for each setting. For example: turn the feature off for tenant A, make the request, and assert that no call reached the provider and no new embedding was written.

What should your contract and trust page say about AI?

Buyers compare your terms with what the big AI vendors already publish, so the baseline has moved. As of 20 November 2026, no training on customer content, customer ownership of output and notice of subprocessor changes are standard in major vendors' public terms, such as the OpenAI Services Agreement and Data Processing Addendum and the Anthropic Commercial Terms. Breach notice in the OpenAI DPA is "without undue delay", with no fixed number of hours, so tighter hour limits are negotiated deal by deal. Check that what you promise customers matches what your model providers promise you, and have your lawyers review the wording. Our post on AI clauses in an enterprise MSA goes through each clause.

A trust page helps you answer once instead of in every questionnaire. A mature AI trust centre shows certificates, the statement of applicability, system cards and subprocessors, and gates audit reports behind a request. OpenAI's trust portal is an example of the pattern. Start with your subprocessor list, a data handling summary and your AI settings documentation. Our AI trust centre guide goes through what to publish.

Do you need SOC 2, ISO 42001 or a compliance platform for the first big customer?

Often not before the first signature, but you should know what each is. A first customer commonly accepts a completed questionnaire, working controls and clear evidence.

  • SOC 2 is an attestation report from a licensed CPA firm, covering controls against the Trust Services Criteria. It is a CPA's opinion, not a certificate. The AICPA has exposure drafts of updated criteria open for comment until 15 January 2027, which add AI considerations.
  • ISO/IEC 42001 is a certifiable AI management system standard. ISO does not certify; accredited certification bodies do. Mono provides readiness work and does not certify anyone. See our ISO 42001 vs SOC 2 comparison.
  • NIST AI RMF 1.0 (a revision is in progress) is voluntary and has no certification. It is useful for structuring your own answers on AI risk across its Govern, Map, Measure and Manage functions.
  • EU AI Act. An Indian SaaS vendor selling an AI feature to EU customers is usually a provider, and the EU customer is the deployer. The Act reaches providers outside the EU whose AI is placed on the EU market. The AI Omnibus, published in the Official Journal on 24 July 2026, moved the Annex III high-risk date to 2 December 2027, but obligations remain, and transparency rules have applied since 2 August 2026. Ask counsel what applies to you.

Compliance platforms such as Vanta and Drata automate evidence collection and control monitoring. As of 20 November 2026, Vanta's pages describe AI governance support mapped to ISO 42001, the EU AI Act and NIST AI RMF, and Drata lists ISO 42001, NIST AI RMF and AIUC-1 among its frameworks. Neither publishes list prices. They track and evidence controls. They do not build tenant AI settings, audit log exports or guardrails in your codebase.

What should you build first?

Order the work by what blocks the deal, then by effort. This is a typical sequence, not a fixed plan.

StageBuildWhy this order
FirstBuild: Audit log events and a tenant AI settings schema; inventory every AI feature and model providerWhy this order: Everything else writes to or reads from these
SecondBuild: Per-tenant SSO with enforcement, then SCIM deprovisioningWhy this order: Identity is often one of the first items buyers' security teams check
ThirdBuild: Roles, scoped API keys, log export, AI toggles enforced through one gateway, with testsWhy this order: Closes the most common questionnaire gaps
FourthBuild: Retention and deletion settings, kill switch, trust page, answer libraryWhy this order: Turns controls into evidence you can reuse
LaterBuild: SIEM streaming, custom roles, SOC 2 or ISO 42001 if buyers demand itWhy this order: Rarely needed to sign the first customer

Before you start, read your questionnaire answers against what the product does. The gaps are your real backlog.

Where Mono helps

We build the enterprise-ready parts of AI products and fix the gaps that stall reviews. For new AI features, our AI product development work includes tenant settings, audit logging and server-side enforcement from the start. For products already in the field, our AI governance work for SaaS reviews what your buyers will ask, closes engineering gaps such as logging, tenant isolation and admin controls, and assembles the evidence so each answer has proof.

This is engineering and evidence work, not legal advice and not a certification. Your lawyers own the contract language, and only an accredited body can certify against a standard.

Sources: SAML V2.0, OASIS Standard. OpenID Connect Core 1.0. RFC 7643 and RFC 7644, System for Cross-domain Identity Management. ISO/IEC 42001:2023, Annex A (A.6.2.8) and ISO/IEC 42006:2025. NIST AI RMF 1.0 and NIST AI 600-1. Regulation (EU) 2024/1689 (Articles 2 and 3) and the European Commission AI Act timeline page, including the AI Omnibus. OpenAI Services Agreement and Data Processing Addendum, openai.com/policies. Anthropic Commercial Terms, anthropic.com/legal/commercial-terms. AICPA SOC suite of services page and 2027 TSC exposure draft announcement. Vanta AI governance and ISO 42001 product pages; Drata ISO 42001 and frameworks pages, as of 20 November 2026. Digital Personal Data Protection Rules, 2025, Rule 6. Last reviewed 20 November 2026.

FAQ

What does enterprise readiness mean for a SaaS product?

It means your product has the controls an enterprise security team checks before signing: single sign-on, automatic user provisioning and deprovisioning, role-based access, exportable audit logs, data retention and deletion controls, and, for AI features, tenant-level AI settings. It also means you can show proof for each one, such as a config, a test or a screenshot.

Do I need SSO to sell to enterprise customers?

Most enterprise buyers ask for it, and many treat enforced SSO as a baseline. Support SAML 2.0 or OpenID Connect per tenant, let admins enforce it, and add SCIM so that deactivating a user in the customer's identity provider removes their access to your product, including sessions and API tokens.

What tenant-level AI controls should a SaaS product have?

Buyers commonly expect an admin switch for each AI feature, a clear no-training default with any exception opt-in, retention settings for prompts and outputs, a choice of which data sources the AI can read, visibility of the model provider, human approval for AI actions that change data, and AI usage in the audit log. Enforce each setting on the server and test it.

What should an enterprise audit log include?

At minimum: a UTC timestamp, the actor and tenant, the action and target, the result, source IP and a request ID. Cover sign-ins, SSO and role changes, API key events, exports, deletions and AI setting changes. Write it on the server, append only, keep secrets and content out, and let customer admins export it with a stated retention period.

Do I need SOC 2 or ISO 42001 before my first enterprise customer?

Not always. Many first customers accept a completed security questionnaire, working controls and clear evidence, especially with a plan for what comes next. SOC 2 is a CPA attestation report and ISO/IEC 42001 certification comes from accredited certification bodies. Mono does not certify anyone. Check what your buyer requires in writing before you commit to either.

Does the EU AI Act apply to an Indian SaaS company?

It can. The Act covers providers placing AI systems on the EU market even when they are established in a third country, and it covers outputs used in the EU. An Indian vendor selling an AI feature to EU customers is usually a provider. The AI Omnibus moved the Annex III high-risk date to 2 December 2027, but obligations remain. Ask counsel how it applies to your product.

Closing your first enterprise deal? Let us review your product against what buyers check.