MEASUREABLE MEDIA LLC • TAILGATE BOOKS
Version 1.1, effective October 27, 2026 Policy Owner: AI Governance Lead (Section 12)
This Artificial Intelligence Governance Policy (the “Policy”) sets the principles, controls, and accountability that govern how Measureable Media LLC (the “Company”) selects, deploys, operates, monitors, and describes artificial intelligence and machine-learning systems (“AI Systems”) in Tailgate Books (the “Service”). It is meant to keep the Company’s use of AI lawful, accurate, secure, and consistent with the promises the Company makes to Contractors in the Tailgate Books Terms of Use (the “Terms”) and Privacy Policy. Capitalized terms not defined here, such as Contractor, Customer, Customer Data, and Your Content, have the meanings given in the Terms.
The Policy is written with the following laws and frameworks in mind (together, the “AI Laws”): Colorado's law on automated decision-making technology in consequential decisions (Senate Bill 26-189, effective January 1, 2027, which repealed the Colorado Artificial Intelligence Act, Colo. Rev. Stat. § 6-1-1701 et seq., after a federal court blocked its enforcement) and the Colorado Privacy Act’s profiling provisions; the California Consumer Privacy Act regulations on automated decision-making technology; the New York Responsible AI Safety and Education Act (RAISE Act); state laws requiring disclosure of AI use; and Section 5 of the FTC Act as the FTC applies it to claims about AI.
1.1 Current Applicability Assessment
As of this Policy’s most recent review, the Company has assessed that the AI Laws’ substantive obligations for high-risk or consequential-decision systems do not apply to the Company’s AI Systems, because:
the Company’s only AI System (receipt reading, Appendix A) reads a receipt image or PDF and suggests four bookkeeping values that the Contractor reviews before anything is saved;
no AI System makes, or is a substantial factor in making, any decision about a person, including any decision about employment, credit or lending, housing, insurance, healthcare, education, legal services, or essential government services;
no AI System scores or profiles any individual, and AI output is never shown to a Customer; and
the Company does not develop, train, or fine-tune AI models. It uses a third-party model on an inference-only basis, so laws aimed at model developers, such as the RAISE Act, do not reach it.
This Policy is therefore adopted as voluntary, forward-looking governance. The Company will reassess applicability under Section 4 before any AI System is added or changed, and at each annual review.
This Policy applies to:
every AI System used in the Service, whether operated by the Company or supplied by a third party and integrated into the Service;
the Company’s use of AI tools to build, test, and operate the Service, including AI coding assistants (Section 11.3); and
every owner, employee, contractor, and agent of the Company who designs, builds, operates, supports, or markets the Service (“Personnel”).
2.1 What Is Not an AI System of the Company
The following features are not AI Systems of the Company, and the Company will not describe them as AI in marketing, in the Service, or in its legal documents:
Estimate building. The Service turns a Contractor’s description of the work into estimate lines on the device and proposes prices by matching the description against the Contractor’s own saved rates. It uses no AI model and no market data.
Dictation. Speech-to-text is performed by the Contractor’s phone operating system or web browser, under that provider’s own terms. The Company receives only the resulting text.
Financial views and reports. The tax set-aside target, sales-tax tracking, mileage value, and profit-and-loss statement are computed by fixed code.
Logo colors. Picking colors from a Contractor’s logo runs on the Company’s own server using fixed code, with no model.
If any of these features is later rebuilt to use an AI model, it becomes an AI System and must go through Section 4 before release.
Lawfulness. AI Systems are operated in compliance with applicable law and with the provider’s usage policy.
Suggestion, not decision. AI output is a suggestion to the Contractor. Nothing an AI System produces is posted to a Contractor’s books until the Contractor reviews and approves it, and no AI output is shown to a Customer.
Minimum data. An AI System receives only the data its task needs. For receipt reading, that is the receipt image or PDF (or, for a forwarded email with no attachment, a PDF copy of its body text), with no identifier of the Contractor, the business, or any Customer.
No training. The Company does not use Your Content, Customer Data, AI inputs or outputs, Usage Data, or data derived from any of them (including de-identified or aggregated data) to train, fine-tune, or improve any AI model, as Sections 15.3, 15.4, 17.3, and 17.4 of the Terms promise.
Transparency. The Company tells Contractors where AI is used, what it sends to the provider, and how accurate it can be. It never claims AI is error-free or calls a feature AI when it isn’t.
Accountability. One named person owns AI governance, keeps the records this Policy requires, and signs off on every change (Section 12).
Security and privacy. AI Systems are operated under the Company’s Written Information Security Program (the “WISP”) and Incident Response Plan.
The AI Governance Lead maintains the AI System inventory in Appendix A. For each AI System, the inventory records its purpose, the provider and model, what data goes in and comes out, who reviews the output, where output is stored, and its risk classification.
Each AI System is classified as follows:
Low risk. Output is a suggestion that a Contractor reviews before use, affects only the Contractor’s own records, is never shown to a Customer, and makes no decision about any person.
Moderate risk. Output reaches a Customer or another third party, is used without item-by-item human review, generates text a Contractor may send under his own name, or processes Customer Data.
High risk. Output makes, or is a substantial factor in making, a decision with legal or similarly significant effect on a person, or the AI System profiles individuals. The Company will not deploy a high-risk AI System without first obtaining legal review and amending this Policy, the Terms, and the Privacy Policy.
Receipt reading is classified as low risk. Classification is reviewed whenever an AI System’s inputs, outputs, provider, model family, or use changes, and at each annual review.
Before releasing a new AI System, or making a material change to an existing one, the AI Governance Lead completes the checklist in Appendix C and keeps it with the inventory. A material change includes: sending the provider any data beyond what the inventory lists; showing AI output to a Customer; using AI output without the Contractor’s review; switching providers; or moving to a new model family. Routine model version updates within the same family are covered by Section 10.2.
For a moderate-risk or high-risk AI System, the assessment must also address foreseeable discrimination or disparate-impact risks, data quality and provenance, privacy, security and misuse risks, how much users are likely to rely on the output, and how a human can review or override it.
5.1 Legal Documents Come First
No new AI System, and no change that alters what data goes to an AI provider, may be released until the Terms and the Privacy Policy describe it. Section 11.1 of the Terms promises Contractors that the Company will update both documents before adding AI features, and a material change to either requires at least thirty (30) days’ notice under Section 25 of the Terms. The Apple App Privacy details and Google Play Data safety form must be updated at the same time.
The Company does not develop or train AI models. It uses a third-party model through a commercial API, on an inference-only basis. For every AI System:
Inputs are limited to the inventory. Receipt reading sends only the receipt image or PDF, as photographed or forwarded. If a forwarded email has no attachment, it sends a PDF copy of the email's body text, without formatting, images, or links, because many stores send the receipt as the email itself. It never sends the email's headers, sender or recipient addresses, or any identifier of the Contractor, the business, or a Customer. It never sends Customer records, rate lists, estimates, or photographs of the work. Very small images in a forwarded email, such as signature logos, are skipped and never sent.
Outputs are kept only as the Terms describe. The values the model returns (store, total, date, and category) are checked on the Company's servers before they are stored: only allowed categories are accepted, amounts and dates are validated, text is shortened to set lengths, and anything else is discarded. The checked values are stored with the receipt, together with the values the Contractor confirms and a record of which model read it, so a misread receipt can be understood later. The provider’s full response is not stored or logged, including in server-function logs and error reports. Stored values are part of Your Content and are deleted with it.
No training or evaluation on Contractor data. Your Content, including receipts and AI inputs and outputs, is never used to train, fine-tune, evaluate, or tune the prompts of any AI model. Test sets used to check a model or prompt (Section 10.2) are built only from receipts the Company owns or creates for that purpose.
Usage Data and de-identified data. Usage Data, and de-identified or aggregated data created from Usage Data or Your Content, are used only as Sections 15.4 and 17.5 of the Terms and Sections 2.3 and 3.5 of the Privacy Policy allow, and never to train, fine-tune, or improve any AI model. That limit applies whether the model is the provider’s, an open-source model, or one the Company might build. Any benchmark or other figure reported from aggregated data, including typical prices for a type of work, must be drawn from at least ten (10) Contractors, and no single Contractor’s data may make up most of the figure (Section 15.4 of the Terms). The AI Governance Lead confirms both conditions before a figure is reported, including internally.
Prohibited information. The Service reminds Contractors not to photograph or forward receipts showing information Section 9.3 of the Terms prohibits. If the Company learns such information reached the provider, it follows Section 10.3.
Keys. The provider API key is held only in the hosting provider’s secret store, is used only by server functions, and never appears in the app, the source code, or logs.
Receipt reading makes no decision about any person, so classic fairness testing (outcome disparities across protected groups) does not apply to it. The comparable risk is that the reader works less well for some Contractors’ receipts than others. Testing is therefore proportionate and practical:
Before a new model or prompt goes live, the Company runs the Section 10.2 test set, which includes printed, handwritten, faded, crumpled, and photographed-at-an-angle receipts; receipts from small local vendors as well as national chains; multi-page and long receipts; and Spanish-language receipts, because the Service is offered in Spanish.
The Company compares accuracy for each field (vendor, total, date, category) across those receipt types, and does not release a change that makes any type materially worse without a documented reason.
Any moderate-risk or high-risk AI System added later must also receive the statistical fairness, error-rate, and counterfactual testing appropriate to its use, documented in its Appendix C assessment.
Every value the receipt reader suggests appears in an editable form. Nothing is posted to the Contractor’s books until the Contractor reviews and confirms it.
AI output is never sent or shown to a Customer and never appears on an estimate, invoice, change order, or Document page.
No AI System may take an action on its own, such as sending a message, recording a payment, or changing a record, without the Contractor’s confirmation.
A Contractor can always enter or correct a receipt by hand without using AI.
9.1 What Contractors Are Told
The Terms (Section 11) and the Privacy Policy (Section 6) tell Contractors where AI is used, which provider powers it, what is sent to the provider, that the provider does not train on it and keeps it only for a limited period, that AI output can be wrong, and that the Contractor must review it. The AI Governance Lead confirms at each review that those descriptions still match the Service.
9.2 In the Service
The receipt form labels each AI reading: “Read by AI. Please check it, then tap anything to fix it.”
9.3 Marketing and Public Statements
Marketing materials, app store listings, the website, and support replies must not:
describe estimate building, dictation, tax figures, or reports as AI, or suggest AI does more than it does;
claim AI output is accurate, error-free, unbiased, or a substitute for the Contractor’s review, an accountant, or tax advice; or
state any accuracy figure the Company cannot support with its own records.
The AI Governance Lead reviews new marketing copy that mentions AI before it is published.
10.1 Ongoing Monitoring
At least quarterly, the AI Governance Lead reviews:
the correction rate for each field, measured as the share of receipts where the Contractor changed a value the reader suggested, using aggregated counts only;
receipt-reading errors and failures in the error-reporting tool; and
support emails about misread receipts.
This monitoring checks quality. It never feeds data into any model. The aggregated counts may be used to decide whether a model or prompt needs to change, but any change is built and checked only against the Company’s own test set (Section 10.2).
10.2 Model and Prompt Changes
The Company pins the model version it uses. Before moving to a new model version or changing the prompt, including when the provider retires a model, the Company runs the test set described in Section 7, compares results with the current version, and records the result and the decision in the inventory. Section 11.7 of the Terms allows the Company to change models; this Section makes sure a change doesn’t quietly make the Service worse.
10.3 AI Incidents
Each of the following is an AI incident, handled under the Incident Response Plan:
data beyond what the inventory lists reached the provider, including a Contractor identifier, email contents other than the body-text PDF of a forwarded email with no attachment, or Customer Data, or any AI tool used by Personnel was given access to production data or production keys, or used that access, in a way Section 11.3 does not permit;
a security breach at the provider affecting data the Company sent;
a sustained or systematic misreading, such as one field wrong across many Contractors after a model change;
AI output reaching a Customer or being posted without the Contractor’s confirmation; or
an attempt, through a forwarded receipt or otherwise, to manipulate the model into returning anything other than receipt values.
Where the Terms require notice of a security incident to Contractors (Section 15.6), the Incident Response Plan’s timelines apply. The AI Governance Lead may pause receipt reading at any time; Contractors can still enter receipts by hand.
10.4 Misuse Safeguards
Because anyone who learns a Contractor's private forwarding address can send mail to it, and the Service does not filter forwarded mail for spam, the Service treats every receipt as untrusted input. Forwarding is limited to fifty (50) receipts per business per day and ten (10) per email, and receipt reading runs only for accounts on a free trial or a paid subscription. The model is asked only for four values, its response is checked on the server against the expected format before it is stored or shown, anything outside that format is discarded, and model output is never executed or used to call other tools.
Using a third-party model does not relieve the Company of responsibility for how the Service uses it. The Company cannot obtain contractual bias-testing or audit rights from AI providers at its current scale, so it relies on diligence proportionate to risk.
11.1 Provider Diligence
Before engaging an AI provider, and at least annually (quarterly for the terms in Section 11.2), the AI Governance Lead:
reviews the provider’s commercial terms, data processing addendum, usage policy, and published documentation on training, retention, security, and model behavior;
confirms the terms support every statement the Terms and Privacy Policy make about the provider, including no training on the Company’s data, limited retention, and where data is processed;
files dated copies of those documents, as retrieved, in the vendor record in Appendix B; and
records any gap, escalates it to counsel, and, where appropriate, raises it with the provider.
11.2 Current Provider: Anthropic
Receipt reading uses Anthropic’s models through its commercial API, under the Anthropic Commercial Terms of Service accepted by Measureable Media LLC, which incorporate Anthropic’s Data Processing Addendum (effective February 24, 2025). The Company has no zero-data-retention arrangement. Under those terms, Anthropic acts as the Company’s processor, does not train its models on the data the Company sends, and keeps that data only for a limited period (longer only where needed to enforce its usage policies). Its terms do not limit processing to the United States. The Company must comply with Anthropic’s usage policy as it changes, and will review the Commercial Terms, DPA, and usage policy each quarter. If Anthropic changes its training, retention, or location terms, the AI Governance Lead notifies counsel before the next release, because the Terms and Privacy Policy may need to change.
11.3 AI Tools Used by Personnel
Personnel may use AI coding assistants and other AI tools listed in Appendix A, Part 2, to build and operate the Service, subject to these rules:
Claude: test data only. Claude (Appendix A, Part 2) works only under the owner’s direction, in sessions he starts, on his Claude Max plan with Anthropic’s “Help improve our AI models” setting turned off, against a local copy of the database that holds only test data. No AI tool may be given access to the production database, production file storage, or production keys. Claude prepares code and database migrations, and the owner reviews each one and applies every database change and every deploy to production himself (WISP Section 11). No Your Content, Customer Data, production database extracts, support emails from Contractors or Customers, or secrets (API keys, credentials, tokens) may be entered into any AI tool, other than the production AI System itself. This rule applies no later than the day the first private-beta Contractor enters real business records. Until then, while the production database holds only test accounts, Claude may reach it under the owner’s supervision; the owner approves every database change before it is applied, and every access is entered in the admin access log (WISP Section 8).
Any later production access. Before any AI tool is given access to production data or production keys, the owner tells counsel. That access requires an Anthropic business plan governed by Anthropic’s Commercial Terms, and the Terms, the Privacy Policy, this Policy, and the WISP must be revised to disclose it before it begins.
AI tools are used only under terms that prohibit the provider from training on the Company's inputs and outputs, such as Anthropic's Commercial Terms. A tool offered only on consumer terms may be used with training turned off, and never with any data or access this Section prohibits.
AI-generated code is reviewed by a person before release and goes through the same testing as any other code.
Personnel record which parts of the codebase were substantially AI-generated, because copyright protection for code can depend on how much of it a person wrote.
The Company’s AI Governance Lead is Michael Cox, Owner, or another person the Company designates in writing. The AI Governance Lead:
maintains Appendices A through C and the records this Policy requires;
approves every new AI System and material change under Section 5;
performs the reviews in Sections 9, 10, and 11;
makes sure anyone who works on AI features or support reads this Policy before starting, and again after any change to it; and
brings to counsel any proposed high-risk use, any change in the AI Laws that may affect the Company, and any AI incident.
Records are kept for at least three (3) years after the AI System or version they concern is retired.
The Company will reasonably cooperate with lawful regulatory inquiries about its AI Systems. The AI Governance Lead reviews this Policy at least once a year and before any AI System is added or materially changed, and updates it to reflect changes in law, technology, providers, and the Service. Each review is recorded in the Revision History below.
AI Systems are probabilistic. This Policy does not require error-free or bias-free output. It requires reasonable, good-faith, risk-based governance and the Contractor review the Service is built around. This Policy is for internal governance and for responding to diligence requests. It does not amend the Terms or the Privacy Policy, and it creates no rights for any third party.
Part 1: AI Systems in the Service
| ITEM | RECEIPT READING (AI-01) |
|---|---|
| Purpose | Reads a receipt image or PDF and suggests store, total, date, and expense category for the Contractor’s expense records. |
| Where used | Receipt photographed or chosen in the app; receipt forwarded by email to the Contractor’s forwarding address. |
| Provider and model | Anthropic, commercial API. Model version: claude-haiku-5-5. Date adopted: November 14, 2026. |
| Model changes | November 14, 2026: claude-haiku-4-5-20251001 to claude-haiku-5-5. The Section 7 test set (50 receipts the Company owns, covering every type Section 7 names, each read three times) was run on both versions. Totals right: 81% against 50%. Totals wrong: 4% against 13%. Dates right: 82% against 47%. Decision: adopt claude-haiku-5-5. Full results are kept with the inventory records. |
| Inputs sent | The receipt image or PDF only, as photographed or forwarded, without redaction; for a forwarded email with no attachment, a PDF copy of its body text. No headers, addresses, or Contractor, business, or Customer identifier. Very small images skipped. |
| Outputs | Four suggested values, shown in an editable form. |
| Human review | The Contractor confirms or corrects every value before anything is saved. Output never reaches a Customer. |
| Stored | The checked returned values, the confirmed values, and the model name, with the receipt, as part of Your Content. The full provider response is not stored. |
| Decision about people | None. |
| Risk classification | Low (Section 4). Last reviewed: November 14, 2026, when the model moved to a new generation in the same family. Still low risk: the inputs, outputs, Contractor review, and use did not change. |
| Disclosures | Terms Section 11; Privacy Policy Sections 2.2 and 6; in-app label “Read by AI. Please check it, then tap anything to fix it.” (Section 9.2). |
Part 2: AI Tools Used by Personnel (Section 11.3)
| TOOL | USE | TRAINING ON INPUTS TURNED OFF? |
|---|---|---|
| Claude (Anthropic), including Claude Code | Code generation and review, running checks, and preparing code and database migrations, against a local copy of the database with test data only; no access to production data or production keys; the owner applies every change and deploy himself (Section 11.3) | Claude Max plan, under Anthropic’s consumer terms, with “Help improve our AI models” turned off |
| ITEM | ANTHROPIC |
|---|---|
| Agreement | Anthropic Commercial Terms of Service, accepted by Measureable Media LLC, incorporating the Anthropic Data Processing Addendum (effective February 24, 2025). No zero-data-retention arrangement. |
| Role | Processor for the Company; subprocessor as to Contractors’ data. |
| Training | No training on the Company’s inputs or outputs (Commercial Terms). |
| Retention | Limited period under its terms; longer only where needed to enforce its usage policies (DPA § H.1.b(iii)). Deletion within 30 days after the agreement ends. |
| Location | Not limited to the United States. |
| Breach notice to the Company | Within 48 hours after Anthropic becomes aware (DPA § G.1). |
| Subprocessor changes | Reasonable notice; 15 days to object (DPA § C.3). |
| Documents on file | Commercial Terms, DPA, and Usage Policy as retrieved on [dates]. |
| Last quarterly review | [date and result] |
Complete before release, sign, date, and file with Appendix A.
Describe the feature, the provider and model, and exactly what data goes in and comes out.
Classify it as low, moderate, or high risk (Section 4). If it is high risk, stop and get legal review.
Confirm it makes no decision about any person, or document why it does and what review exists.
Confirm no Customer Data goes to the provider, or get legal review of what goes and why.
Confirm the provider’s terms prohibit training on the data and state retention and location; file copies (Appendix B).
Run the accuracy test set (Sections 7 and 10.2) and record the results.
Confirm the Contractor reviews every output, and that no output reaches a Customer without that review.
Confirm the Terms, Privacy Policy, subprocessor page, and app store privacy forms are updated, with thirty (30) days’ notice where required.
Confirm the in-app label and any marketing copy follow Section 9.
Update Appendix A and record the approval below.
Approved by: ______________________ Date: ______________