Draft · not in forceVersion 0.3Last changed 2026-09-20

AI Use and Disclosure Statement

Where AI sits in Troy’s work, where it does not, what gets recorded and labelled, and who decides whether a piece of advertising says so.

Who this applies to
  • Clients Troy does marketing work for, and the people at a client who approve work
  • Work Troy makes, or helps make, for a client to publish
  • Troy’s own staff, contractors and anyone working on Troy’s behalf
  • Anyone who wants to know where AI sits in work Troy has touched

In plain EnglishThis describes how Troy works with AI. It is not legal advice, and it is not a claim to be compliant with anything.

Troy is a marketing company that uses AI tools in its work. This statement sets out where those tools are used, where they are not, what a person is answerable for, what Troy records, what a client is asked to confirm before work ships, and what is disclosed to the public.

It is published so that a client, a platform, a regulator or a member of the public can see where the machine is without having to ask. Anyone may ask anyway — clause 12 says how.

  • It is not a claim of compliance. Troy holds no certification, has not been audited against any standard, and does not describe itself as compliant with any law or framework.
  • It is not a contract. Where it differs from a signed agreement, the agreement governs. The service terms and the data processing agreement say what Troy may do with a client’s material and data.
  • It is not a list of tools. It describes what the tools are used for and who is answerable, which is what matters to a reader and does not go stale every time a vendor is changed.
  • It is not finished. It is a draft, version 0.3, and no lawyer has read it.

In plain EnglishSix words that do a lot of work below. The three content categories are the ones the rest of the document turns on.

Troy
Troy Agency Inc., trading as Troy, and the people who work for it. Used throughout for both.
AI tool
A model or service supplied by another company that generates or edits text, images, audio or video, summarises data, or ranks options. Troy builds no models of its own and trains none.
Fully human
Material made without an AI tool at any stage.
AI-assisted
Material a person wrote, briefed, shot or selected, where an AI tool did part of the work: a rewrite, a summary, a resize, a background removal, a spelling pass.
AI-generated
Material an AI tool produced in substance, where the person’s part was the brief, the review and the decision to use it.
Provenance record
The internal record Troy keeps against a piece of work: which category it falls in, which tool was involved and for which part, who reviewed it, and who approved it.
Ships
The moment work leaves Troy for an audience — a live ad, an email send, a published page. Work shown to a client for approval has not shipped.

In plain EnglishAI drafts, summarises and ranks. It does not decide what ships, and it does not spend money.

Troy’s strategists run client marketing on an internal console assembled from third-party tools, the ad platforms’ official connectors and AI workflows. That is where the AI is. The list below is what it does and does not do in that work.

In plain EnglishDrafting, summarising, ranking and routine production. All of it lands in front of a person.

  • Drafts copy and creative variants under the client’s brand rules, for a person to edit, cut or reject.
  • Summarises account data that has already been read from a connected account, so a strategist can see a month of activity quickly.
  • Ranks fixes. The Troy Score’s findings are put in order of how much they matter, against the published formula, and a person reviews that order before a client sees it.
  • Drafts the commentary around a report: the sentences, never the figures.
  • Turns repeated internal work into written playbooks, and drafts the internal notes strategists keep.
  • Does routine production work on material the client already owns, such as resizing and cutting creative down for different placements.

In plain EnglishThe limits. No tool at Troy decides, spends, supplies a number, sends, or judges a person.

  • No AI decides what ships. A named person at Troy approves every piece of work before a client sees it, and a named person at the client approves it before it reaches an audience.
  • No AI spends money. Troy never holds a client’s media money — clients pay the ad platforms directly. Where Troy has spend authority, a change is proposed with its evidence, approved by a named person at the client, and held under a hard cap.
  • No number in a report comes from a model. Every figure traces back to the API response it came from. A model may write the sentence around a number; it never supplies the number, and a figure that cannot be traced is deleted rather than corrected.
  • No AI sends. Drafts land in the client’s own email tool as drafts. A send needs the client’s approval and the client’s confirmation that the list has consented. Troy never imports lists.
  • No AI talks to a client’s customers. Troy runs no consumer-facing chat agent.
  • No AI makes a decision about an individual person. See clause 11.

In plain EnglishParts of what this site describes are not built. This says which parts.

  • The agents are not a product a client can switch on. They open by invitation from 2027. Until then the same workflows are run by strategists, which is the work this statement describes.
  • Every Troy Score is prepared by a strategist by hand from read-only access. It is automated later, and the Score page says so.
  • Peer comparison does not exist. It needs enough connected accounts to compare against, and Troy does not have them.
  • Troy has no models of its own, and none are planned in the first version.

In plain EnglishA person is accountable for everything that reaches a customer. Not a tool, not a vendor, not a model.

One rule sits above the rest of this document, and the rest of it is mostly machinery for keeping that rule true.

  1. Everything that reaches an audience has a named person at Troy who approved it and a named person at the client who approved it. Both are written down, and the record is kept as clause 5.2 describes.
  2. Approval is a decision, not a formality. It means a person read the work, checked what it claims, and is answerable for it. A tick applied without reading is a breach of this statement, not a shortcut.
  3. An AI tool’s output is a draft until a person says otherwise. A tool sounding confident is not a reason for anything.
  4. Accountability does not move to a vendor. A model vendor’s terms may disclaim everything; that changes nothing about who answers to the client.
  5. If a person cannot explain why a piece of work says what it says, it does not ship.

In plain EnglishThree categories. Each has a record Troy keeps, a confirmation Troy asks you for, and a public disclosure question that is yours to answer.

Every piece of work Troy makes is put in one of three categories by a person, and the category is written down before the work is shown to the client. The categories are defined in clause 2.

What Troy records, asks and discloses, by content type
Content typeWhat Troy records internallyWhat the client confirms before it shipsWhat is disclosed publicly
Fully humanWho made it, who reviewed it, who approved it, and the date of each.That they approve the work. No AI question arises.Nothing about AI, because there is nothing to disclose.
AI-assistedThe above, plus which tool was involved and which part of the work it did.That they have seen which parts were AI-assisted, and that they accept the work as theirs to publish.The client’s decision. Troy applies whatever they choose — a visible label, a platform’s own AI declaration, both, or neither — and records which was chosen and who chose it.
AI-generatedThe above, plus the brief or prompt, the version that was kept, and the person who reviewed the output line by line.That they know the material is AI-generated, which parts, and that they have decided whether to label it where it runs. Troy asks this in writing and keeps the answer.The client’s decision, applied before the work ships. Troy will say when it believes a label is needed, and will not ship work whose labelling it thinks is misleading. See clause 6.

Where a piece sits between two categories, Troy records the higher one. The doubt is resolved towards disclosure, not away from it.

In plain EnglishWorked examples, because the categories are easy to argue about in the abstract.

  • A spelling and grammar pass over copy a strategist wrote: AI-assisted. The words are the person’s.
  • Five headlines a model produced from a brief, three of which a strategist kept and edited: AI-generated. The words came from the model, and editing them does not change where they came from.
  • A photograph the client supplied, with the background removed by a tool: AI-assisted. The subject is real and was really photographed.
  • A scene that never existed, built by an image model: AI-generated, and subject to clause 7.
  • A weekly report where every figure came from an API response and a model drafted the commentary: AI-assisted, with the commentary checked line by line against the figures.
  • A Troy Score: AI-assisted at most. The formula is published, the findings are evidenced, and a strategist prepares and reviews it.

In plain EnglishTroy writes down what was AI-made at the time, not afterwards from memory. Where that record is held, and who can see it, is here too.

  • The record is made when the work is made. A record reconstructed later is marked as reconstructed.
  • It holds the category, the tool and the part it did, the brief or prompt where one exists, the reviewer, the approver, the client’s labelling decision, and the dates.
  • A client may ask for the record behind any piece of work Troy still holds, and Troy gives it. It is the client’s work.
  • The record is kept for as long as the client record, so a question about a piece of work can always be answered (a proposed default, not yet approved). The client’s file is kept for 7 years after the engagement ends (a proposed default, not yet approved).
  • Changes made through Troy’s approval queue are written to a permanent change log: a revert adds an entry, it never removes one. Until that queue exists as a product a client can use, the same record is kept by hand against the client’s file.

The record outlives the campaign because the questions do. A complaint, a platform review or a claim about a piece of work can arrive long after it stopped running: in Ontario, where Troy is based, the basic limitation period is two years from the day the claim was discovered, with an outside limit of fifteen years from the act complained of, though whether either period runs, and from when, depends on the facts and on where the claim is brought. Troy keeps the record so the question can be answered, not so it can be argued.

Where the record is held matters as much as what is in it, so this is said plainly. Troy, Louin and Lunari Group share systems and data. Client data is held on systems Lunari Group runs, and the companies in the group can see each other's data. The provenance record, and the client file it sits against, are on those systems. Troy is based in Ontario, and not all of the systems its work sits on are in Canada — the privacy policy says where they are and what carries the transfer.

In plain EnglishWhether an advertisement says it was made with AI is the advertiser’s call and the advertiser’s legal responsibility. Troy’s job is to make sure you can make it.

Troy is not the advertiser. The client is. What a given piece of advertising must disclose depends on what is in it, which platform it runs on, where the audience is, and rules that differ between those markets and are changing. Troy does not make that decision for a client and does not claim to know the answer for every market.

  • Troy tells the client what was AI-made, and which parts, before the work ships. That is what the provenance record is for.
  • Troy applies the client’s decision — a visible label, a platform AI declaration, both or neither — and records which was chosen and who chose it.
  • Troy will not remove a label a client asks for, whatever it might do to performance, and will not argue a client out of one.
  • Troy will add a label on the client’s instruction to work that does not strictly need one. Over-disclosure is the client’s prerogative.
  • Troy will not describe material as human-made when it is not, will not put a client’s name to a claim it cannot evidence, and will not present an AI-made image as a photograph of a real event.
  • Troy makes no promise that AI-made material can be owned, registered or is free of other people’s rights. Ownership of work product is dealt with in the service terms.

In plain EnglishTroy does not make fake people, fake voices or fake endorsements. Whoever asks.

Some AI output is a tool doing production work. Some of it is a machine manufacturing a person who does not exist and putting words in their mouth. Troy does the first and not the second.

  • No photoreal AI people. No synthetic person presented as a real customer, model, employee, founder or expert.
  • No AI voices that imitate a real person, and no voice clone of a client, an employee or a public figure.
  • No synthetic endorsements. No invented testimonial, review, quote, rating or case study, written, spoken or shown. Troy publishes no result it cannot evidence, for a client or for itself.
  • No real person’s face, voice or likeness generated or altered without that person’s written permission.
  • No depiction of an event that did not happen, presented as though it did.
  • No AI-made image or video of a person used to imply that a real person used, endorsed or was harmed by a product.

Illustration, pattern, texture, background, type treatment and product-in-scene composition may be AI-generated, under the same record, the same review and the same approval as everything else. The test is whether a reasonable viewer would take the result for a photograph of something that really happened.

These limits are a floor, not a preference, and they do not move for a client who asks nicely or pays more. The Acceptable Use Policy sets out the work Troy refuses, and these limits sit inside it.

In plain EnglishTroy does not train models on your data. Where a vendor’s own terms would let it, Troy will say so on the sub-processor list rather than leave you to find it.

  • Troy builds, trains and fine-tunes no models. It uses tools supplied by other companies.
  • Troy does not train, fine-tune or otherwise improve any model on a client’s content, account data, customer lists, creative or results.
  • Troy does not hand a client’s data to a vendor for that vendor’s own training, and will not use a vendor or a tier that reserves the right to train on what is sent to it. Where a vendor offers a plan or a setting that keeps content out of training, that is the one Troy takes.
  • Vendors keep prompts and outputs for a while even when they do not train on them, usually to check for abuse. Where that is so, Troy states the period against the vendor on the sub-processor list, in plain words, and a vendor changing its training or retention terms is a change to that list.
  • Comparing a client’s efficiency with peers uses aggregate figures only, needs separate opt-in consent, and that consent is never tied to a discount or a price. A client who declines loses nothing and is charged the same.
  • Troy does not use one client’s material to make another client’s work. That is a statement about Troy’s clients, not about the group Troy sits in: who can see a client’s data is dealt with at clause 5.2 and in the privacy policy.

The model vendors are OpenAI and Anthropic. Which plan each is used under, and whether that plan lets the provider train on what Troy sends, is To be completed: Which plan is each of OpenAI and Anthropic used under, and does that plan allow the provider to train on what Troy sends? Both say their business and API terms exclude training by default, but the answer depends on the plan and has to be confirmed rather than assumed — a client will ask, and the sub-processor list has to state it.. Both are United States companies, so material Troy sends to a model leaves Canada. The transfer rests on the Standard Contractual Clauses, with the UK Addendum where the United Kingdom is involved, entered into through each supplier's own data processing agreement (a proposed default, not yet approved), and the sub-processor list is where each vendor’s own terms, including how long it keeps prompts, are set out.

In plain EnglishAn AI draft is checked before you see it. Some things only you can check, and this says which.

AI tools are wrong confidently and often. Everything below exists because reading well is not evidence of being right.

In plain EnglishWhat a draft is checked against before a client is asked to look at it.

  • Every figure, against the response it came from. A number a model produced is deleted, not corrected.
  • Every claim about the client’s product, price, stock, delivery or terms, against something the client supplied. An unsupported claim is cut.
  • Every claim of cause. A person or an agent may say what changed. Saying why needs an experiment, and without one the claim is removed.
  • Brand rules: the client’s words, the client’s tone, and the client’s do-not-say list.
  • Obvious infringement: a competitor’s trademark, a recognisable face, a logo, a lyric, or a style passed off as another company’s.
  • The category and the label, so both are right before the client is asked to approve.
  • Whether the draft can be checked at all. Work that cannot be traced back to something real is not sent.

In plain EnglishThe things Troy has no way of knowing, so it asks the client.

Troy asks for these because Troy cannot know them, not to move responsibility.

  • Facts only the client holds: stock, price, delivery times, terms, and whether an offer is still live.
  • Claims that need evidence the client keeps, including anything a regulator would ask the client to substantiate.
  • Permissions: the right to the images, names, logos, music, footage and people in the work.
  • Consent for any list a message would be sent to. Troy never imports lists, and a send needs the client’s confirmation.
  • Whether the work should carry an AI label where it will run, which is the client’s decision under clause 6.

One of those is worth naming. Canada’s Anti-Spam Legislation turns on consent, on the sender identifying itself, and on a working unsubscribe in a commercial electronic message. The sender is the client, not Troy: Troy drafts, the draft lands in the client’s own tool, and nothing goes out without the client’s approval. Whether a particular message is caught by that law, and what it has to carry, is a question for the client and the client’s own advisers — Troy does not answer it and this paragraph is not an answer to it.

In plain EnglishRules about AI disclosure differ by market and are changing. This section is short on purpose.

Troy does not claim to be compliant with any of this. It claims to work in a way that leaves the client — who carries the obligation — able to comply.

One date is worth naming, as context rather than as an obligation Troy is under. The EU AI Act’s Article 50 transparency obligations apply from 2 August 2026 (European Commission, https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act). Troy’s markets are Canada and the United States. Troy does not offer its services to people in the European Union or the United Kingdom yet, so Troy does not treat Article 50 as applying to the work it does today — whether it reaches a particular campaign depends on where that campaign runs, which is the client’s question and not Troy’s to answer. It is named because it is the most worked-out AI-labelling rule anywhere, because it is where platform and client expectations are drifting, and because Troy would have to meet it before taking work aimed at the EU.

Beyond that, obligations differ by country, by platform and by the kind of content, and several regimes are still being written. Ad platforms set their own AI disclosure rules too, which can be stricter than the law and change without notice. Troy tracks what applies to the work it does and tells clients what it finds. It does not advise on the law, and a client who needs a view on their own position should take their own advice.

Where Troy is based is settled; the company itself is not, yet. Troy Agency Inc. is incorporating federally, under the Canada Business Corporations Act, with its registered office in Ontario, with its office at Unit 6, 10651 Keele Street, Maple, Ontario L6A 3Y9, Canada. The incorporation has not completed: the corporation number is To be completed: There is no corporation number yet, which means the federal incorporation has not completed. Until it has, Troy Agency Inc. is an intended name rather than a company: it cannot sign a contract, and calling itself “Inc.” in one would be wrong. This is the first thing to finish., and until there is one, nothing on this site should be read as saying the company already exists. None of that decides what a client’s advertising has to say.

So far as Troy is aware, no Canadian statute in force imposes an AI-content transparency duty of the kind Article 50 sets out; advertising here is caught instead by the general prohibition on false or misleading representations, and rules differ again from province to province and from state to state. Troy records what was AI-made and applies the client’s labelling decision in every market it works in, including the ones that do not ask for it. That is a description of practice, not a reading of the law, and it can be overtaken by a rule Troy has not seen.

Privacy law is a separate question with a clearer answer. Troy is subject to Canada’s federal private-sector privacy law, the Personal Information Protection and Electronic Documents Act, and the authority that oversees it is the Office of the Privacy Commissioner of Canada. People in the EU or the UK may also complain to their own national authority. That much follows from where Troy sits. The GDPR and the UK GDPR are named on this site as context, not as regimes Troy answers to today. In the European Union, none appointed. Troy does not offer its services to people in the European Union, and will appoint a representative there before it does. In the United Kingdom, none appointed. Troy does not offer its services to people in the United Kingdom, and will appoint a representative there before it does. The privacy policy is where that stands. None of those is an AI-labelling law, and naming them is not a claim to comply with any of them.

In plain EnglishThe Score is a published formula about an account, reviewed by a person. It makes no decision about any individual.

  • The Score grades a business’s marketing accounts. It produces a number out of 100 and a ranked list of fixes. It is free.
  • The formula is published and versioned. It is at v0.1 and marked a draft. When it changes, the change and the reason are published.
  • Every deduction carries the evidence behind it. No evidence, no deduction.
  • Today a strategist prepares every Score by hand from read-only access, and reviews it before the client sees it. No Score is sent unread, and that does not change when more of the reading is automated.
  • The Score does not predict results, and it will never claim to.
  • A fix the client dismisses with a reason stops costing points. The client knows things about their business that an audit does not.
  • The do-it-yourself instructions are complete and free. The diagnosis is never a paywall for the cure.

On automated decision-making: Troy makes no decision about an individual person by automated means, and nothing in the Score or the agents profiles, ranks or scores a client’s individual customers. Decisions about who Troy works with, what it charges and whether an engagement ends are made by people. Where a reader has rights about automated decision-making and profiling, the privacy policy explains them and where they depend on the reader’s country.

One rule is worth naming, because a Canadian client will ask about it. Quebec’s Law 25 gives a person rights where a decision about them is based exclusively on automated processing, and it may apply where a client or a client’s own customers are in Quebec. Whether it does is a question about the client’s processing rather than Troy’s, and Troy has not assessed it for anyone. It is mentioned here because the answer for Troy’s own tools is the short one: nothing in the Score or the agents makes a decision about an individual at all.

In plain EnglishAsk, and Troy tells you, for any piece of work.

  • A client may ask about any piece of work Troy made for them: the category, which tool was involved, which parts it did, who reviewed it, who approved it, and what the client decided about labelling. Troy answers in writing.
  • The answer comes from the provenance record described in clause 5.2, not from memory.
  • Anyone else may ask about work Troy has published in its own name, at troy@louin.io.
  • For a client’s own advertising, Troy passes the question to the advertiser rather than answering for them, and tells the person asking that it has done so.
  • If Troy cannot answer because a record is incomplete, it says so instead of guessing.
  • Asking costs nothing, is not a complaint, and has no effect on a client’s price or service.

In plain EnglishIt will change. Every change is dated in the table at the bottom of this page.

  • This is a draft at version 0.3. It is not in force and no lawyer has reviewed it.
  • It changes when Troy’s practice changes: a new kind of use, a category redrawn, a limit added or removed, a vendor rule settled.
  • A change that narrows what Troy will do takes effect when it is published. A change that widens what Troy may do with a client’s material is told to affected clients before it applies to their work.
  • Earlier versions are kept, so it is possible to see what Troy said at the time a piece of work was made.
  • Nothing here overrides a signed agreement. Where the two differ, the agreement governs.

Related reading: the Acceptable Use Policy for the work Troy refuses, the sub-processor list for the vendors behind the tools, the service terms for approvals and ownership, and the privacy policy for what happens to personal data.

Changes to this document
VersionDateWhat changed
0.320 Sep 2026The founder answered the open questions, and two of the answers change what this document says. Troy, Louin and Lunari Group share systems and data, so the provenance section now says where a client’s record is held and who can see it, and says that the basis for the sharing is not in place. Troy’s markets are Canada and the United States only, so the EU AI Act is now named as context rather than as an obligation Troy is under, and no EU or UK representative is claimed. The model vendors are named; what remains open is the plan each is used under. The 0.2 note below said the incorporation was settled — it is not: the federal incorporation has not completed and there is no corporation number yet. CASL and Ontario’s limitation period are named where they bear on the work. Still not reviewed by a lawyer.
0.220 Sep 2026Troy’s incorporation in Ontario settled, so the jurisdiction question in section 10 is answered and the Canadian position on AI labelling is stated. The open questions about model vendors and provenance marking moved out of hand-written boxes and onto the checklist at /legal. Still not reviewed by a lawyer.
0.119 Sep 2026First draft, written alongside the site. Not reviewed by a lawyer.