How SeedBasis uses AI
What AI generates in SeedBasis, which models do the work and through whom, how findings are checked, where reports can go wrong, and what happens to your data.
Last updated 23 October 2026 · Markdown version
Effective 28 September 2026. SeedBasis is an AI system: it uses large language models to research your idea, or a company you run, and to answer your questions about the results. This page explains what that means for what you read.
The short version
- Reports and the agent's answers are generated by AI and checked by automated rules. No person reviews them before you see them.
- Every finding shows its source, how reliable that kind of source is, and a confidence level.
- AI can occasionally make mistakes, even when a finding has passed our checks. Treat a report as a starting point: check what matters against the sources, do your own research, and ask a qualified professional before you act on legal, financial or tax questions.
- We do not train AI models on your data, and we only use AI providers whose terms do not allow them to train on it.
1. What AI generates, and what it does not
| In SeedBasis | Made by | How you can tell |
|---|---|---|
| Findings, summaries, charts, tables and plans in a report | AI models working from public sources we collect, then automated checks | Everything in a report is AI-generated research unless it is labelled as edited by you |
| Quotes that support a finding | Copied from the source, and kept only if they appear there word for word | Each quote links to its source |
| Numbers in charts | Taken from findings, never typed in by a model | Chart values lead to the findings behind them |
| Numbers in a section's text | Written by an AI model, and kept only if the findings a sentence cites carry them, or they are plain arithmetic on those numbers | Each sentence links to the findings it rests on |
| Company logos | Copied from each company's own website and resized, never generated | They sit beside the company's name |
| Your onboarding answers, notes and edits | You | Edits are labelled as yours, with the researched original kept beside them and in the archive |
| The agent's replies | An AI model reading your reports | They appear in your conversation with the agent |
| Changes the agent makes to a page | An AI model | Marked as the agent's, listed in the answer that made them, and undone in one click |
| A new report, or deleting a page of yours | An AI model proposes it | Nothing happens until you confirm it |
2. Which companies run the models
| Job | Run by |
|---|---|
| Turning your answers into a research brief | Fireworks AI or Together AI |
| Deciding how deep to go in each area | |
| Planning the research in each area | |
| Pulling the passages that matter out of long sources | |
| Writing findings from the sources | |
| Checking that each finding's quotes really support it | Fireworks AI or Together AI |
| Drawing charts, diagrams and tables from the findings | |
| Writing each section around its findings and figures, the plans and each group's page | Fireworks AI or Together AI, or Google when that answer cannot be used |
| Writing the report summary and noting where sections disagree | Google, or Fireworks AI or Together AI when that answer cannot be used |
| Checking plans for generic advice that could apply to any idea | |
| Checking the claims a report's answer rests on against web pages read that day, before its decision pages are written | Anthropic, or Google when Anthropic cannot answer |
| Rewriting the sentences that stated a claim that check corrected | |
| Reading the finished report once, after every page is written, for places that disagree or state what their findings do not give, and rewriting those places | Anthropic, or Google when Anthropic cannot answer, and Google for a report about a company you run that carries your figures or documents |
| Checking each of those rewrites against the side the evidence favours | |
| SeedBasis Agent's answers to your questions | Google, or Fireworks AI or Together AI when Google is too busy to answer |
| SeedBasis Agent's answers when you talk with a report | |
| Turning what you say into text when you talk with a report | ElevenLabs, or Google when ElevenLabs cannot |
| Speaking the agent's answers when you talk with a report | ElevenLabs, or OpenAI when ElevenLabs cannot |
| Finding public posts on X for some parts of a report | xAI |
| Asking which products AI assistants recommend for your kind of product, for the Channels section | OpenAI, Anthropic and Google, each with its own web search |
We reach every model through OpenRouter, Inc., and each request names the only companies that may run it, all United States companies: Google LLC; xAI (SpaceXAI LLC); Anthropic, PBC; OpenAI; ElevenLabs (Eleven Labs Inc.); and, for an open model made by Z.ai, Fireworks AI, Inc. and Together Computer, Inc. Z.ai receives nothing. xAI's model only finds posts: it returns links, each post's text is read from X itself, and nothing it writes appears in a report. Once a report, models at OpenAI, Anthropic and Google answer one question: which products they recommend for your kind of product. They receive only those few words, and their answers are quoted as what AI assistants say, never taken as fact. Anthropic's model also runs once a report to check the claims its answer rests on: it reads the research brief's description of the idea, the claims with their quotes and passages of the pages our searches reached that day, and says whether each page confirms a claim, says otherwise or is newer; code holds every quote it gives to its page. When you talk with a report, ElevenLabs hears you and speaks the agent's answers, with Google's model to hear and OpenAI's to speak when it cannot. Which model does each job changes as we measure quality and cost; we update this page before a new company receives your data. The subprocessors page lists these providers.
3. How findings are checked
- A finding is one specific claim with its evidence: the exact quote, the source, the source's tier from T1 (a primary statement by the responsible party) to T4 (community signal), and a confidence from 0 to 1.
- A quote must appear word for word in the source it cites, or the finding is dropped.
- A second model, from a different family than the one that wrote the finding, checks that the quotes support the claim. A finding they do not support is dropped, and one they only partly support has its confidence lowered.
- A number stated as a fact must appear in the quotes it rests on, allowing for rounding, or the finding is dropped. Estimates show their arithmetic instead.
- Confidence is capped by the weakest source a claim rests on. A number from a company's own marketing page never reads above 0.6, and high confidence needs two independent sources.
- Numbers in charts come from findings, never typed in by a model.
- Plan steps that could apply to any idea are removed, and every step must cite findings.
- Before the decision pages are written, the claims the answer rests on are checked again against web pages read that day by a model of another family. A claim a page contradicts or updates is kept, marked weak and labelled with whose page says otherwise and when, what the page says is added beside it, and the sentences that stated the old claim are rewritten.
- Claims about you must rest on an answer you gave during onboarding.
- Each area shows how many sources it reached and which ones failed, so gaps are visible.
The full rules are in our method.
4. Where reports can go wrong
- AI models can occasionally state things that are not true. The checks in section 3 catch many of these mistakes, not all of them, so a finding can pass every check and still be wrong.
- Sources can be wrong or out of date. A finding is only as good as the page or record behind it, and a report reflects its sources on the day it was researched.
- Some sources cannot be reached. Paywalled, blocked or unavailable sources are left out, so an area can look thinner than the real picture. Reports show what failed.
- A quote can be accurate and still misleading. The word-for-word check proves a quote exists in the source, not that it was read in the right context.
- Estimates are reasoning, not facts. They are labelled and show their working, but their inputs can be wrong.
- A company seen from outside is partly estimated. In a report about a company you run, numbers about other companies, such as a private competitor's revenue, are worked out from public data by at least two methods and shown as ranges with those methods; where the methods disagree by more than three times, the report says the number cannot be estimated reliably. The research cannot interview your customers or see inside another company, and it says what it could not find.
- Public data is uneven. It covers some countries, languages and industries far better than others, particularly English-language and United States sources.
- Legal and tax detail depends on place and changes often. The Legal risks and Capital groups tell you what to check and who to ask, not what the law requires of you.
- Finding nothing is not proof. No competitors or complaints found does not mean there are none.
- Web pages can hide instructions aimed at AI. The agent can change your report's pages, so a hidden instruction could at worst produce an edit you did not ask for. Every change it makes is marked as the agent's, listed in the answer that made it and undone in one click; it cannot use a report credit, delete one of your pages, or add a link to a site your report does not already cite; and connected assistants can only read.
5. People stay in charge
- Nobody at SeedBasis reviews a report before you see it, and nobody reads your idea to run the service.
- You decide what to do with a report. You can edit any finding or chart; your edits are labelled as yours, and the researched original is kept.
- Before you act on legal, financial, tax or investment content, have it reviewed by a qualified professional. Anyone you share a report with may read it as research: the decision stays theirs, and it is not advice. A report you share says it is AI-generated research by SeedBasis, and a copy you make of it should say so too.
- Before checkout opens, you confirm that you understand reports are research generated by AI, that they can be wrong, and that they are not legal, financial, tax or investment advice.
6. Transparency under the EU AI Act
Article 50 of the EU AI Act requires AI systems that interact with people to make that clear from the first interaction, and AI-generated text to be marked in a machine-readable format. This is how SeedBasis does it:
- Our public pages, this disclosure and the acknowledgement before checkout say that reports are generated by AI.
- In the app, every conversation with the agent opens by saying that it is an AI, that it can be wrong, and that it is not a lawyer or financial adviser, and a line under the message box repeats it.
- Every finding and block is stored with its origin:
researchedfor content produced by the AI research,user_editedfor your edits. The origin travels with the content in your data export and in everything a connected assistant reads. - On report pages, AI-generated content is also marked in the page itself: every finding and block the research produced carries
data-ai-generated=trueanddata-origin=researchedin the page's HTML, your edits carrydata-origin=user_edited, and every report page declaresai-disclosure: ai-originatedin a meta tag. - Every PDF you download says on its cover and its summary that it is AI-generated research, and so do the cover of every Word copy and set of slides; each file carries the same inside it, the IPTC digital source type that document tools read, and every PDF a mark sealed by SeedBasis. Anyone can check a PDF at seedbasis.com/check, where it is read on their own device and never uploaded.
- A report you share by a view-only link says on every page that it is AI-generated research by SeedBasis and not advice, and its pages declare
ai-disclosure: ai-originated.
7. What happens to your data at our AI providers
| Provider | What it says it keeps | Training |
|---|---|---|
| OpenRouter | Does not store prompts or responses unless logging is turned on, which we leave off. It keeps request details such as the model used, token counts and timing. | Does not train on them |
| Fireworks AI and Together AI, for the open model | Keep no copy of our requests: they run only on their endpoints with zero data retention. | Nothing is kept, so nothing can be trained on |
| Keeps our requests for up to 55 days to detect abuse, then deletes them. A request that carries a document you attached is different: it goes only to Google endpoints that keep no copy, as does every request of a company report that carries your own figures or documents. | Its paid API terms do not allow it to train models on customer content | |
| xAI | Keeps no copy of our searches: they run only on its endpoints with zero data retention. | Its terms do not allow it to train models on our data |
| ElevenLabs | Keeps no copy of a call's recordings or the answers it speaks: they run only on its endpoints with zero data retention. | Nothing is kept, so nothing can be trained on |
| OpenAI | Keeps API requests for up to 30 days to monitor abuse, then deletes them. | Does not use API data to train its models |
| Anthropic | Deletes API requests within 30 days. | Its commercial terms do not allow it to train models on customer content |
These are the providers' own statements. What we allow and do not allow:
- We do not let requests go to model providers that may train on your data.
- Every request names the only companies that may run it. No other company that hosts the same models receives it.
- xAI's model receives only search words, never your idea text or anything about you, and a search runs only on xAI's endpoints with zero data retention.
- OpenAI receives only a few words naming a kind of product, never your idea text or anything about you. Anthropic receives those words and, for the check of a report's claims, the research brief's description of your idea and its customers, the claims with their quotes and passages of public pages, never your name, email address or account details. Zero data retention is not required of either: their own endpoints, the only ones whose search and answers match their assistants' and the only ones we let run Anthropic's model, do not offer it through OpenRouter, and Anthropic deletes API requests within 30 days. A report about a company you run that carries figures you typed or documents you attached skips that check.
- Zero data retention is required of the companies that run the open model and of xAI. For most requests to Google we use Google's standard endpoints, which keep a request for up to 55 days to detect abuse and may not train on it: Google offers a single endpoint for its model that keeps nothing, and a report should not depend on one endpoint with nothing to fall back to. Your own documents, and a company report's own figures and documents, go only to the endpoints that keep nothing.
- Your documents are the exception. Every request that carries a document you attached, whether the agent reads it for an answer, deep research reads it for your question, or Google's model copies out the words of a scanned PDF's pages, runs only on endpoints that keep no copy, whichever model answers it. So does a report about a company you run that carries figures you typed or documents you attached: every request of the models that research, write and check it goes to such endpoints, from its research plan to its last check. If none is available, the request waits or fails; it never falls back to one that keeps a copy.
- When you talk with a report, every request of the call, the agent's included, runs only on endpoints that keep no copy, with one exception: if ElevenLabs cannot speak an answer, OpenAI's audio model speaks it from its own endpoint, which does not offer zero data retention through OpenRouter. It receives only the agent's answer, never your voice. We never keep a recording of your voice.
- We do not send your data to any model provider that is not on the subprocessors page, and we update that page before we add one.
8. Questions
If a report looks wrong, or you have questions about how SeedBasis uses AI, write to support@seedbasis.com. The Privacy Policy covers your data more broadly.