How to Choose an AI Agency Without Buying the Pitch
Judge the proposed job, the people, and the evidence before you judge the sales deck.
The sales deck is the easy part. Your job is to find out what the agency has built, who will do the work, and what happens when the system is wrong.
Write the job down first
Do not begin with “we want AI.” Describe the work that needs to change: the support queue that takes too long, the documents people rekey by hand, or the inventory decision that lacks a usable forecast. Name the users, inputs, systems, and decision the project must support. That gives every agency the same job to price and gives you something concrete to compare.
- State what people do today and which part of that workflow should change.
- List the data and systems the work will touch.
- Write down the errors, security limits, and approval steps the solution must handle.
- Decide what evidence would show that the project is useful before you approve a larger rollout.
Ask for evidence, not adjectives
Ask for a project that resembles your job, then ask what the agency shipped, what went wrong, and what the client had to maintain. Industry experience matters when the work has specific constraints. A team working in healthcare, finance, or legal should be able to explain the relevant compliance and operating limits without hiding behind the industry label.
Case studies, client references, Google ratings, and review counts are different kinds of evidence. A 4.5+ rating with 20+ reviews can be a useful public filter, but it does not prove fit for your project. No reviews may mean a firm is new or that its clients do not post public reviews. In either case, ask for work you can examine and a client you can question.
Technical names are not proof by themselves. If an agency mentions PyTorch, TensorFlow, LangChain, AWS, Azure, or GCP, ask why that choice fits your systems and what tradeoffs it creates.
Useful answer: “We would test a RAG pipeline against your existing documents and deploy it in your AWS account. Here is how we would measure retrieval errors.”
Weak answer: “Our proprietary platform handles the whole process automatically.”
Find out who will do the work
Meet the people who will lead discovery, build the system, test it, and manage the project. Ask how much of their time is included and whether any work will be handed to another team after the contract is signed. The credentials in a pitch matter less if those people will not be on your project.
The commercial model should match the work. Treat these figures as planning ranges, not quotes:
- Fixed project: $10K–$200K when the scope and acceptance criteria are defined.
- Hourly or daily work: $150–$400/hr for senior AI engineers during exploration or changing scope.
- Monthly retainer: $5K–$50K/mo for maintenance, support, or continued improvement.
For a fuller list of assumptions and ranges, read the AI agency pricing guide.
Check delivery and maintenance
A paid discovery phase or small proof of concept lets you examine the working relationship before a larger commitment. If the proposed build is $200K, a $5K–$15K discovery phase is a planning option, not insurance against a bad project. It should end with a usable scope, open risks, and a clear decision about whether to continue.
Put the handoff terms in writing:
- Who owns the code and models when the project ends?
- What can your team maintain, and what still depends on the agency?
- What happens to your data during the work and after the contract?
- Will the system run in your infrastructure or the agency’s?
- Who monitors errors, updates dependencies, and approves changes after launch?
Make the final call
Compare each agency against the same written job. Look for relevant evidence, named delivery people, separated costs, testable acceptance criteria, and a handoff you can live with. A polished pitch cannot make up for a vague scope or an unknown delivery team.
Review AI agency profiles to build a shortlist, then verify the claims that matter to your project. If the brief is ready, send it to us.