Guide for SaaS teams

How to build an AI trust center.

Enterprise buyers ask the same AI questions in every deal. A trust center answers them once: what your AI features do, what data they touch, and what you can prove. Here is what to publish, what to keep gated, and a lightweight model card to start with.

Publish onceModel card template
AI trust center

What buyers look for

DocumentsCertificates and reports, with the scope and date.
AI summaryFeatures, models and the data each one touches.
Model cardsPurpose, limits and oversight, per feature.
SubprocessorsIncluding your model providers.
Incident contactWho to call and what you promise.
Public and gated partsOnly what you can prove
Keep it current

What a trust center is, and is not.

It is a page where a buyer’s security team finds answers before they ask. It is not a certification and not a legal document. A good one shows your certificates, your Statement of Applicability if you have one, your model cards and your subprocessors, and puts audit reports behind a request step. OpenAI’s public trust portal, checked on 20 November 2026, is built that way.

Public

Your AI use summary, model cards, subprocessor list, data handling terms and security contact.

Behind request access

Audit reports, penetration test summaries, the Statement of Applicability, and completed questionnaires.

On request, in a deal

Anything specific to one buyer, such as a signed DPA or a named test result.

Never listed

A certificate or report you do not hold, or a claim you cannot show evidence for.

This page is general information, not legal advice.

What to publish, and what proves it.

Every line on a trust center is a promise. Write only what you can back with a setting, a log or a document. Our AI security questionnaire help uses the same rule for each answer.

Certificates and reports

What to sayWhat you hold, the scope, the date and who issued it. ISO/IEC 42001 only if an accredited certification body has certified you.

What proves itA table of documents with a request-access step for audit reports.

AI use summary

What to sayWhich features use AI, which model providers sit behind them, and what customer data reaches each.

What proves itOne short page per feature, linked from the trust center, kept in step with your data flow diagram.

Data handling terms

What to sayWhether customer data is used for training, how long it is kept, how it is deleted and where it is processed.

What proves itPlain statements that match your contract and your settings, with the setting or log that proves each one.

Subprocessors

What to sayEvery vendor that handles customer data, including model providers, and how you tell customers about changes.

What proves itA list with dates, plus a way for customers to subscribe to changes.

Human oversight and limits

What to sayWhere a person reviews AI output, and the known limits of each feature.

What proves itA short statement on each model card, backed by the product setting that enforces it.

Incident handling

What to sayWho customers contact, what you promise to tell them and when.

What proves itA published contact and a process you have tested at least once.

The model card, for a SaaS feature.

Model cards come from a 2019 paper by Mitchell and colleagues, “Model Cards for Model Reporting”. It describes them as short documents that accompany trained machine learning models and report how they perform across conditions and groups. For a SaaS product we suggest one card per AI feature, not per model, so the buyer has one stable page even when you change provider. The fields below are our suggestion, not a standard.

Feature, owner and version

What the feature is called, who is accountable, and the date of this card.

Purpose and out-of-scope use

What it is for, and what it must not be used for.

Model and hosting

Which model and provider sit behind it, and where it runs. Update the card when you change provider.

Data in and data out

What goes into the model, what is stored, and whether anything is used for training.

How it was tested

What you checked, on what data and when. Give only results you can reproduce.

Known limits and risks

Where it can be wrong, such as made-up facts, bias or leaked context, and what you do about each.

Human oversight

Who reviews the output, and how a customer admin switches the feature off.

Change log

What changed, when, and whether customers were told.

AI terms buyers expect to see.

Major model vendors now state no-training, output ownership and subprocessor notice in their public terms. Buyers expect the same from you. The points below come from public templates, checked on 20 November 2026. Have your own lawyer review anything you publish. For the wording of the training question, see how to answer it and prove it, and for the controls a first enterprise buyer checks, see enterprise readiness for your first big customer.

No training on customer data

OpenAI’s Services Agreement says it will not use Customer Content to develop or improve the Services unless the customer agrees. Anthropic’s Commercial Terms say it may not train models on Customer Content from its Services.

Who owns the output

Both vendors’ public terms leave Inputs with the customer and give the customer ownership of Output. The Common Paper AI Addendum, a free template, takes the same line.

Deletion

OpenAI’s Services Agreement deletes Customer Content within 30 days of termination. Pick a figure you can meet, and state it.

Subprocessor changes

OpenAI’s DPA gives notice of subprocessor changes, 30 days to object, and a right to terminate if the objection is not resolved.

Breach notice

OpenAI’s DPA says breach notice is “without undue delay”, with no hour figure. Tighter limits are a point buyers negotiate, so decide what you can promise before you are asked.

Say how you handle an AI incident.

Buyers want to know who they contact and when you will tell them. Your deadlines come from the law that applies to you, not from the AI standard.

ISO/IEC 42001

Annex A has controls for incident communication, external reporting and AI event logs. Clause 10.2 asks you to correct a nonconformity, find its cause and keep records. The standard sets no reporting deadline.

CERT-In and DPDP

CERT-In requires listed cyber incidents to be reported within 6 hours. DPDP Rule 7 asks for a detailed report to the Data Protection Board within 72 hours, and applies from 13 May 2027.

EU AI Act

Providers of high-risk AI systems report serious incidents under Article 73: within 15 days of becoming aware, with shorter limits for a widespread infringement, a critical-infrastructure disruption or a death.

AI-specific rules in India

As of November 2026, India has no AI-specific incident reporting rule. Publish the clocks that do apply to you, and do not invent a new one.

Which frameworks to map to.

You do not need all of them. Pick the ones your buyers name. See ISO 42001 and SOC 2 compared for what each proves, and the SIG AI domain and CSA AI-CAIQ for the questionnaires.

NIST AI RMF

AI RMF 1.0 is voluntary and has no certification. It has four functions: Govern, Map, Measure and Manage. NIST says a revision is in progress. Its generative AI profile, NIST AI 600-1 (July 2024), lists 12 risks, including confabulation, data privacy and information security, which map well to a model card’s limits section.

ISO/IEC 42001

A management system standard for organisations that build, provide or use AI. Only an accredited certification body can certify against it. Do not mention it on a trust page until you hold a certificate or have a dated plan you are willing to share.

EU AI Act

The AI Omnibus moved deadlines, but obligations remain. Annex III high-risk rules now apply from 2 December 2027, and Annex I from 2 August 2028. An Indian SaaS vendor selling an AI feature to EU customers is usually a provider, and the EU customer is the deployer.

A platform, or build it yourself?

What platforms do

Compliance platforms such as Vanta and Drata automate evidence collection and control monitoring, and Vanta offers a hosted trust center. As of 20 November 2026, Vanta lists support for ISO 42001, the EU AI Act and NIST AI RMF, and Drata lists ISO 42001 and NIST AI RMF. Features vary by plan, and neither publishes list prices.

What they do not do

A platform can host the page and track the controls. It does not build the product features behind the claims, such as tenant-level AI settings, audit logs and guardrails in your own code. That engineering is where we work, and it works alongside a platform.

How we help.

Mono does not certify you, audit you or give legal advice. We prepare the company, build the engineering and collect the evidence. If you are also heading for an audit, we can cover the same ground for ISO 42001 readiness.

  1. Step 1

    Inventory

    We list every AI feature, the model behind it, and the customer data that reaches it.

  2. Step 2

    Check the claims

    We test each statement you want to publish against your settings, contracts and logs. Anything we cannot show, you leave out.

  3. Step 3

    Build what is missing

    Tenant-level AI switches, audit logs and guardrails, built into your product, so the claims stay true.

  4. Step 4

    Write and publish

    Model cards, the AI summary and the evidence index. You publish, and your lawyer reviews the terms.

  5. Step 5

    Keep it current

    A short checklist run on each model or provider change, so the page does not go stale.

AI trust center FAQs.

What is an AI trust center?

It is a page, public or partly gated, where a SaaS company shows enterprise buyers how its product uses AI and how it protects their data. It usually holds your certificates, your list of subprocessors, a summary of each AI feature, data handling terms and a security contact. Its job is to answer the common questions once, so each deal does not start from a blank questionnaire.

Do we need one before our first enterprise customer?

No law or standard requires a trust center. It is a sales and security tool. It starts to pay off when the same AI questions arrive in several deals. If you are answering them for the second or third time, write the answers down once and publish what you can.

Should our SOC 2 report go on a public page?

Usually not. A SOC 2 report is a CPA firm’s opinion on your controls and is commonly shared only on request, often under an NDA. Publish the fact that you have one, the scope and the date, and put the report itself behind a request-access step. Never list a certificate or report you do not hold.

What is a model card, and do we need one per model?

A model card is a short document that travels with a model and says what it is for, how it was tested and where it fails. For a SaaS product, write one for each AI feature rather than each underlying model. A feature can change model provider and the buyer still needs one stable page that says what it does and what data it touches.

Can Mono certify us or write our trust center for us?

Mono does not certify anyone. We are not a certification body, a law firm or an auditor, and we do not give legal advice. We help you work out what is true about your AI features, build the product controls behind the claims, and collect the evidence. You publish the page, and your own lawyer should review the terms on it.

Put the answers where buyers can find them.

Sources: Mitchell et al., Model Cards for Model Reporting, FAT* 2019 (arXiv 1810.03993). NIST AI RMF 1.0 and NIST AI 600-1 (26 July 2024). Regulation (EU) 2024/1689, Articles 2, 3 and 73, and the European Commission AI Act timeline as amended by the AI Omnibus (Official Journal, 24 July 2026). OpenAI Services Agreement and Data Processing Addendum (effective 1 January 2026). Anthropic Commercial Terms (17 June 2025). Common Paper AI Addendum v1.0. OpenAI Trust Portal, Vanta and Drata product pages, checked 20 November 2026. CERT-In Directions, 28 April 2022. Digital Personal Data Protection Rules, 2025, Rule 7. Last reviewed 20 November 2026. See also our AI governance page.