How to judge an AI startup pitch at a London demo night

Separate present AI capability, customer evidence, commercial claims and future plans when assessing a startup pitch at a London demo night.

Direct answer: Separate what the product does now from customer evidence, commercial claims and future plans. Record the exact task, available product access, data used, human work hidden around the AI and one failure the pitch does not resolve. Use a London demo night to discover a candidate, then verify the company and rerun a representative task before buying, partnering or recommending it.

A short pitch is designed to compress. It usually combines a problem, product demonstration, customer story, market argument and future vision in a few minutes. Compression helps an audience follow the idea, but it can erase the boundary between present capability and intended capability.

The correct response is not cynicism. It is careful classification. A startup can be early and useful. You simply need evidence appropriate to the decision you are considering.

Applies to: public startup pitches, accelerators and AI demo nights in London. This is a product-adoption method, not investment advice, and no current startup or event is recommended.

Use the pitch evidence separation

The pitch evidence separation is an editorial method with five columns:

1. Problem: the user and task the startup says it serves.
2. Present capability: what the available product demonstrated now.
3. Customer evidence: what named or described users have achieved under stated conditions.
4. Commercial position: current access, support, price basis and contractual route.
5. Future plan: features, markets or performance not yet available.

Write each claim in one column. A future integration does not become present capability because it appears on the same slide. A customer logo does not tell you which product version the customer used or whether it paid.

The parent article How to Choose the Right AI Tool Without Being Sold by the Demo asks for task fit, difficult examples, privacy, complete cost and exit. The pitch is a lead for that process, not an exception to it.

Identify the product inside the story

Reduce the pitch to an input, process and output. “An AI platform for professional insight” is not enough. “A web application that takes a supplied public report and drafts paragraph-referenced questions for a reviewer” can be tested.

Ask who performs each step. A founder may prepare data, configure prompts, correct output and assemble the final view before the audience sees it. Human service is not a defect; many useful products include it. The buyer needs to know whether the same help is part of the ordinary offer and cost.

Record the application, model or external service, account conditions and connected sources when disclosed. Do not assume the startup trained the underlying model or owns every component. What matters for adoption is who is responsible for the complete system and what happens when a dependency changes.

Ask whether the displayed product is generally available, limited beta, private pilot or prototype. Those labels have no universal technical definition, so request the practical consequences: who can access it, what support exists, which terms apply and whether work can be exported.

Test customer and performance claims

For a customer claim, capture:

• customer type and task;
• previous process or baseline;
• product configuration and support;
• period measured;
• output reviewed and by whom;
• setup, correction and maintenance time;
• failures or excluded cases;
• commercial relationship and permission to use the story.

A statement that work became “ten times faster” is incomplete without start and finish points. Did timing include source preparation, checking and exceptions? Was the comparison manual work by a beginner or an established process? Ask for the method rather than challenging the number rhetorically.

Current UK advertising guidance says marketers should hold evidence for objective claims which consumers are likely to understand as factual. The ASA's substantiation guidance also explains that objective claims in testimonials need supporting evidence. Whether a particular stage statement is legally an advertisement depends on context; this article makes no legal finding. The editorial habit remains sound: attach the evidence to the claim.

My recommendation is to shortlist a startup only when the present capability solves a defined task and the founder can identify the limits and verification route. The recommendation changes for an exploratory conversation where the purpose is co-design rather than purchase, but roles, data and ownership still need written boundaries.

Ask one question from each missing column

An eight-minute pitch rarely answers everything. Choose the first missing fact which changes your next step.

Do not demand confidential customer information in public. Ask for a permission-cleared case study or a reproducible test. A responsible refusal to disclose can be appropriate; it means the claim may remain unverified for your decision.

For data handling, ask what information leaves your device, which organisations receive it, where current terms are published and how deletion or export works. Do not upload a real file at the demo night.

Verify identity without confusing registration with quality

Confirm the legal entity named in the terms or contract. The Companies House service provides free access to basic company information and filings. Its search guidance describes the available public records. The live register also warns that Companies House does not check the accuracy of information filed. Registration therefore helps establish identity and status; it does not validate product performance, solvency, security or every claim.

Check the official product domain, privacy notice, terms, support contact and company details are consistent. Be cautious when a pitch uses a trading name without showing who supplies the service.

For a material contract, security assessment, investment or partnership, use qualified professional due diligence. Requirements vary by transaction and jurisdiction. A London presentation and UK registration do not show where data is processed or which law governs every service relationship.

Work a pitch-to-test calculation

This example uses illustrative observations and costs, not a real startup. An attendee records six claims during a demo night:

• two present capabilities shown on stage;
• one customer result with a named baseline but no review-time figure;
• one current price claim;
• two future product plans.

Only two of six, or 33.3%, are observed present capabilities. That is not a pitch quality score. It tells the attendee not to repeat the other four as demonstrated functions.

Now calculate the event commitment:

• preparation: 20 minutes;
• return travel: 80 minutes;
• event: 150 minutes;
• source and company checks: 50 minutes;
• synthetic follow-up test: 90 minutes.

Total is 20 + 80 + 150 + 50 + 90 = 390 minutes, or 6.5 hours. At an illustrative time value of £24 per hour, time value is 6.5 × £24 = £156. Add £10 travel for a complete decision cost of £166.

The follow-up test finds that the tool completes the ordinary case but loses one source reference in the difficult case. That observation is more useful to adoption than increasing the “present capability” proportion. The next step is to ask whether the source loss is detectable and correctable, not to calculate a general accuracy rate from two cases.

Time value is not cash earned or saved. The arithmetic makes the discovery process visible.

Distinguish an early product from an unsafe dependency

An early startup may change features, providers, pricing and support quickly. That can be acceptable for a reversible, low-risk experiment. It is a poor fit for essential work when no export, fallback, owner or support route exists.

UK government AI assurance guidance says organisations may need different techniques across design, deployment and operation, and should consider governance alongside system and data evidence. The official introduction emphasises proportionate assurance rather than a single approval method.

Before a pilot, define:

• a safe synthetic or minimised dataset;
• the task and acceptance test;
• who checks every material output;
• maximum spend and duration;
• a manual fallback;
• export and deletion checks;
• a stop condition.

Do not let proximity to founders replace documentation. Founder access can improve support, but a process which works only when the founder intervenes is a service dependency that should be priced and recorded.

Run a seven-day follow-up

On the night, classify the material claims into the five columns and ask one missing-column question. Within two working days, verify the legal supplier, current product terms and the official documentation for the demonstrated capability.

Within seven days, run one ordinary and one difficult synthetic case under access you could actually obtain. Record raw outputs, review time, corrections, support and export. Compare with the current non-AI process.

Proceed to a bounded pilot only when the present capability, responsibility, data terms and exit fit the task. Stop when the product depends on an unavailable future feature, the legal supplier is unclear, a material claim cannot be supported or the only successful path requires founder intervention not included in the offer.

Related guides

• AI policy panel versus product workshop: which London session should I attend?
• How to choose a London AI event when you need an accessible format

Frequently asked questions

Does an early-stage startup need the same evidence as an established vendor?

It needs evidence proportionate to the claim and consequence. An early team may have fewer customers and shorter operating history, but it should still distinguish present capability from plans, explain the test conditions and disclose material limits. You can accept more uncertainty for a reversible experiment with synthetic data than for essential or regulated work. Do not lower the standard for privacy, security or rights because the founders are accessible. The condition that changes the decision is not company age by itself; it is whether the risk, controls, ownership and fallback fit the proposed use.

What does “beta” tell me about product readiness?

Very little without further detail. Ask who can access the product, which features and terms apply, what support and uptime are offered, whether data or behaviour may change, and how users export or delete work. A beta can be stable enough for a low-risk test or far too uncertain for dependency. Do not assume it is free, private or safe. Record the version and date because conditions can change. The exception is when the startup publishes a precise beta policy and service description; use those facts rather than the label alone.

Is a customer logo evidence that the company uses the product?

Not enough evidence for your decision. The relationship might be a paid deployment, trial, partnership, supplier arrangement or permission to mention a brand. Ask what task, configuration, period and result the logo represents and whether a public case study exists. Do not contact customer staff or infer endorsement without permission. A detailed, permission-cleared case can be useful when its conditions match yours. The exception is a simple statement directly confirmed by the customer through an authoritative source, but it still does not prove wider product quality.

Should I sign up for a pilot at the event?

Only if the pilot terms, data boundary, cost, owner, task, review method and exit have already passed your checks. An expression of interest is different from authorising access or uploading information. Event excitement and limited spaces can compress judgement. Take the written offer away, compare it with alternatives and confirm internal approval. A low-risk demonstration account using synthetic data may be reasonable immediately when no payment or continuing commitment is hidden. Do not connect work systems or accept consequential terms from a phone screen in a crowded venue.

Can Companies House tell me whether the startup is trustworthy?

No. It can help identify a UK company and show public filings and status, but the service itself warns that filed information is not checked for accuracy. Registration does not validate an AI product, financial health, cyber security, customer claim or founder competence. Use it as one identity check alongside contract, domain, people, product and evidence review. Overseas entities or sole traders may require different checks. For a material purchase, partnership or investment, obtain professional due diligence appropriate to the transaction and jurisdiction.

How should I treat a feature described as “coming soon”?

Record it as a future plan with no current decision credit. Ask what works today without it, whether access, price or terms are known and what evidence will confirm release. Do not design an essential workflow around a date the supplier can change. A roadmap can help you understand direction, but it is not a delivered capability or contractual commitment unless your agreement makes it one. The exception is a reversible exploration whose purpose is to help shape the feature; define the co-design role, data, rights and stopping point in writing.

Sources and verification

• Substantiation, Advertising Standards Authority and CAP, checked 6 October 2026 for current guidance on documentary support for objective claims and testimonials.
• Searching the Companies House register, GOV.UK, checked 6 October 2026 for the current description of company information available through the official service.
• Find and update company information, Companies House, checked 6 October 2026 for the live register and its warning that filed information is not checked for accuracy.
• Introduction to AI assurance, GOV.UK, checked 6 October 2026 for current guidance on proportionate, lifecycle and governance-aware assurance.

No current London demo night, startup, customer, product, price or company record was assessed. Every claim count, cost and test result in the example is illustrative. This article is not investment or legal advice.