How do I compare AI agency proposals without being misled by demos?

9 May 2026

Start by separating the demo from the delivery promise

A demo is not proof that an AI agency can deliver your project. It is proof that they can make something look useful under controlled conditions. That may still be valuable, but it should only count for 10-15% of your decision.

The proposal should carry the weight. It should explain the business problem, the workflow, the data sources, the integration route, the risks, the acceptance tests, the rollout plan, the support model and the real cost. If the demo is exciting but the proposal is vague, believe the proposal. Vague on paper usually becomes expensive in delivery.

This matters because UK AI adoption is still uneven. The Department for Science, Innovation and Technology's 2025 AI Adoption Research, based on 3,500 business interviews, found that around 1 in 6 UK businesses currently use at least one AI technology. Among adopters, 85% use natural language processing and text generation, while agentic AI adoption is only 7%. In plain English: lots of agencies can demonstrate generative AI, but far fewer have repeated evidence of production-grade agentic systems.

The ONS Management and Expectations Survey found that AI was adopted by 9% of UK firms in 2023 and projected to reach 22% in 2024. The most common barriers were difficulty identifying use cases at 39%, cost at 21% and AI expertise or skills at 16%. Those numbers explain why demos are persuasive. Buyers want confidence. The problem is that confidence is not the same as delivery evidence.

Use a weighted scorecard instead of choosing the most impressive presentation

Do not compare proposals by feel. Build a simple weighted scorecard and tell each agency how they will be assessed. Good agencies will welcome it. Weak agencies will try to steer you back to the demo.

Score each category from 1 to 5, multiply by the weighting, and compare totals. Do not allow one dazzling demo to compensate for a missing data protection plan or unclear ownership. If a project touches personal data, customer communication, HR, finance, regulated advice or operational decisions, risk and governance should carry more weight than interface polish.

For a typical UK SME project, a credible first implementation might cost £8,000-£25,000 for a focused internal workflow, £25,000-£80,000 for multi-system automation, and £80,000+ for deeper custom development with complex data, security or change management. If one proposal is half the price of the others, do not celebrate too quickly. Ask what is excluded.

Ask every agency to prove how the demo becomes a live system

The most important question is simple: what has to be true for this demo to work in our business every day?

Make each agency answer that question in writing. Their answer should cover data access, permissions, prompt or model design, integration with tools such as Microsoft 365, Google Workspace, HubSpot, Xero, Shopify, Sage, Slack, Teams, your CRM or your helpdesk, exception handling, human approval and monitoring after launch.

A good agency will say things like: this will not work until your CRM fields are cleaned, this customer workflow needs human approval at two points, this should start as an internal assistant before it becomes customer-facing, or this part is better solved with rule-based automation than AI. That honesty is a strength, not a weakness.

A weak agency will keep the conversation at demo level. They will say the model can understand anything, integration is straightforward, or the system learns over time without explaining from what data, under what controls, and who is accountable when it gets something wrong.

Ask for an implementation map with five stages: discovery, prototype, pilot, production launch and support. Each stage should have a decision gate. For example, a £10,000 pilot should not automatically become a £50,000 rollout. The proposal should define what success looks like, what failure looks like, who signs off and when you stop.

Check the proposal for the expensive things demos usually hide

Demos hide the boring work because the boring work is what makes the system usable. That includes data cleaning, permission design, integration testing, staff training, error monitoring, documentation and support. These are not extras. They are the project.

Look for seven hidden cost lines:

• Data preparation: cleaning spreadsheets, CRM fields, documents, product data or helpdesk content.
• Integration work: connecting APIs, middleware, webhooks, Zapier, Make, n8n, Power Automate or custom code.
• Security review: access controls, vendor checks, data processing terms and role-based permissions.
• Testing: real user testing, edge cases, hallucination checks, load checks and fallback scenarios.
• Change management: training staff, redesigning the workflow and updating internal processes.
• Ongoing running costs: model usage, licences, cloud hosting, monitoring, support retainers and maintenance.
• Exit costs: export of data, documentation, IP ownership, handover and removal of vendor dependencies.

If those lines are missing, the proposal is incomplete. The agency may still be competent, but you do not yet know the cost of buying from them.

This is also where UK regulation matters. The Competition and Markets Authority's AI Foundation Models review highlights principles around competition, consumer protection, transparency, accountability and governance. For buyers, the practical lesson is not abstract. Avoid proposals that lock you into one model, one platform, one black-box workflow or one supplier without a sensible exit route.

Demand evidence of production work, not just technical talent

Technical talent matters, but production evidence matters more. A clever prototype built by one senior developer is not the same as a reliable system used by a team every day.

Ask for three types of evidence. First, live or recently live examples that are similar in risk and complexity to your project. They do not have to be in your exact sector, but they should involve comparable data, users and operational pressure. Second, named references you can speak to. If confidentiality prevents naming the client publicly, a private reference call should still be possible for serious procurement. Third, evidence of support after launch: service levels, response times, monitoring routines, documentation and training materials.

Also ask about failures. A serious AI agency should be able to explain projects that did not proceed, pilots that failed, assumptions that turned out wrong, and what they changed as a result. If every story is a perfect success, you are probably hearing sales copy rather than operational truth.

Be fair to different agency types. Accenture, Deloitte, PwC and IBM can handle large enterprise programmes, governance-heavy environments and global change management, but they may be too expensive and slow for a 20-person SME. Specialist boutiques can move faster and cost less, but may have limited capacity or narrower expertise. No-code automation consultants can be excellent for workflow speed, but may not be right for regulated, security-heavy or deeply custom systems. The right choice depends on the risk of the work, not the shine of the demo.

The questions that reveal whether the proposal is real

Put these questions into your procurement process. Do not save them for the final call.

• What exact business metric will this project improve, and what is the current baseline?
• What data will the system use, where is that data stored, and who can access it?
• What happens when the AI gives a wrong answer?
• Which parts are deterministic automation and which parts are probabilistic AI?
• What will we own at the end: prompts, workflows, code, documentation, data structures and accounts?
• What will still cost money after launch?
• What is excluded from the fixed price?
• How do we leave if the relationship does not work?
• Who on your side is actually doing the work?
• Can we speak to a client whose project is live, not just demonstrated?

Listen for specificity. A strong answer includes names, numbers, tools, stages and trade-offs. A weak answer uses phrases like enterprise-grade AI, seamless integration, cutting-edge intelligence or transformation at scale without explaining what will happen on Tuesday morning when your staff log in.

If two proposals are close, choose the one that is clearer about limits. The agency that admits a pilot may show no ROI is usually safer than the agency that promises a guaranteed productivity uplift before seeing your workflow.

When this does NOT apply

You do not need a heavy procurement process for every AI purchase. If you are buying ChatGPT Team, Claude Team, Microsoft Copilot, Gemini for Workspace, Grammarly, Fireflies, Fathom, Zapier, Make or another low-cost SaaS tool for a small group, a practical trial is usually enough. Set a budget, test for 30 days, check data rules, then keep or cancel.

You also do not need a full agency proposal if the work is genuinely tiny, such as a £500 prompt workshop or a one-day automation tidy-up using non-sensitive data. The cost of over-procurement can exceed the risk.

This framework does apply when the agency is proposing custom workflows, customer-facing AI, data integration, CRM automation, internal agents, document processing, financial analysis, HR support, regulated content or anything that could damage trust if it fails. At that point, the risk is no longer the software subscription. The risk is operational dependency.

If you want a practical next step, ask each agency to resubmit its proposal using your scorecard. Give them one page for business outcome, one page for delivery plan, one page for risk and governance, one page for pricing and one page for evidence. The best partner will become easier to spot very quickly.

If you want a second pair of eyes on proposals before you sign, book a free call. No pitch, no pressure. We will help you separate real delivery evidence from demo theatre.

Is This Right For You?

This applies if you are comparing two or more AI agencies, automation consultants, Microsoft Copilot partners, OpenAI integrators, no-code automation specialists or broader digital agencies that now sell AI services. It is especially useful if the demos look impressive but the proposals feel hard to compare.

It does not apply if you are buying a simple off-the-shelf subscription for one user. For that, run a small internal trial and cancel if it does not work. Formal proposal scoring matters when the commitment is material, usually from £5,000 upwards, or when the system will touch customer data, staff workflows, financial decisions, operational systems or UK GDPR obligations.

Frequently Asked Questions

How much weight should I give an AI agency demo?

Give the demo 10-15% of the decision. It proves what might be possible, but the proposal, delivery plan, references, risk controls and commercial terms should decide whether the agency is safe to hire.

What is the biggest red flag in an AI agency proposal?

The biggest red flag is a proposal that shows the end result but not the path to production. If there is no detail on data, integration, testing, ownership, support and failure handling, you are buying a promise, not a delivery plan.

Should I choose the cheapest AI agency proposal?

Not automatically. A cheaper proposal may be right if the scope is genuinely simpler, but it may also exclude data preparation, testing, training, support or integration work. Compare total cost, not headline setup fee.

What should an AI agency proposal include?

It should include the business problem, measurable outcome, scope, exclusions, milestones, data requirements, integration plan, risk controls, UK GDPR considerations, acceptance tests, pricing, support terms, IP ownership and exit process.

How do I know if an AI demo is misleading?

A demo is misleading if it uses perfect data, avoids edge cases, skips security questions, hides human work behind the scenes, or shows a polished interface without explaining how the system will operate inside your business.

Should I ask for client references before signing?

Yes, for any meaningful engagement. If the project is above £5,000 or touches operational data, ask to speak to at least one client with a live system. Case studies are useful, but reference calls reveal more.

What if every agency proposes a different approach?

That is normal. Force comparison by asking each agency to map its approach against the same outcome, budget, data assumptions, risk controls and support requirements. If they cannot do that, they are not ready for the work.

Do UK GDPR and data protection matter during proposal comparison?

Yes. If the AI system will process personal data, the proposal should explain data roles, access, retention, security, vendor terms and human oversight. If data protection only appears after you ask, treat that as a warning sign.