Updated 16 October 2026. New buyer guide with 15 questions, red flags and a scoring sheet; AI provider data policies checked on this date.
To choose an AI development company, first confirm that an off-the-shelf tool will not do the job. Then send the same written brief to three firms, ask each the same 15 questions, compare quotes on first-year cost (build plus 12 months of running), and pick the one that tests on your real data and leaves you owning the code, prompts and data.
Good answers are specific: a metric, a test plan, a named model with a reason, and a monthly running cost. Weak answers are vague promises, demos that never touch your data, and contracts that keep the work in the vendor's accounts. This guide aims to be fair. Apex Logic is one option among many, and sometimes the right answer is not to hire anyone.
Custom AI tool or off-the-shelf product?
Many AI needs are already covered by products you can buy today. Use this table to decide:
| Buy off-the-shelf when | Build custom when |
|---|---|
| The task is common: meeting notes, writing help, general chat | The process follows your own rules, documents or pricing |
| A tool you already pay for has an AI feature that fits | It must read from and write to your own systems |
| Volume is low and per-seat pricing is affordable | Volume is high and per-seat or per-task fees add up |
| The vendor's standard data terms are acceptable | You need control over where data goes and how long it is kept |
| Good enough is good enough | You need review steps and accuracy you can measure |
There is also a middle path: connect your existing tools with an automation platform and add one AI step. Our comparison of n8n, Zapier and Make and our list of AI tools for small business are worth checking before you commission anything.
How to run the selection
- Write a one-page brief: the process, volumes, systems involved, what success looks like and your budget range.
- Collect 20 to 50 real examples (emails, documents or questions) with personal data removed, plus the answers you would expect.
- Send the brief to three firms and ask the questions below in writing or on a call.
- For larger projects, pay for a short discovery phase or pilot at a fixed fee before you commit to the full build.
Our guide to hiring a web development agency covers references, proposals and payment terms in more depth.
15 questions to ask an AI development company
For each question, here is what a good answer sounds like and the red flag to watch for.
1. What problem are we solving, and how will we measure it?
Good answer: they restate your process in their own words and propose a metric with a baseline, such as hours per week or the share of documents processed without edits. Red flag: they talk about models and features before asking how the work is done today.
2. Would an off-the-shelf tool do this?
Good answer: they name one or two products and explain where those fall short for you, or tell you to buy one. Red flag: every problem needs a custom build.
3. What similar work can you show, and can we speak to a past client?
Good answer: relevant examples, a demo on data like yours and a reference you can call. Red flag: only generic chatbot demos, with confidentiality given as the reason for every gap.
4. Which model will you use, and how hard is it to switch?
Good answer: they compare two or three models on your examples and keep the model choice in configuration, so a switch means a settings change and a re-test. Red flag: one provider for everything, or a proprietary platform with no export.
5. How will you test accuracy before launch?
Good answer: a test set built from your real examples, a pass rate reported before launch, and the same tests re-run after every change. Red flag: "it works well in our demo" with no numbers.
6. What happens when the AI is wrong or unsure?
Good answer: confidence checks, rules in code for things like totals and dates, human review for unclear cases, and a clean handoff to a person. Red flag: a promise of 100% accuracy.
7. Where does our data go, and is it used for training?
Good answer: they name every provider and plan tier, the hosting region and the retention period, and offer a data processing agreement. Red flag: free tiers or personal accounts used for client data, or vague answers.
8. How do you handle security risks such as prompt injection?
Good answer: least-privilege access, approval before the AI takes actions, checks on its output, and familiarity with the OWASP Top 10 for LLM Applications. Red flag: they have not heard of prompt injection.
9. Who owns the code, prompts, test sets and data?
Good answer: you do, with the code repository and cloud accounts in your name from day one. Red flag: you get a license to their platform and nothing you could take elsewhere.
10. What will it cost to run each month?
Good answer: a per-unit estimate (per document, chat or lead) at your volume, covering model usage, hosting and monitoring, plus spending alerts. Red flag: only a build price. Our LLM API costs guide shows how usage adds up.
11. Is the price fixed, and what is in scope?
Good answer: a written scope with milestones, assumptions and a change process; see fixed price versus time and materials. Red flag: a vague scope with a large upfront payment.
12. Who will actually do the work?
Good answer: named people, including a senior engineer you can talk to, with any subcontractors disclosed. Red flag: a senior sales team that disappears after signing.
13. How will it connect to our systems?
Good answer: they name the APIs and access they need, check rate limits early, and have a plan for systems without a good API. Red flag: "we will figure it out" for your oldest system.
14. What does support look like after launch?
Good answer: a warranty period for bugs, a monthly plan with response times, monitoring, and a plan for when a provider retires a model. Red flag: "it will not need maintenance".
15. What happens if we stop working together?
Good answer: documentation, a handover session and every credential already in your accounts. Red flag: hosting, API keys and domains held in their accounts.
How to compare quotes
Quotes for the same brief can differ widely because they include different things. Put them side by side on these lines:
| Line item | What to check |
|---|---|
| Discovery and design | Included, or billed first? |
| Build and integrations | Which systems are in scope and which are not |
| Evaluation | Test set size, target pass rate, re-testing |
| Deployment | Whose cloud account, which environments |
| Running cost | Monthly estimate at your volume |
| Support | Warranty length, monthly fee, response times |
| Payment schedule | Tied to milestones you can verify |
Compare the first-year total, not the build price: build cost plus 12 months of running and support. A low quote that skips evaluation or monitoring usually costs more later. Our guide to calculating AI automation ROI shows how to check that the numbers pay back.
Contract points that protect you
Code and data ownership
The contract should assign you the code, prompts, test sets and configuration on payment, and list them as deliverables. If an open-weight model is fine-tuned for you, the weights belong on that list too. Your data stays yours and is deleted from the vendor's systems at the end. Our guide to protecting IP when outsourcing software covers the clauses.
Model and vendor lock-in
Ask for the model provider to sit behind a thin layer in your code, with the choice kept in configuration, and for data stored in standard formats you can export. If the vendor builds on its own platform, get a perpetual license or the source code.
Data privacy
Name the providers and plan tiers in the contract, because terms differ. As of 16 October 2026, OpenAI says data sent to its API has not been used to train its models since 1 March 2023 unless you opt in. Anthropic says inputs and outputs from its commercial products, including its API, are not used for training by default. Google's Gemini API terms say content sent to unpaid services can be used to improve Google's products and may be read by human reviewers, while prompts and responses on paid services are not used to improve its products. Add a data processing agreement, retention limits and a breach notice period.
Evaluation and acceptance criteria
Write the test into the contract, for example "at least 95% of fields correct on our 200-invoice test set". State who builds the test set, how long you have to accept the work, and what happens if the target is missed.
Support
Agree a bug warranty after launch, the monthly support fee and response times, and who handles provider changes such as model retirements and price changes.
A simple scoring sheet
Score each firm from 1 (poor) to 5 (excellent) on each line, multiply by the weight and add up. In a spreadsheet with weights in B2:B8 and one firm's scores in C2:C8, use =SUMPRODUCT(B2:B8,C2:C8).
| Criterion | Weight | Firm A (example) | Firm B (example) |
|---|---|---|---|
| Understands your process and goal | 20% | 4 | 3 |
| Relevant work and references | 15% | 3 | 5 |
| Test and acceptance plan | 15% | 5 | 2 |
| Data privacy and security | 15% | 4 | 3 |
| Ownership, lock-in and exit | 10% | 5 | 2 |
| First-year total cost | 15% | 3 | 5 |
| Support and communication | 10% | 4 | 3 |
| Weighted score | 100% | 3.95 | 3.35 |
In this example, Firm B is cheaper and has the stronger portfolio, but Firm A wins on testing and ownership, which matter most once the tool is live. Treat a score of 1 on privacy or ownership as a deal breaker, whatever the total.
If you would like a second opinion on a brief or a quote, send it to us. We build custom AI tools on fixed quotes, you own 100% of the code, and we will tell you when an off-the-shelf tool is the better buy. You can also browse the AI tools we build to see typical scopes and starting prices.
Choosing an AI development company: FAQ
How much does it cost to hire an AI development company?
It depends on scope, integrations and testing. As one reference point, Apex Logic's published ranges as of 16 October 2026 are $1,500 to $10,000 for a custom AI chatbot or LLM integration, $500 to $1,500 for a single automation workflow, and $3,000 to $25,000 for a custom web app or SaaS MVP. Add the monthly running cost. Our AI chatbot cost guide breaks the numbers down.
How long does it take to build a custom AI tool?
At Apex Logic, a chatbot or a single AI feature typically takes 2 to 4 weeks once the scope is agreed. Larger tools with several integrations take longer, and the slowest part is often getting access to your systems and sample data, not the AI itself. Ask each firm for a week-by-week plan with a demo at each milestone, so you can see progress early.
Should I hire freelance AI developers or a company?
A strong freelancer can be a good fit for a small, well-defined tool. A company adds cover for holidays and illness, more skills such as design and security, and usually a support plan. Either way, ask the same 15 questions, insist on code in your own repository and check references. The real risk is a missing test plan or exit plan, not the size of the supplier.
Do we own the AI model if we pay for custom development?
Usually you own the application, not the underlying model. Most business tools call a model from a provider such as OpenAI, Anthropic or Google under that provider's terms. What you should own is everything built around it: code, prompts, test sets, configuration, your data and, if an open-weight model is fine-tuned for you, the resulting weights. Write that list into the contract.
What are the biggest red flags when hiring AI developers?
Promises of 100% accuracy, no plan to test on your data, quotes with no monthly running cost, code or accounts kept in the vendor's name, client data on free tiers or personal accounts, and pressure to sign before a written scope exists. One of these deserves a hard question. Two or more is a good reason to keep looking.
Is a paid pilot worth it before a full AI project?
For anything beyond a small, well-understood tool, usually yes. A short pilot at a fixed fee, run on your real examples, shows how accurate the approach is, what it will cost to run and how the team works before you commit the full budget. Agree in advance which result leads to the full build and which does not, so the decision is not made on feel.
Loading comments...