Try the demo

How Canon uses AI

Canon advises; people decide. Here is exactly what the AI does, what it sees, and what keeps it honest.

Checked without AI

Banned words, preferred terms, required disclaimers, punctuation rules, and numbers that don't match the fact ledger. These are exact, repeatable checks with no AI involved.

Checked with AI

Tone, unsupported or exaggerated claims, what the writer meant, and the wording of a suggested fix. Every AI finding is labelled and shows a confidence score.

Measured in images, without AI

Text contrast (WCAG 2.1 AA), retired logos and missing alt text are measurements, not opinions. The same image always gives the same result.

Reading text in images

Canon uses Tesseract OCR when it's installed. Where it isn't (as on this hosted demo), the image is sent to the AI to read its text and draft alt text. People can correct what was read before it's checked.

What keeps the AI honest

  • Every finding must quote the draft exactly and cite a real rule or fact ID. Anything else is dropped before you see it, and the number dropped is shown.
  • If the rule engine already found an issue, its exact finding wins over the AI's.
  • Low-confidence and high-severity findings are marked for human review.
  • Nothing is applied, approved or published without a person clicking.
  • If the AI is down or returns bad output twice, Canon still shows rule-engine results and says the AI review was unavailable.
  • The reliability scorecard measures misses and false flags on a fixed test set.

What is sent to the AI service

When AI review is on, Canon sends the draft text (and, where no local OCR is installed, uploaded images), the brand voice, the active rules and the facts the author is allowed to see to the AI service (Google Gemini API). Before sending, 2 confidential terms are replaced with placeholders. Names and emails of users are never sent.

Masking only covers terms on the confidential list; it can't recognise confidential information that isn't listed, and it can't mask text inside an image that is sent for reading. Agency partners' drafts are checked only against facts shared with partners, so internal facts never leave the company through an agency draft.

This demo uses Google's free Gemini API tier. Under Google's terms for unpaid services, submitted content may be used to improve Google's products and may be read by human reviewers, and Google asks users not to submit sensitive or confidential information. That's acceptable here because all demo data is fictional. A real deployment would use a paid tier, where these terms differ.

Status in this deployment: AI review on (Google Gemini · gemini-3.1-flash-lite).

Accountability

Every check, fix, override (with its reason), approval and fact change is recorded with who did it and when. Admins can see the full audit log; everyone can see the history of the content they work on.