How we use AI
This page lists every feature that calls a language model, what the model is given, and what fixed rules decide instead.
I. AI LTD trading as InvestorUniverse · Scoring rules version 2.1.2 · Read on 5 October 2026, 22:30 UK time
Where a language model is used
The model is supplied by an outside provider, named on the sub-processors page while it is connected. These are all of its uses.
| Feature | What the model is given | What comes back, and the check |
|---|---|---|
| Reading your company's website | The public text of the website you enter. | A one-sentence description, a sector, a geography and a business model. Each is kept only if it is a value we recognise. Fixed rules read the site first, and their reading stands if the model fails. |
| Pitch deck feedback | The text of the deck you upload, and the sections a rule-based check found missing. | A summary, strengths, improvements and suggested profile details. You choose which details to keep. Nothing is applied to your profile without you. |
| First-email drafts | Your company name and description, sector, raise and stage. The investor's firm and type, the recipient's name and job title, up to two portfolio companies and the reasons for the fit. | A subject and a body for you to edit. If the model fails you get a fixed template. |
| Anonymous investor briefs | Structured facts only: investor type, location band, stages, sectors, cheque range and activity band. No name. | Two sentences and a short thesis. The wording is thrown away if it could identify the investor, and the rule-written brief is used. |
| Questions about your results | Your question and your own list. Investors you have not unlocked appear only as a number. | A short answer, shown only if it names no locked investor, states no figure that is not in your data and carries no link. Questions about contact details are refused by rule and never reach the model. |
| The assistant in your workspace | Your question and the names of the tools the assistant has. | A choice of which tools to run. The model does not write the answer: the answer is what the tools read from your own workspace. Used only when fixed rules cannot route the question. |
| Meeting prep and the simulated committee | Facts the rules have already selected about your raise and the investor. | A short rewording of those facts, checked again before it is shown. |
| Staff tool: whom to ask before a public listing | Firm or company names with counts our rules computed. | An order and a one-sentence reason. A member of staff decides whether to ask, and approves each request. |
What the model is never given
- An email address or a phone number, of an investor or of anyone else.
- The name of an investor you have not unlocked, when you ask about your results.
- The investor database as a whole. Each call carries the facts for one task.
- Another customer's data. A call is built from the workspace of the person it is for.
- Payment details. We do not hold card details at all.
Controls on every call
- Every call goes through one client, with a named prompt that carries a version.
- Each call has a time limit and is tried at most twice.
- The reply is parsed and checked against a fixed shape before anything uses it. A reply that fails is dropped.
- There is a daily spending ceiling. Once it is reached, calls are refused and the rule-based fallback is used.
- Every feature above works without the model, from rules or a template.
- We log each call's prompt name and version, the model, token counts, cost, time taken and whether the reply was valid. That log does not hold the text sent or the reply.
What is computed by rules, not by a model
- Matching and scoring. The Fit Score is a weighted average of 16 factors, computed by code. The scoring engine makes no model call. The weights are published in the methodology, and the version of the scoring rules, 2.1.2, is stamped on every list.
- Fit and confidence are not probabilities. Neither figure says how likely an investor is to invest.
- Exclusions, plan limits, prices and what you may see are decided by rules.
- One input can come from a model. The model's reading of your website can set your company's sector, geography and business model, and the scoring rules then use them. You see that reading and can correct it.
A model output is never verified
Every fact we hold carries a source tier from A to F. Tier F is “inference or model output”. A tier F fact can carry only the state Inferred or Estimated. It can never be Verified. The code enforces that, and the database rejects a row that breaks it. The tiers are set out in the methodology.
What needs your approval, and what does not
- A first email in your name. A first email to an investor is sent only after you approve that message. You can edit it first, and you can approve several drafts in one action.
- Follow-ups follow your setting. With follow-ups switched on, up to 2 short follow-ups in fixed wording are queued after a first email is sent. You can cancel each one before it goes, and a follow-up is dropped if the investor has replied or you have moved them on. With follow-ups switched off, none is sent.
- Sending has its own switch. Sending on a founder's behalf is not switched on at present, so approved emails wait in the queue. No founder can send more than 40 emails a day through us.
- Changes the assistant suggests. When the assistant proposes a change to your workspace, it takes effect only when you confirm it.
- Removal requests. A removal takes effect at once by rule, and a person then reviews every request.
Asking for a human review
If you think a score, a classification, a brief or a draft is wrong, ask for a person to look at it. Write to us through the help centre or email support@investoruniverse.uk, and say which investor or which part of your profile it concerns. If it is about your own personal data, use the data request form or email privacy@investoruniverse.uk.