To choose an AI development company, check three things: proof they’ve shipped a live AI system like yours, a clear plan for your data and code ownership, and a fixed scope you can test before you pay in full. This guide gives you the exact questions to ask, the answers to listen for, and what a project should really cost.
Key takeaways
- Ask for a live system like yours, not slides or mockups.
- Make them explain how answers come from your data (RAG) and how they measure accuracy.
- Get an NDA and a DPA, and keep code, cloud, and AI accounts in your name.
- Know the real costs: Emergen builds AI chatbots for $1,500–12,000, and a typical RAG assistant costs about $30–60 a month to run.
- Start with a paid pilot or first milestone, not the full build.
About this guide. Written by Huzaifa Imran, Emergen’s CTO, and reviewed by Hashir Siraj, our CEO. We build AI assistants such as Abee, which has 600+ users at a hospitality business in France. Disclosure: we’re one of the vendors you might compare, so we publish our own prices and hold ourselves to the same checklist. AI prices checked 10 October 2026.
How to choose an AI development company in 5 steps
Step 1: Write down the job
One paragraph: who will use the AI, what it should answer or do, and where your data lives. Vendors can only quote accurately against a clear job.
Step 2: Shortlist 3 to 5 vendors
Pick teams that show live AI products, not only mockups. Fewer than three gives you nothing to compare; more than five wastes your time.
Step 3: Ask the same 15 questions
Use the questions below on a 30-minute call with each vendor. Write down their answers word for word.
Step 4: Score them side by side
Use the weighted scorecard further down. Same questions and same scale for everyone keeps the decision fair.
Step 5: Start with a paid pilot
Begin with a small paid pilot or first milestone. You learn how the team works before you commit the full budget.
Agency, freelancer or in-house team: which do you need?
| Option | Best for | Watch out for |
|---|---|---|
| Freelancer | One clear, small task, like adding a chatbot to an existing site | One person is a single point of failure. Check who covers testing, design, and support. |
| Agency or studio | A full product: design, build, AI, testing and launch | Some agencies pass work to a junior team or subcontractor. Ask who actually builds it. |
| In-house hire | AI is your core product and needs work every week | Hiring takes months, and you pay salaries before anything ships. |
Our take: for a first AI project, most small businesses don’t need an in-house team. An agency or strong freelancer gets you to a working version faster. Hire in-house once AI is central to your product.
15 questions to ask an AI development company
Ask every vendor the same questions in the same order. Vague answers tell you as much as good ones. Hiring for a chatbot? Pay extra attention to questions 4, 5 and 13: they show whether the bot will answer from your data or make things up.
Experience and proof (questions 1–3)
1. Can you show me a live AI system you built that is like mine?
Why ask: Slides are easy to make. A live product shows they have handled real users and real data.
Good answer: They open a working app and walk you through it, then offer a case study or a past client you can talk to.
Warning sign: Every example is “under NDA”, or all they have is a prototype.
2. Which parts did your team build, and which came from off-the-shelf tools?
Why ask: Most AI products mix libraries with custom work. You’re paying for the custom part.
Good answer: A clear split, for example: “The chat window is a library. We built the document pipeline, the search, and the admin panel.”
Warning sign: “We built everything from scratch”, or they can’t tell you.
3. What went wrong on a past AI project, and how did you fix it?
Why ask: Every real AI project hits problems. How a team talks about them shows real experience.
Good answer: A specific story, like “answers were wrong on scanned PDFs, so we changed how we split the text.”
Warning sign: “Nothing went wrong.”
Technical approach (questions 4–6)
4. How will the AI find answers in our data?
Why ask: This shows if they build real RAG (retrieval-augmented generation): the AI first looks up the right parts of your documents, then writes its answer from them.
Good answer: They describe the steps in plain words: upload documents, split them into chunks, turn chunks into embeddings (numbers that capture meaning), store them in a vector database, search for the closest matches, then answer from those matches.
Warning sign: “We connect ChatGPT to your files.”
Then go one level deeper. These follow-ups separate teams that have tuned a real system from teams that followed a tutorial:
- “What chunk size and overlap do you use, and why?” There’s no single right number, but they should have one and a reason.
- “Do you use keyword search as well as meaning-based search?” Mixing them (hybrid search) catches product codes and names that meaning-based search can miss.
- “How do you measure if answers are right?” Look for a fixed test set of real questions with expected answers, re-run after every change.
5. What happens when the AI doesn’t know the answer?
Why ask: A confident wrong answer costs you more than “I don’t know”.
Good answer: It says so, shows where its answers come from (sources or citations), and hands off to a person.
Warning sign: “It always knows the answer.”
6. Which AI model will you use, and can we switch later?
Why ask: Model prices and quality change every few months.
Good answer: They name a model, explain why it fits your budget and data rules, and build it so you can swap models without a rebuild.
Warning sign: Locked to one provider with no clear reason.
Data and security (questions 7–9)
7. Where will our data be stored and processed?
Why ask: Data location affects GDPR, your contracts and your customers’ trust.
Good answer: They name the hosting region and every outside service that touches your data. For UK or EU clients, they raise GDPR before you do.
Warning sign: “On secure servers”, with no details.
8. Will you sign an NDA and a DPA?
Why ask: A DPA (data processing agreement) sets out how a vendor may use personal data for you. Under Article 28 of the GDPR, you need one with any vendor that processes personal data on your behalf. The UK ICO explains what the contract must cover.
Good answer: “Yes to both, before we see any of your data.”
Warning sign: Hesitation, or “we don’t usually do that.”
9. Will our data be used to train anyone’s model?
Why ask: Some AI services can use inputs to improve their models unless settings or contracts prevent it.
Good answer: No. They explain the provider settings or contract terms that keep your data out of training, and offer a self-hosted model for very sensitive data.
Warning sign: They don’t know.
Pricing and contract (questions 10–12)
10. Is this fixed price or hourly, and what exactly is included?
Why ask: Budget surprises usually start with an unclear scope.
Good answer: A written scope with milestones and a price for each. Fixed price suits a clear scope; hourly suits open-ended research.
Warning sign: One total number with no breakdown.
11. Who owns the code, prompts, data, and any trained model?
Why ask: Ownership decides whether you can switch vendors later.
Good answer: You do, once you pay. The repository, cloud accounts, and API keys sit in your name.
Warning sign: “You get a licence to use the software.”
12. What will this cost to run each month after launch?
Why ask: The build price is only half the bill.
Good answer: A breakdown of hosting, model usage, and database fees based on your expected traffic. Compare it with the worked example below.
Warning sign: No estimate at all.
Delivery and support (questions 13–15)
13. How will we test it before launch?
Why ask: AI can shine in a demo and fail on real questions.
Good answer: They test on your real questions and show you the test list, the answers, the failures, and the fixes.
Warning sign: “We’ll test as we go.”
14. How often will I see progress, and who is my contact?
Why ask: Silence is usually the first sign a project is slipping.
Good answer: A named contact, a weekly demo, and a shared board or channel.
Warning sign: Monthly updates only.
15. What happens after launch?
Why ask: Documents, models, and user questions change, so answers drift over time.
Good answer: A support plan with bug fixes, model updates, and a set response time.
Warning sign: Support “as needed”, with no terms.
Red flags that should slow you down
- No live demos. Only slides, Figma screens, or “it’s under NDA” for every project.
- Promises of 100% accuracy. No honest AI team makes that promise.
- A price before they see your data. A quote in the first five minutes is a guess.
- They can’t explain RAG in plain words, or how they stop the AI from making things up.
- Code or accounts in their name, not yours.
- No mention of a DPA when you handle personal data from UK or EU users.
- A big upfront payment with no milestones you can test.
- “Fully autonomous” claims. Any AI agent that acts on its own needs limits and human checks.
How to score and compare vendors
Score each area from 1 (weak) to 5 (strong) and multiply by its weight. The maximum is 80. One rule matters more than the total: any vendor scoring below 3 on proof, technical depth or data security is out, whatever its total.
| Area | What you’re checking | Weight | Example vendor A | Example vendor B |
|---|---|---|---|---|
| Proof | Live AI systems like yours, references | 3 | 4 | 5 |
| Technical depth | Clear RAG plan, handles “I don’t know” | 3 | 4 | 2 |
| Data and security | Hosting, NDA, DPA, no training on your data | 3 | 3 | 4 |
| Ownership | You own code, accounts, prompts and data | 2 | 5 | 3 |
| Pricing clarity | Written scope, milestones, running costs | 2 | 4 | 5 |
| Communication | Named contact, weekly demos | 2 | 4 | 5 |
| Support | Plan after launch, response times | 1 | 3 | 4 |
| Weighted total | 62 / 80 | 63 / 80 |
Result: vendor B has the higher total but scores 2 on technical depth, so it fails the rule. Vendor A is the safer choice. Copy the table into a spreadsheet and score your own shortlist.
How much does an AI project cost?
Build cost
Prices vary a lot between vendors, so here are Emergen’s own published ranges as one data point for your comparison.
| Project type | What you get | Emergen price | Typical timeline |
|---|---|---|---|
| Website & Support Chatbot | Answers common questions on your site and hands off to your team | $1,500–3,000 | 1–2 weeks |
| RAG Knowledge Assistant | Answers from your own documents, with sources | $3,000–6,000 | 3–5 weeks |
| Enterprise AI Assistant | Connects to your CRM or HR system, with user roles and stronger security | $6,000–12,000 | 6–10 weeks |
Messy data, many connected systems, strict security rules, and mobile apps push a project toward the top of its range. For a clear scope, fixed price per milestone is usually safer for you than hourly.
Monthly running cost: a worked example
Here is what a typical RAG knowledge assistant costs to run, using list prices checked on 10 October 2026: 2,000 documents, 3,000 questions a month, about 3,500 tokens sent and 400 returned per question (a token is roughly ¾ of a word).
| Cost item | How it’s worked out | Budget model (GPT-5.6 Luna) | Stronger model (GPT-5.6 Terra) |
|---|---|---|---|
| Embed documents (one-off) | 4M tokens × $0.02 per 1M | $0.08 | $0.08 |
| Model input | 10.5M tokens × $0.20 or $2.00 per 1M | $2.10 / month | $21.00 / month |
| Model output | 1.2M tokens × $1.20 or $12.00 per 1M | $1.44 / month | $14.40 / month |
| Vector database | Supabase Pro plan | $25.00 / month | $25.00 / month |
| Total, before app hosting | ≈ $29 / month | ≈ $60 / month |
What this tells you: the AI model is rarely the high cost. If a vendor quotes hundreds of dollars a month for this kind of usage, ask what’s in it. Most of the ongoing cost is people: refreshing documents, reviewing wrong answers, and updating the test set. Sources: OpenAI API pricing, Supabase pricing. More detail in our guide to LLM integration cost and timeline.
How to brief vendors: a sample RFP outline
An RFP (request for proposal) is a short brief you send to every vendor, so they quote on the same thing. Two pages is enough:
- About us: what your business does, in two or three lines.
- The problem: what is slow, costly, or broken today.
- Who will use it: customers, staff, or both, and roughly how many.
- What it must do: 5 to 10 must-have tasks, e.g., “answer product questions from our 200-page manual and cite the page.”
- Your data: where it lives and whether it holds personal data.
- Rules and limits: GDPR, HIPAA, or others, and where data must be stored.
- Systems to connect: website, CRM, helpdesk, HR tools.
- Budget range and deadline: sharing a range gets better quotes, not higher ones.
- What you want back: scope, price per milestone, timeline, team, and two similar projects.
To cover AI risks in your brief, the NIST AI Risk Management Framework is a free, practical reference.
How to outsource AI development safely
Outsourcing works well, including to a team in another country, when you control four things:
- Contracts: an NDA, a DPA where personal data is involved, and a clause that says you own the IP.
- Accounts: code repository, cloud hosting, and AI accounts in your company’s name; give the vendor access, not ownership.
- Access: share only the data needed, and use test or masked data where you can.
- Milestones: pay in stages tied to software you can try.
How Emergen answers these questions
We’re a software studio building web apps, SaaS, mobile apps, and AI integrations for startups and small businesses: 46+ products shipped, Top Rated on Upwork with 100% Job Success. Here’s how we answer the key questions:
- Proof (Q1): Abee is an enterprise AI assistant we built for a hospitality business in France, with 600+ users. It has a real RAG pipeline, connects to the client’s CRM and HR systems, and uses MFA and role-based access. Built with Flutter, Node.js, and OpenAI models, including Whisper for voice. The client reports 50% faster task completion and zero security incidents since launch.
- Not sure (Q5): our assistants cite their sources and hand off to a person when they’re unsure.
- Data (Q8, Q9): we sign an NDA before we start; a DPA and self-hosted models are available.
- Testing (Q13): we test on your real questions before launch.
Frequently asked questions
Ask for a live system like yours, how the AI finds answers in your data, and what happens when it doesn’t know. Then ask who owns the code and data, where the data is stored, and what it costs to run each month.
No live demos, promises of 100% accuracy, a quote before they’ve seen your data, code or accounts kept in the vendor’s name, no DPA, and big upfront payments with no milestones.
Ask them to explain their pipeline in plain words and show it working. A real RAG team talks about chunking, embeddings, vector search, citing sources, and how they test accuracy.
For a first AI project, an agency or skilled freelancer is usually faster and cheaper. Hire in-house once AI is central to your product.
At Emergen, a website support chatbot costs $1,500–3,000, a RAG knowledge assistant $3,000–6,000, and an enterprise AI assistant $6,000–12,000. Running costs for a typical RAG assistant are about $30–60 a month before hosting.
At Emergen: 1–2 weeks for a support chatbot, 3–5 weeks for a RAG knowledge assistant, and 6–10 weeks for an enterprise assistant.
Yes, if you sign an NDA and a DPA, keep accounts in your name, share only the data needed, and pay by milestone.
Get a second opinion on your AI project
Comparing AI vendors right now? We’re happy to be one of them. Bring your use case, and we’ll tell you what it takes, what it costs, and whether you need AI at all.
Keep reading: AI chatbot development cost in 2026 · How to integrate GPT into your app · AI chatbot development services · AI integration services
