How to Compare AI Agency Proposals: A Scorecard and Red Flags to Spot Before You Sign

To compare AI agency proposals, put them on the same scorecard: scope clarity, testing plan, handoff and ownership, integrations, pricing structure, and evidence of past work. Then check for red flags such as vague scope, only demos, no testing plan and no handover. The cheapest quote is rarely the comparable one.

Key takeaways: Normalise the scope first, then score each proposal on the same six criteria, and treat missing testing, handoff or ownership terms as a reason to pause, not a detail to sort out later.

Why are AI agency proposals so hard to compare?

They are hard to compare because each agency interprets your request differently, so you end up with quotes for different projects. One proposal may include a WhatsApp chatbot with CRM sync and a testing phase. Another may cover only a website widget that answers FAQs. A third may be a single line item called "AI agent" with no breakdown at all. If you put the three totals side by side, you are comparing numbers, not offers. Before scoring anything, you need to find out what each one actually includes and what it leaves out.

The best defence is to start from your own brief. If you sent every agency the same document, you can check each proposal against it and see which requirements were answered, reinterpreted or ignored. If you did not, our guide on how to write an AI project brief will help you build one now. You can send it to all the agencies and ask them to revise their quote against it. That one step often makes the differences between proposals much clearer, and it shows you which agencies listen.

How do you put proposals on equal footing first?

Build a simple scope checklist from your brief and mark, for every proposal, what is included, what is excluded and what is not mentioned. The third category matters most, because silence usually becomes an extra charge or a gap in the finished work. Do this in a spreadsheet with one row per requirement and one column per agency.

1. Channels: website, WhatsApp, Instagram, email, phone. Which ones are covered?
2. Integrations: which of your tools (CRM, calendar, Shopify, Slack) are connected, and in which direction does data flow?
3. Knowledge: who prepares and loads your FAQs, policies and documents, and how are they kept up to date?
4. Human handoff: what happens when the agent cannot answer, and who receives the conversation?
5. Languages: are all the languages you need explicitly named?
6. Post-launch: is there a support or adjustment period, and is it described?

Once the rows are filled in, send each agency a short list of questions about the blanks. How they answer is useful data in itself. A good provider replies clearly and updates the proposal in writing. A weak one says "we will see that during the project," which usually means the cost of that item is still unknown. For the reasons behind big price gaps, see what drives the cost of a custom AI agent project.

What should an AI agency proposal scorecard include?

A useful scorecard has six criteria, each scored from 1 to 5, with a weight that reflects what matters most to your business. Keep it simple enough that you will actually fill it in. Weights are your call: a clinic handling patient data may weight security heavily, while a shop may weight integrations and cost of changes more. Write the weights down before you read the proposals, so a polished presentation does not shift them.

Add a seventh line for fit if you want it: did they understand your business, or did they send a template? Fit is hard to quantify, but you can feel it in whether the proposal uses your real situation (your volume of enquiries, your tools, your hours) or generic language. Score it honestly and keep it separate from the technical criteria so it does not hide a weak point elsewhere.

What red flags should you look for before signing?

The main red flags are vague scope, a proposal built around demos, no testing plan, no handoff terms, and pressure to sign quickly. None of them proves bad intent, since some agencies simply write thin proposals. But each one points to a gap that tends to turn into an argument, a delay or an extra bill later. If you see two or three together, ask for a revised proposal before moving forward.

• Vague scope: phrases like "full automation," "complete AI solution" or "as needed" with no deliverables. If you cannot tell what you will receive, neither can the agency be held to it.
• Only demos: an impressive demo on generic data says little about how the agent behaves with your messy real questions. Ask how they will test with your own conversations and documents.
• No testing plan: no mention of test cases, review rounds or go-live criteria. Compare it with our pre-launch chatbot testing checklist.
• No handoff or ownership terms: nothing about accounts, credentials, documentation or what happens if you leave. See who owns your AI agent.
• Guaranteed results: promises of specific savings or conversion gains before they have seen your data. Outcomes depend on your process and volumes.
• No mention of human escalation or limits: an agent that never says "I don't know" is usually one nobody planned for failure.
• Pressure tactics: a discount that expires in 48 hours is a sales tool, not a project need.

How do you judge the testing plan and handoff?

Read these two sections more carefully than the price, because they decide whether the agent works on day 30 and whether you can keep using it on day 300. A good testing plan names who writes the test cases, how many rounds there are, what counts as a pass and who signs off on launch. It also leaves room for your team to try the agent with real questions before customers do. If the proposal says only that the agent will be "tested and optimised," ask for those details in writing.

For handoff, look for concrete items: who owns the accounts where the agent runs, who holds the credentials, whether you receive documentation of prompts, flows and integrations, and whether someone on your side is trained to make small changes. Also ask what monthly costs continue after launch, such as model usage, messaging fees and maintenance, and who pays them. Third-party platform fees, for example for WhatsApp, are set by the platform, so check them on Meta's WhatsApp pricing documentation instead of trusting a figure in a proposal. An agency that is comfortable with these questions is usually one that has done a handover before.

How do you weigh price against everything else?

Weigh price last, after you have normalised scope and scored quality, and compare it as total cost over a realistic period rather than the headline number. A low quote that excludes testing, integrations and support can end up more expensive than a higher one that includes them. Add up setup, monthly running costs and the likely cost of changes in the first year, and ask each agency how it charges for changes after launch.

It also helps to test the proposal against the value of the work. If the agent will save a few hours a week, a large build may not pay back, and a simpler option may fit better. The ROI worksheet for process automation gives you a way to estimate this before you commit. And if two proposals score close, consider whether the project suits a freelancer or an agency, as covered in freelancer or AI agency.

What should you do after scoring the proposals?

Shortlist the top one or two, send them your remaining questions in writing and ask for a short call to walk through the proposal. Use the call to check that the people who will do the work are the ones talking to you, and that the phases match what you expect. Our overview of what a custom AI agent project involves is a handy reference for that conversation.

If the numbers are close, a small paid pilot or a first phase limited to one process can reduce the risk before you commit to the full scope. Whatever you choose, get the final scope, testing plan, ownership terms and costs into the contract, not only into the sales deck. If you want a second opinion on a proposal, or want to see how Kuan AI would approach your case through diagnosis, design, implementation and measurement, you can contact us through the form.

A scorecard will not make the decision for you, but it stops the decision from being made by whoever presented best. Fill it in before the final calls, keep the weights fixed, and be willing to walk away from a proposal that cannot answer basic questions in writing.