Website governance for AI agents: legal, brand and accessibility checks before anything publishes

When people wrote every page, governance lived in habits: the web manager knew which claims legal had cleared, and changes were slow enough to review. AI agents remove the slowness. An agent can rewrite forty product descriptions in an afternoon. If the review model was a policy document nobody opens, volume alone will push mistakes through.

The fix is to govern the moment a change becomes public. Agents can draft, edit and propose as much as you like. A change only goes live after it passes a defined set of checks and a person with the right authority says yes. That is the whole model. The rest of this article is the detail: what to check, who approves, what to automate, what to log and how to roll it out.

If you are still deciding which tasks to hand to agents at all, start with our list of what AI agents can and cannot do on a corporate website. If you want the agent roster, branches and review board, read our AI agent web team operating model. This article sits between the two: the governance layer that legal, compliance and brand reviewers sign.

We group pre-publish checks into six areas. Each one has a rule an agent can apply and a decision a person owns.

• Legal and regulatory claims. In Canada, the Competition Bureau looks at the general impression a representation creates, in addition to its literal meaning, so a claim can be literally true and still mislead. Performance claims must rest on an adequate and proper test. Agents may only use statements from an approved claims file, and any new claim, comparison, certification or environmental statement goes to legal.
• Brand and tone. Voice rules, banned words, product naming, capitalisation and positioning statements. Agents check against the written guide; the brand lead decides anything new, such as a tagline, a category name or a change in how you describe the company.
• Accessibility to WCAG 2.2 AA. WCAG 2.2 became a W3C Recommendation in October 2023, with the current version dated 12 December 2024, and added nine success criteria, six of them at level A or AA: Focus Not Obscured (Minimum), Dragging Movements, Target Size (Minimum), Consistent Help, Redundant Entry and Accessible Authentication (Minimum). It also removed 4.1.1 Parsing. Automated tools test part of this; people test keyboard and screen reader flows.
• Privacy and consent. Any new form, tracking tag, chat widget or embedded tool changes what personal information you collect. The Office of the Privacy Commissioner of Canada says consent should generally be express and must be express when information is sensitive, outside reasonable expectations or carries a meaningful risk of significant harm. In Quebec, consent must be clear, free and informed, given for specific purposes and requested for each purpose, and anyone using technology that can identify, locate or profile a person must first say so and explain how to activate it.
• Proof and testimonials. Logos, quotes, case study figures and review snippets. The Competition Bureau notes that consumers assume a testimonial comes from someone who used the product, and that an undisclosed material connection between reviewer and advertiser is a concern even when the review is honest. An agent cannot confirm that a customer agreed to be named, so proof is always a human check.
• Pricing and commercial terms. Prices, discounts, plan inclusions, free trial conditions and currency. Agents may only show prices that match the approved price source, and any change goes to the commercial owner.

The line is simple to state: if a rule can describe it, automate the check; if it needs judgment or accountability, a person signs. Automated checks run first, so reviewers only see changes that already pass the basics.

Checks we automate on every agent change:

• Claims match: every claim, number and certification on the page matches an entry in the approved claims file, or the change is flagged for legal.
• Banned words and naming: product names, trademarks and banned terms checked against the brand guide.
• Accessibility: contrast, labels, heading order, alt text presence, target size and focus visibility on every changed template.
• New data collection: any new form field, script, pixel or third-party embed is detected and routed to privacy review.
• Proof match: every logo, quote or figure maps to a record showing permission and date.
• Price match: every price on the page equals the value in the price source.

The W3C itself warns that evaluation tools cannot check every accessibility requirement, and the same is true of every category above. Automated checks reduce what reaches a reviewer. They do not replace the reviewer, and a passing check should never be presented as legal, privacy or accessibility conformance.

This is general information, not legal advice. Confirm your obligations with qualified counsel.

Agents are never accountable. They draft and run checks; a person approves. Use this RACI as a starting point and replace roles with names. R is responsible for doing the work, A is accountable and signs, C is consulted, I is informed.

Two rules keep the table honest. One person holds A per row, never a committee. And the approval tool, whether a pull request, a CMS workflow or our human-in-the-loop approval workflow, must block publishing until that person has signed.

Legal and compliance reviewers will ask one question after an incident: who changed this, why and who approved it. The audit trail should answer it in a minute. For every agent change, record the brief or rule that triggered it, the diff, the result of each automated check, the sources behind any new fact, the approver, the time of approval and the time it went live.

Version control gives you most of this when every change is a reviewed commit. On a traditional CMS, turn on revision history and approval logs, and keep agent accounts separate from human accounts.

Rollback is the other half. Any published change should be reversible in one step, by the web owner, without a developer. Test it before the first agent change goes live, and again whenever the stack changes. A rollback that has never been tried is a hope, not a control.

Agents follow written rules far better than unwritten ones, and reviewers can only audit what is written. These are the documents we set up with clients before agents touch production pages.

• Approved claims file: every claim, figure and certification you may use, with its source, owner and review date.
• Brand and voice guide in a form an agent can apply: rules, examples and banned terms, not a mood board.
• Accessibility standard: WCAG 2.2 AA as the target, which checks are automated and who performs manual testing.
• Privacy and consent rules: which data each form may collect, which tags need consent, and who approves new collection.
• Proof register: each customer logo, quote and figure, with permission, date and expiry.
• Price source: the single place prices come from, and who may change it.
• Agent permissions: read by default, write to drafts or branches only, no publish rights, no access to production secrets.
• Incident and rollback procedure: who acts, how fast, and how affected teams are told.

If your company already uses a formal AI framework, map these documents to it. NIST’s AI Risk Management Framework, released in January 2023 for voluntary use, is organised around four functions: Govern, Map, Measure and Manage. NIST added a Generative AI Profile, NIST AI 600-1, in July 2024. The documents above are the website-level Govern and Manage pieces of that picture.

The aim of the first month is a working gate on real changes, not a perfect policy set. Keep scope narrow and let evidence decide what comes next.

1. Days 1 to 5: inventory the site by risk. Tag each section as low risk (blog, events), medium (product and solution pages) or high (pricing, legal, security, customer proof, regulated product claims).
2. Days 1 to 5: name one accountable approver per change type and fill in the RACI with real names.
3. Days 6 to 10: write the claims file, proof register and price source from what is live today, with legal reviewing the claims file once.
4. Days 6 to 10: set agent permissions to drafts or branches only, and turn on the audit trail and approval blocking in your CMS or repository.
5. Days 11 to 15: configure the automated checks for claims, naming, accessibility, new data collection, proof and price, and test a rollback on staging.
6. Days 16 to 25: run agents on low-risk sections only. Every change goes through the gate, and approvers record whether they approved, edited or rejected it and why.
7. Days 26 to 30: review the log with legal, brand and privacy: which checks caught issues, which ones reviewers caught that checks missed, and how long approvals took. Add rules for the misses, then decide whether medium-risk sections join next month.

High-risk sections stay with people drafting and approving until the log shows the gate works on lower-risk pages. That evidence is what convinces a compliance reviewer, more than any vendor promise.

Good website governance with AI agents is less about restricting agents and more about making the publish step strict. When the checks are written down, the approver is named and every change leaves a record, agents can move fast where it is safe, and reviewers can see exactly what they signed.

Our website development service builds this gate on your stack, with the claims file, RACI, automated checks and audit trail in place before the first agent change ships. If you want to see what it would take on your site, talk to our team.