How to compare an AI agency, freelancer and in-house team

The label on an AI supplier's website tells you little about who will deliver your project. An agency may use subcontractors; a freelancer may work with specialists; an internal team may still depend on external platforms. Compare the proposed arrangement against your actual job.

This guide is about delivery models and commercial questions. For the different types of work you might buy, see what an AI consulting proposal should include.

Start with the work you need covered

List the capabilities the project requires: understanding the process, designing the system, building integrations, reviewing security, testing, training and ongoing operation. One person may cover several roles, but the proposal should say who is responsible for each.

A supplier's size is not a substitute for that information. Ask who will actually work on the project, who reviews their work and what happens if someone becomes unavailable. Find out which parts are subcontracted and who remains accountable.

For help delivering the work, see our AI automation agency and AI consultancy services.

Understand the commercial model

Fixed scope and price

A fixed price can make comparison easier when the deliverables and acceptance criteria are clear. Check the assumptions, exclusions, number of review rounds and process for changes. “An AI assistant” is not a sufficiently precise deliverable on its own.

Time and materials

This can suit work where discovery will change the requirements. Agree reporting, spending limits and decision points. Ask how unfinished work is handed over and how you can stop without losing access to what has already been produced.

Retained support

Specify the work the retainer covers, the route for reporting an issue, response arrangements and exclusions. Distinguish keeping an existing workflow operating from building new features. Confirm whether platform and usage fees are included.

Outcome-related fees

If a fee depends on an outcome, define the baseline, measurement period, data source and factors outside either party's control. Extra capacity, lower costs and new revenue are different measures. Avoid a payment formula based on a number neither side can independently reproduce.

Compare evidence rather than promises

Ask for a relevant example the supplier is entitled to share. Understand what they built, their role, the starting point and how results were measured. A named customer, a prototype and a production system are different kinds of evidence.

Where confidentiality prevents a public case study, ask for a permitted reference, a redacted process explanation or a demonstration using fictional data. Do not pressure a supplier to reveal another client's information. Apply the same care to your own material during selection.

Make the proposal testable

For each deliverable, write down how you will accept it. An integration should be checked against the actual destination, including duplicate handling and failure recovery. Training should identify the audience and tasks they need to perform. Documentation should be usable by the person taking ownership.

Fictional comparison: two suppliers quote for enquiry triage. One offers a chatbot demonstration. The other includes approved source material, staff review, CRM integration, test cases and a fallback. These are different scopes, even if both proposals use the same headline.

Ask both suppliers to price the same required outcome and clearly mark optional work. That is more informative than deciding whether one fee looks expensive against an unsourced market range.

Check ownership and the exit route

• Who owns the code, configuration and documents?
• Which accounts must be in your business's name?
• Can you export the information in a usable format?
• What licences or supplier-owned components remain dependencies?
• Who maintains connections when providers change their services?
• What support and access remain when the engagement ends?

Have these points addressed in the written agreement. A live system can become difficult to maintain if the only working credentials, source files or operating knowledge remain with a departing supplier.

Use a consistent selection record

Score each proposal against the same requirements: fit to the task, delivery evidence, named responsibility, acceptance, information handling, total cost and maintainability. Record uncertainties separately rather than turning missing information into an optimistic score.

Agree your own project owner before signing. Suppliers need timely decisions and access to the people who understand the work. A capable team cannot resolve an internal disagreement about the process by choosing a model.

See Blue Canvas AI consultancy, current pricing and published case studies, or book a free conversation about the work you need covered.