AI Use and Data Policy

Part 1. AI Use and Data Policy

1. Summary

Pitchlume uses artificial intelligence ("AI") to turn a manufacturer's catalog material (drawings, data sheets, CAD files) into a visual, configurable catalog, and to help people write, present and follow up. We use AI with five rules:

1. US providers on paid terms that do not train on your data.
2. Your files go only to the providers we chose, or nowhere.
3. A person confirms before anything is published.
4. Everything AI makes is labeled, and never passed off as an engineering value.
5. We never train AI models on your content.

2. Where AI is used

• Catalog engine: reads drawings and data sheets, extracts dimensions and attributes with their source, builds 3D models and generated drawings, and checks its own work.
• Assistants: the assistant panel, the viewer copilot, case and catalog chat, and the campaign assistant.
• Content: decks, print pieces, documents, email drafts, renders, images and marketing video.
• Meetings: transcription and summaries of recordings a user uploads (the uploader confirms that everyone agreed to be recorded).
• Research: web search on public sources, for example trade show details.

Pitchlume does not use AI to make decisions about people (hiring, credit, housing or similar), and it does not email your customers or prospects on its own. People send that email.

3. Providers

• We use US-based AI providers on paid commercial terms under which the provider does not train its models on our data, running in US regions.
• Today these are OpenAI, Anthropic and Google Cloud Vertex AI, each through our own direct account. Marketing video uses Runway, which receives only the prompts and images a person chooses for a video. The current list, with purposes and locations, is on the Subprocessor List.
• We ask each provider for zero data retention where it offers it. Where it is not yet confirmed for our account, the provider's standard API terms apply, which may keep inputs for a limited period (up to 30 days under the terms we have seen) for abuse monitoring, not for training.
• We do not use providers or hosts based outside the United States.
• Our assistants do not name the model behind a given answer. We choose models per task and change them as better ones arrive, always within the providers on our Subprocessor List.

4. How requests are routed

• Your files take the strictest route. Drawings, CAD files, extracted values and anything else from your uploads go only to providers we hold direct accounts with, in US regions, with no-training terms.
• Fail closed. If a request for your files cannot be sent on a route that meets these requirements, it fails. It never falls back to a route that does not.
• No substitutions we did not choose. A fallback, when there is one, is a model we selected in advance. No intermediary may swap in a different model or vendor.
• Requests without your files (for example chat about public information, or research from public sources) may use a US gateway (Vercel AI Gateway) with zero-retention and no-training settings, where your workspace's AI pathway is set to it. Your files are never sent through it.
• One workspace at a time. Each AI request carries only one workspace's data.
• Each AI request is recorded with the provider, model and route, so we can check that the rules held.

5. A person confirms what is published

Uploaded and AI-generated content goes live in your catalog, and is shown to guests, only after a person in your workspace confirms it. That person may be an engineer, an administrator or a marketing manager, as your workspace's roles allow. There is no automatic approval. Pitchlume records who confirmed and when. Whoever publishes takes responsibility for that content within your company.

6. Labels, disclaimers and provenance

• Representations, not engineering documents. Generated 3D models, drawings, renders and modified designs are representations. They carry a disclaimer such as: "Illustrative model. Refer to the manufacturer's sealed drawing for dimensions and ratings." The manufacturer's sealed drawing remains the engineering document of record.
• Preliminary is labeled preliminary. Modified or concept designs carry: "Preliminary sketch for engineering review. Not for construction. Values marked preliminary are not approved."
• We never invent engineering values. Where a value is missing, estimated or unverified, Pitchlume still shows the model, with a clear warning such as "Incomplete data: illustrative only." A guessed or default value is never shown as a real one.
• Fidelity labels tell you how far a 3D model can be trusted (for example CAD-exact, From the drawing, Representative, Concept).
• Provenance. Where we can, each extracted value keeps a record of where it came from (which file, which page), so a person can check it.
• Originals are never changed. Uploaded files are stored as they were received, with a SHA-256 hash. Split pages and renders are separate, derived copies.
• Exports carry the label. Exported files carry the disclaimer, any preliminary label and an embedded notice naming the workspace, the time and an export ID; the audit trail records who exported. See the Engineering Content Disclaimer and Export Terms.
• Generated videos carry the workspace watermark but no on-screen disclaimer. They are marketing pieces, and the rule in Section 12 applies to them as to everything else.

7. Your rating tables and formulas

Rating tables and calculation methods are your intellectual property. They stay on our servers. AI assistants can ask for one answer at a time (for example, the rating for one size) through a server function; they cannot read or reveal the whole table or formula, and they are instructed never to.

8. Safety rules for our assistants

• Content from files, web pages, emails, transcripts and tool results is treated as data, never as instructions, which limits prompt-injection attacks.
• Assistants never reveal internal prompts, keys, other workspaces' data, or rating tables and formulas.
• Assistants take actions only through permission-checked tools, and ask for confirmation before anything outward-facing or irreversible.
• Replies are checked for secrets before you see them.
• We test these rules against a set of attack prompts.

These measures reduce risk; they do not eliminate it.

9. Your controls

• Workspace roles decide who may upload, who may consent when an upload is declared or detected as restricted, and who may publish.
• Feature switches: features are switched per workspace. EraShift sets them, and workspace admins can adjust the ones their plan allows.
• Budgets: workspace admins can set AI credit budgets for people and agents.
• Downloads are off for non-admin users until a workspace admin turns them on.
• Removal: you can delete content you uploaded or generated. "Report a misclassified upload" in the app removes a file and what was derived from it (extracted values, models, renders, cached research and search entries) and tells your workspace's admins. You can also write to help@pitchlume.ai.

10. Restricted and regulated data

Pitchlume is not designed for restricted or regulated data, and we do not encourage or allow it. That includes ITAR technical data, export-controlled technology with an ECCN, Controlled Unclassified Information and similar categories. Our AI providers' own terms prohibit export-licensed and ITAR input.

• We ask first. Before files are processed, the person uploading them declares whether any file contains restricted or regulated data.
• Code checks before any AI call. Our own code (not an AI model) checks file names, text and title blocks for restricted-data markings. If it finds any, the upload pauses and nothing is sent to an AI model.
• Consent is the uploader's, on the customer's behalf. On "Yes", "Not sure" or a detected marking, the app shows how files are processed and by which services. A person whose role allows it either declines the files (they are removed and never processed) or accepts responsibility and continues. That choice is the customer's own decision and risk. It is not our approval and it does not determine export classification.
• Everything is recorded: who decided, when, which files, and which version of the wording they saw.

See the Restricted Data Policy.

11. Who owns AI output

• Your content stays yours. You own your Customer Content (your catalog material, drawings, CAD files, rating tables and calculation methods).
• Output generated from your content (generated 3D, drawings, renders, decks, text) belongs to you, subject to EraShift's rights in the underlying platform, engine, generators and templates. The exact terms are in the Terms of Service and in Part 2 below.
• No training, no sharing. We never use your content to train AI models, and we never use one customer's content for another customer.
• We may use aggregated, de-identified usage data (for example, how many pages a run read and how long it took) to run and improve the Service.
• AI output may not qualify for copyright protection, and similar output may be generated for others from their own content. We assign you whatever rights we have in output made from your content; we do not promise that it can be protected.

12. Limits of AI

AI can be wrong. It can misread a drawing, miss a note, or produce something that looks right and is not. That is why Pitchlume labels its output, keeps the sealed original as the document of record, and requires a person to confirm before publishing. Do not use generated content for engineering, fabrication, installation or safety decisions. Use the manufacturer's sealed drawing and qualified engineering judgment.

13. Where we are today

We want this policy to be accurate, so here is what is not finished as of the version date of this document:

• Zero data retention is requested from our AI providers and is not yet confirmed in writing for every account. Standard API retention applies until then (Section 3).
• The Gemini API through Google AI Studio is not limited to US regions. We use it only with public or test data, never with your files or workspace data. Gemini models for your data run on Google Cloud Vertex AI in the US.
• The strict routing rule (direct providers only, fail closed) is enforced in the catalog engine today. We are applying the same routing check to every other feature that sends your files to an AI model (decks, documents, meeting transcription, renders). Those features use our direct provider accounts today.
• Runway's processing region is not confirmed as US-only in our contract. It is used only for marketing video and receives only what a person chooses for that video.
• Automatic deletion of AI run records after the retention period (12 months by default) is not switched on yet.

14. Contact

Questions about how we use AI: info@pitchlume.ai or help@pitchlume.ai.

Change log

Part 2. AI Terms

These AI Terms (the "Addendum") are part of the Master Subscription Agreement at pitchlume.ai/legal/terms (the "Agreement") between EraShift Studio LLC ("EraShift") and Customer, and apply to Customer's use of AI Features. Part 1 of this document describes how AI is used; these terms are the contract. If this Addendum conflicts with the Agreement on its subject, this Addendum controls, except that Agreement §5 (Restricted Data), the Restricted Data Policy and the Engineering Content Disclaimer and Export Terms control on their subjects. Capitalized terms not defined here have the meanings in the Agreement.

1. Definitions

1.1 "AI Provider" means a third-party provider of AI models that EraShift uses to deliver AI Features, listed on the Subprocessor List.

1.2 "Input" means Customer Content and other data that Customer, its Authorized Users or the Service on their instruction send to AI Features, including files, prompts, instructions and context.

1.3 "Agent" means an AI Feature that carries out a multi-step task, such as building a catalog from drawings, drafting a deck or campaign, or researching an account.

2. AI Providers and data handling

2.1 US providers, commercial terms. EraShift uses AI Providers that are based in the United States, under paid commercial terms that do not permit the AI Provider to train its models on Input or Output, and in US regions.

2.2 Drawings and customer files. Drawings, CAD and other files Customer uploads are sent only to AI Providers through EraShift's direct accounts with those providers (currently OpenAI, Google Cloud Vertex AI and Anthropic), with zero data retention where the provider offers it. If that route is not available, the call does not go elsewhere; it fails.

2.3 Non-sensitive calls. EraShift may route calls that do not contain Customer's files (for example chat, research from public sources, or metadata) through an AI gateway service in the US, with zero data retention and no-training settings, as listed on the Subprocessor List.

2.4 No substitution. EraShift will not let a model or provider that it has not chosen and listed as a Subprocessor process Input as a fallback.

2.5 Provider retention. Where zero data retention is not available, an AI Provider may keep Input and Output for a limited period under its terms (for example for abuse monitoring), as described in the Subprocessor list.

2.6 Current state. As of the version date of this document: (a) zero data retention is requested from the AI Providers and is not yet confirmed in writing for every account, so Section 2.5 applies; (b) the video provider on the Subprocessor List receives only the prompts and images a person chooses for a marketing video, and its US-only processing region is not confirmed by contract; and (c) a Gemini API route that is not limited to US regions is used only with public or test data and never receives Input.

2.7 No training by EraShift. EraShift does not use Input or Output to train or fine-tune any AI model, and does not use one customer's Input or Output for another customer (Agreement §13.7).

2.8 Workspace scope. Each AI job runs with access only to the Workspace it was started in.

3. A person confirms what is published

3.1 No auto-publish. Uploaded and AI-generated content goes live in Customer's catalog, portal, share pages or other published surfaces only after a person in Customer's Workspace confirms it. The Service records who confirmed it and when. There is no setting that publishes AI Output automatically.

3.2 Responsibility. The person who publishes acts for Customer. Customer is responsible for what it publishes, including its accuracy, and for its use with Guests and customers.

3.3 Agents propose; people decide. Agents may draft, propose and prepare changes; publishing follows Section 3.1, and email is sent only by a person. AI Features never send email to Customer's customers or prospects automatically.

4. Labels

4.1 Output that represents a product carries the labels described in the Engineering Content Disclaimer and Export Terms, including the disclaimer, "Preliminary" and "Incomplete data" labels.

4.2 AI-enhanced images and other generated media are labeled as generated in the Service (for example "AI-enhanced photograph made from the 3D model").

4.3 Customer will not remove or alter these labels, and will not present Output as made without AI where that would mislead.

5. Customer's review of Output

5.1 Review before use. AI Features can produce Output that is inaccurate, incomplete, inconsistent with the Sealed Original, out of date, or offensive, and that may look confident when it is wrong. Customer is responsible for reviewing Output for accuracy and suitability before relying on it, publishing it or sharing it.

5.2 Values. Values that AI Features extract from drawings or documents are unverified until a person confirms them against a source. They are not quality records and are not engineering documents.

5.3 Research. Output based on web or public-source research may be wrong or out of date, and may come from third-party sites EraShift does not control.

6. No guarantee of accuracy

6.1 AI FEATURES AND OUTPUT ARE PROVIDED WITHOUT ANY WARRANTY OF ACCURACY, COMPLETENESS, RELIABILITY OR FITNESS FOR A PARTICULAR PURPOSE, in addition to the disclaimers in Agreement §16. Agreement §15.2(a) does not apply to Output.

7. Credits

7.1 Consumption. AI Features consume Credits as described in Agreement §8.2, at the rates EraShift shows to Workspace admins in the Service. Some Agents charge a fixed number of Credits per unit of work (for example per deck, per drawing page or per image); others are metered by the computing cost of the work.

7.2 Estimates and confirmation. Where the Service shows an estimate before a run, the estimate is a guide and actual consumption may differ. The Service may ask for confirmation before a run estimated above a threshold shown in the Service.

7.3 Budgets and stops. Workspace admins can set budgets per person and per Agent. When the balance or a budget is used up, new AI work stops; work already running may finish (Agreement §8.2(e)).

7.4 Failed runs. If a run fails because of an error in the Service, EraShift will, on request to help@pitchlume.ai within 30 days, restore the Credits it consumed. Credits are not restored for runs that completed but did not produce what Customer wanted, or that Customer canceled.

8. Ownership of Output

8.1 As between the parties, and subject to EraShift's rights in the Service, Customer owns Output generated from Customer Content (Agreement §13.3).

8.2 Output may be similar to output generated for others, and EraShift does not warrant that Output is unique or protectable by copyright or other intellectual property rights.

8.3 EraShift's indemnity for IP Claims (Agreement §17.1) does not cover Output to the extent derived from Customer Content, prompts or instructions.

9. Prohibited uses of AI Features

Customer will not, and will not let anyone, use AI Features to:

(a) process Restricted Data (Agreement §5 and the Restricted Data Policy);

(b) create documents that appear to be sealed, signed or certified by an engineer or other licensed professional, or that imitate a seal, stamp or license number;

(c) present Output as an engineering document, test result, certification or quality record;

(d) make decisions that produce legal or similarly significant effects on individuals (for example employment, credit, housing or insurance decisions);

(e) generate content that is unlawful, defamatory, deceptive, infringing, harassing or sexually explicit, or that impersonates a person or business;

(f) create or train a competing AI model or product, or extract model weights, prompts or system instructions;

(g) circumvent safety measures, rate limits, Credit metering, labels or the human confirmation step;

(h) send unsolicited bulk messages, or contact people without a lawful basis; or

(i) violate the usage policies of the AI Providers, which EraShift will make available on request and which apply to Customer through this Addendum.

10. Meetings and recordings

10.1 Before uploading a recording or using transcription, the uploader confirms that everyone recorded agreed to the recording. Customer is responsible for that consent and for recording laws (Agreement §4.4).

11. Changes to models and AI Features

11.1 Models change. AI models and providers change often. EraShift may change, add, upgrade or retire the models and AI Providers used for AI Features to keep them current, secure and cost-effective. A new AI Provider is added only with notice as a new Subprocessor under the DPA (Section 5).

11.2 Effect. A model change may change the style, quality, speed or Credit consumption of Output. EraShift will not make a change that materially reduces the core functionality of the AI Features Customer has paid for during a Subscription Term, except under Agreement §2.3.

11.3 Provider withdrawal. If an AI Provider withdraws or restricts a capability, EraShift may change or withdraw the affected AI Feature under Agreement §2.3.

12. AI laws

12.1 Each party will comply with laws on AI that apply to it. If a new law requires a change to AI Features, labels or this Addendum, EraShift may make it under Agreement §26.4.

Change log