On this page 8 sections
- Ask who writes the code
- Fixed price needs a written out-of-scope list
- You own the code, accounts and keys
- AI work needs evals and security tests
- Proof means live products and named clients
- When a larger agency is the better fit
- How to choose an AI development company
- How FormulaBot got to 1.5M+ users with two engineers
How to choose an AI development company comes down to seven checks before signing: who writes the code, how many layers sit between you and them, what is out of scope in writing, who owns the code and the accounts, how they test model output, and what proof you can verify. Ask all seven early.
Rankings of agencies are easy to find and hard to trust, because the criteria are rarely stated. A checklist works on any vendor, including us, and it surfaces the problems that cost founders months: a team that changes after the contract, a quote that grows, code you can’t legally reuse. Zeroic meets these seven, and we’ll say so once, in the section on our work with FormulaBot. This page names no agency and ranks nobody.
Ask who writes the code
Ask for the names of the engineers who will build your product, and meet them before you sign. This is the cheapest test there is. The person on the sales call is often not the person who commits code.
Count the layers between you and the engineer. A project manager, an account manager and a team lead each relay your message once, and each relay loses detail. For a single product or a few AI features, you want the engineer on the call. We run senior-only for this reason: senior engineers only, no project managers, no account managers, and clients talk directly to the people building.
Ask what seniority the people on your project have. “Senior” on a website means little. Ask how many years each engineer has shipped production software, and for one product they built that is live today.
Fixed price needs a written out-of-scope list
A fixed price is good news only when it follows scoping. Ask for the quote in writing with three parts: what is included, what is excluded, and how a change gets priced. The excluded list is the one that protects you, because every dispute about scope starts with something both sides assumed.
Be suspicious in both directions. A fixed quote from a one-page brief means padding or a fight later. An hourly rate with no cap on an AI project means you carry every estimating mistake. We cover what makes the first model work in fixed-price MVP development, and a paid audit is a cheap way to scope before committing. Our AI audit starts from $1,000 and ends with a ranked report you can take to any vendor.
You own the code, accounts and keys
Ownership is the check founders skip and regret. Under 17 U.S.C. 201(a), copyright vests initially in the author, which is whoever wrote the code, not whoever paid. The Copyright Office’s Circular 30 explains that a commissioned work counts as a work made for hire only if it falls into nine listed categories and both parties sign an agreement saying so. Custom software is not named on that list.
So a clause that says “work made for hire” may not be enough. Ask for an assignment instead. Under 17 U.S.C. 204(a), a transfer of copyright ownership is not valid unless it is in writing and signed. The contract should assign all deliverables to you, state when it transfers (on final payment, for example), and cover code written by subcontractors. This is general information, not legal advice. Have a lawyer read the clause.
Ownership also covers the accounts the product runs on:
- Repository: it lives in a GitHub organization where you are the owner. GitHub’s docs on transferring organization ownership show that ownership and billing are separate steps, so check both.
- Model API accounts: the OpenAI or Anthropic account is yours, with your billing. OpenAI’s data controls say API data is not used to train its models by default and that abuse-monitoring logs are kept up to 30 days. Those terms attach to the account holder, so you want to be that holder.
- Cloud, database and domain: registered to your company, with the vendor added as a collaborator you can remove.
- Credentials: handed over in a password manager at handoff, then rotated.
AI work needs evals and security tests
A vendor who has shipped AI products can tell you how they know the output is right. A vendor who hasn’t will talk about prompt tuning. OpenAI’s evaluation best practices say generative AI is variable, so the same input can give different outputs and traditional software testing falls short. The same guide recommends tests built from production usage, and says to “log everything” so you can mine logs for good eval cases.
Ask three questions:
- Where do the evals live? In your repository, so they outlive the vendor. As of October 16, 2026, OpenAI’s docs say its hosted Evals platform becomes read-only on October 31, 2026 and is scheduled to shut down on November 30, 2026. Tests locked in a dashboard disappear with it.
- How is output scored? Automatically where possible, calibrated against human review, and rerun on every prompt or model change.
- How do you test for prompt injection? OWASP lists prompt injection first in its 2025 LLM risks. Its mitigations include deterministic validation of output formats, least-privilege access, human approval for high-risk actions and regular adversarial testing. A good vendor names those without prompting.
Proof means live products and named clients
Ask for a product that is live now, a named client who will take a call, and what the vendor did on it. Unnamed “enterprise clients” can’t be checked. A case study with a number and a person attached can.
Review sites help if you read how they work. Clutch says every review is checked for the reviewer’s identity and work history, and publishes a review it can’t verify with a “Not Verified” label. Filter for verified reviews and read the ones that describe a problem and how it was handled. Our own approach to these projects is on the AI development company page.
Then run the reference call. Ask the client three things: did the price change, did the team change, and would they hire again for the next project.
When a larger agency is the better fit
A small senior studio is not the answer to every project. A larger agency is the better choice in three cases:
- You need 24/7 support with an on-call rotation and a contractual response time.
- The work needs a large parallel team, such as ten engineers across web, mobile and data for one launch.
- You need delivery or compliance coverage across several countries, with local entities.
If none of those apply, you are paying for overhead you don’t use. If one does, the price of the extra layers is worth it.
How to choose an AI development company
Score each vendor on these seven. A “no” on the first four is a reason to walk away.
| # | Check | Passes when | Walk away if |
|---|---|---|---|
| 1 | Who builds it | You meet the named engineers before signing | You only meet sales |
| 2 | Layers | At most one person between you and the engineer | PM, account manager and lead all relay |
| 3 | Price | Fixed after scoping, with a written out-of-scope list | Quoted from a brief, or hourly with no cap |
| 4 | Ownership | Signed assignment of all deliverables to you | ”Work made for hire” is the only clause |
| 5 | Accounts | Repo, cloud and model accounts in your name | The vendor holds the keys |
| 6 | Testing | Evals in your repo, plus prompt injection tests | No answer on how output is checked |
| 7 | Proof | Live product, named client, reference call | Only unnamed logos |
How FormulaBot got to 1.5M+ users with two engineers
Here is how these checks look in practice. FormulaBot, now Better Analyst, has 1.5M+ users. Two Zeroic senior engineers worked embedded with its founding team for more than three years. They built the English-to-SQL pipeline that replaced the rules-based first version, then the dashboards, data flows and 40+ data connectors. The founder worked with the engineers directly, with no project manager between them.
The same structure shows up on a fixed scope. We built Artswrk, a two-sided marketplace, from zero in 8 weeks, and it grew to 10k+ creators before its acquisition. If you are scoping a first version, our MVP Sprint is fixed scope and fixed price, and you own 100% of the code, infrastructure and credentials. For a longer engagement, compare it with the math in retainer vs full-time hire.
Already have a codebase from a previous vendor? Read Code Rescue vs a full rewrite before you hire the next one.
Frequently asked questions
- How do I choose an AI development company?
- Run seven checks before you sign: who writes the code and whether you meet them, how many layers sit between you and the engineer, a fixed price tied to a written out-of-scope list, signed code ownership, accounts in your name, evals and security testing, and verifiable proof. A vendor that dodges any of the seven is telling you something.
- Who owns the code an AI development company writes for me?
- Under US copyright law, copyright starts with whoever writes the code, and a transfer is valid only if it is in a signed writing. Unless your contract contains a signed assignment, you may hold only a license. Ask for a clause that assigns ownership of all deliverables to you, and have a lawyer read it.
- Is a fixed-price quote a good sign?
- Only when it comes after scoping and lists what is out of scope in writing. A fixed price quoted from a one-page brief means the vendor will either pad it or fight you over changes later. A fixed price with a named boundary, a delivery date and a change process puts the estimating risk on the vendor.
- Should I pick a large agency or a small studio?
- Pick a large agency when you need 24/7 support, a big parallel team or delivery across several countries. Pick a small senior studio when the work is one product or a few AI features and you want the people building it on every call. Match the vendor's structure to the size of the job.
- What should I ask an AI development company about testing?
- Ask how they measure whether the model's output is correct, where those tests live, and how they test for prompt injection. A good answer names a test set built from production usage, scored automatically, kept in your repository, plus adversarial tests and human approval on risky actions. Vague answers about prompt tuning are a red flag.
References
- US Copyright Office - Circular 30: Works Made for Hire
- 17 U.S. Code 201 - Ownership of copyright
- 17 U.S. Code 204 - Execution of transfers of copyright ownership
- GitHub Docs - Transferring organization ownership
- OpenAI API docs - Data controls in the OpenAI platform
- OpenAI API docs - Evaluation best practices
- OWASP GenAI - LLM01:2025 Prompt Injection
- Clutch Help Center - How Clutch verifies reviews