AI Transparency

GoPresent Ltd uses generative artificial-intelligence ("AI") models extensively to deliver the Services. This notice explains, in plain English, which models are used, what they receive as input, what they produce as output, what the human-in-the-loop and review steps are, and the known limits of AI-generated content.

We publish this to support the disclosure obligations under Article 50 of the EU Artificial Intelligence Act 2024 (effective 2 August 2026), the equivalent guidance from the UK Information Commissioner's Office, and the broader expectation that AI-assisted services should be transparent about the role AI plays in shaping their output.

This notice is not a substitute for our Privacy Policy at /legal/privacy-policy or our Data Processing Agreement at /legal/data-processing-agreement — both of which contain the formal data-protection terms.

The GoPresent Platform uses AI to produce or assist with the following:

Founder Portal — gap analysis (assessment of a pitch deck against an investor-readiness rubric); pitch-deck generation (a fully drafted deck from an intake form + uploaded source materials); conversational coaching (a multi-turn AI assistant that answers founder questions about their fundraise); presenter-script generation (per-slide talking points); fact-check (automatic verification flags on claims in a generated slide); voice-note transcription; automatic memory extraction (structured facts persisted across sessions); quick-score (a fast first-pass assessment from minimal input).

Scout Portal — quick-screen analysis of a third-party founder's deck; deal-brief / IC-memo drafting; competitor extraction; site extraction (research of a portfolio or prospect company's public web presence).

Consultant Portal — CRM helpers (drafting outreach emails, summarising client conversations) used by the GoPresent staff team to operate the consulting service.

Every surface that displays AI-generated content carries an in-product 'Generated by AI' badge or banner. The badge is the in-product expression of this notice — it tells you, at the moment you're reading the content, that the content was produced by a model rather than human-authored.

Where AI-generated content makes specific factual claims (numbers, named entities, market sizes, benchmarks, dates), the badge is accompanied by a 'Hallucination disclaimer' banner that asks you to verify the specific claims against your own sources before relying on or sharing them externally.

Where the model itself indicated low confidence in its output (e.g. confidence score below 60 / 100), an additional 'Low confidence' banner appears that suggests how to improve the result (more source material, a clearer deck, a longer transcript).

Primary provider: Anthropic PBC (USA) — the Claude family of large-language models (currently Claude Sonnet 4.5 / 4.6 for most workloads, Haiku 4.5 for fast / low-cost calls, Opus 4.6 / 4.7 priced for completeness but rarely defaulted). Anthropic is configured under enterprise terms with zero-retention where contractually offered: prompts and outputs are not used by Anthropic to train future models.

Failover provider: Together Computer Inc. (USA) — the Llama family, used as a secondary provider when Anthropic returns rate-limit (429) or server-error (5xx) responses. Failover is logged so we can observe how often it engages.

Voice-to-text: Deepgram Inc. (USA) — for transcribing voice notes uploaded by founders. Deepgram receives the audio file and returns transcribed text; the audio is not retained beyond the transcription window.

Embedding model: OpenAI text-embedding-3-small (USA) — used solely to compute 1536-dimensional vectors for the project_corpus retrieval-augmented conversation feature. We send the text of your uploaded materials + the text of in-conversation messages used as retrieval queries; we receive vectors back. OpenAI is not used for any chat completion, deck generation, or other generative output. Per OpenAI's API terms, embeddings inputs are not retained beyond the request and are not used to train OpenAI models.

We do NOT use OpenAI ChatGPT for chat completions, Google Gemini, or any custom-fine-tuned model in production today. If we change providers we will update this notice and notify active customers.

When you trigger an AI-powered feature, the relevant prompt is constructed from a combination of:

(a) Your account profile and project context — name, company name, sector, stage, current round target, profile fields you have filled in.

(b) Materials you have uploaded to that project — pitch deck text extracted via PDF parser, transcript text from voice notes, intake-form answers, source documents you attached.

(c) Prior conversation in that project — when you use the conversation feature, the prompt includes the recent conversation history and a curated 'project memory' of structured facts the AI has previously extracted.

(d) Limited platform context — the prompt template itself (which we maintain at apps/web/src/server/ai/prompts/ in our codebase, version-pinned so a given output can always be re-traced to the prompt that produced it).

Your data is NEVER combined with another customer's data in a single prompt. All AI calls are scoped to your organisation and project.

Outputs from AI calls are stored in our Supabase database, scoped to your organisation under our Row-Level Security policies. They are also written to an immutable AI-runs audit log so we can trace any output back to the prompt and model that produced it.

We do not use customer prompts or outputs to train any model — neither our own (we don't train models) nor any third-party provider's. Anthropic's enterprise terms prohibit them training on our prompts; Together's contract is reviewed periodically and confirmed in writing (see Section 11 of the DPA).

We retain AI run records under the same retention policies described in our Privacy Policy — generally for the life of your account plus the periods required by HMRC for billing records.

The Services are designed for the user (founder, scout, or consultant) to exercise judgement over AI output. The user always sees the output before any external action is taken — no AI-generated content is sent on your behalf to an investor, a portfolio company, or another third party without your explicit confirmation.

Higher-tier services delivered by GoPresent (Analyst Deep Dive at £1,999, White Glove at £7,499, and bespoke consulting from £20,000) include a formal human review step where a GoPresent consultant reviews the AI output before it is shared with the founder. These services are explicitly marketed as 'human-reviewed'.

For the £499 / £746 self-service tier (Raise Ready), the AI output is delivered to the founder directly. The founder is the human-in-the-loop. The disclaimers described in Section 3 set this expectation explicitly.

Generative AI models can hallucinate: produce plausible-sounding output that is incorrect. Specifically, in the context of GoPresent's Services, the limits we are aware of are:

(a) Specific numbers (TAM / SAM / SOM, growth rates, revenue, market shares, valuations) should be verified against your own primary sources before being cited externally. The model's training data has a cut-off and does not contain real-time market data.

(b) Named entities (specific competitor names, named investors, named customers) should be verified — the model can confuse similarly-named entities, attribute acquisitions to the wrong acquirer, or invent names that sound right but don't exist.

(c) Time-sensitive claims ("X just raised a Series B", "the FCA's current guidance is Y") will not reflect events after the model's training cut-off. The fact-check feature flags time-sensitive claims as 'needs_source' to nudge verification.

(d) Sector-specific regulatory analysis (financial services, healthcare, defence) should not be relied upon without independent professional review. Section 4 of our Acceptable Use Policy at /legal/aup makes this explicit.

(e) Investment advice — outputs are commentary and analysis, not investment advice. GoPresent is not authorised by the FCA.

GoPresent does NOT make solely automated decisions that produce legal effects on you or similarly significantly affect you in the sense of UK GDPR Article 22.

AI-generated 'scores' (e.g. capital-readiness score on the founder portal, deal score on the scout portal) are advisory inputs to a human decision (the founder deciding what to revise, the scout deciding whether to take a meeting). They do not, in themselves, alter your contractual entitlement to the Services or your access to any feature, except where you explicitly opt in to score-gated workflows for your own use.

Where score-driven gating is offered as an opt-in (e.g. 'upgrade your score by 10 points to unlock advanced advisor pack'), the gating is documented in the relevant in-product surface and the score itself is always inspectable.

If you believe an AI-generated output has caused you harm — defamatory content, misleading numbers used in an external pitch, personal data of a third party leaking into your output — please contact privacy@gopresent.ai immediately. We will investigate and, where appropriate, take corrective action including deletion of the output and notification of any affected third party.

Routine feedback on output quality (a gap analysis you think missed a key point; a deck slide that didn't read well) is welcomed at feedback@gopresent.ai and is used to improve the prompt registry.

We will update this notice when we change AI providers, add a new AI feature with materially different data flows, or when applicable law changes (notably the staged commencement of the EU AI Act). The version and last-updated date at the top of this page indicate the current version.