How to Choose an AI Contractor in Ukraine: 12 Questions Before You Start

"Build us some AI" in 2026 is heard as often as "build us a website" was in 2010. And just as often it ends in disappointment: a bot that confuses prices, a demo that never made it to production, or an invoice for $20,000 for something that comes together for $1,600. At MaxICo Labs we regularly take on projects "after someone else," so we put together a concrete filter: 12 questions worth asking any AI contractor before signing a contract, and the answers you should hear.

Why choosing an AI developer isn't a website tender

With a website everything is simple: there's a mockup, there's the build, and the result either matches the mockup or it doesn't. With an AI system it's different, and this is exactly where the familiar selection process breaks down:

• The result is non-deterministic. The same bot can answer the same question differently. Without a testing system (an eval set of scenarios), you'll learn about problems from your customers.
• Costs don't end at release. Every dialogue with an LLM costs money. A bot that "works" can burn $600/mo on tokens where $40 would do, if no one optimized the architecture.
• A demo is easy, production is hard. A chatbot prototype comes together in an evening. A bot that correctly handles 95% of real requests, doesn't invent discounts, and doesn't leak data is weeks of work.

So the main task at the selection stage is to tell apart the team that built a demo from the team that kept systems running in production.

Questions 1-4: experience and case studies

1. "Show me three projects that have run in production for over six months"

Not screenshots and not videos, but live links. Ask them to message the bot, pose an awkward question. If the portfolio is only "internal projects under NDA" with not a single public one, that's a reason to be wary. A solid AI agency in Ukraine always has 2-3 cases you can touch: look at how it's presented, for example, in our case studies.

2. "What does monthly maintenance of your typical solution cost?"

If the contractor can't name a figure, they didn't calculate it. A normal answer is specific: "a bot for 2,000 dialogues/mo is $30-80 on API plus $10-20 hosting."

3. "What happens when the model answers incorrectly?"

The right answer contains the words "fallback," "escalation to a human," "logging," "topic boundaries." If they tell you "GPT barely makes mistakes anymore," you're facing a salesperson, not an engineer.

4. "Who owns the code, the prompts, and the data?"

The answer should be "you, after the final payment," and it should be fixed in the contract. Separately, ask about access: API keys should be on your accounts, otherwise when the partnership ends you'll be left with a black box.

Questions 5-8: technology and process

5. "Why this model and this stack specifically?"

A good contractor explains the choice in terms of money and the task: "for classifying requests a cheap model at $0.30/M tokens is enough, we use the expensive one only for generating answers." A bad one says "we work on GPT because it's the best."

6. "How do you test answer quality?"

Listen for the word "eval." Before release the system should be run against a set of 30-50 real scenarios: typical questions, provocations, attempts to push the bot off-topic. Without this, "we tested" means "we clicked around ourselves for 10 minutes."

7. "Where is my data and my clients' data stored?"

You should hear specifics: which database, which jurisdiction the server is in, whether personal data flies into the model's API, whether there's anonymization. For niches with sensitive data (medicine, finance, legal) this is question number one.

8. "What will you make deterministic code, and what AI?"

This question filters out the most. Checking an order status, calculating shipping, sending reminders, that's ordinary code that works for $0. AI is needed where there's natural language and decisions. A contractor who shoves an LLM into every operation is building you constant costs out of thin air.

Questions 9-12: money and accountability

9. "How much will tokens cost at my volume?"

Give them your numbers: requests per month, average dialogue length. The contractor should calculate it with you or by the next day. That's 15 minutes of work, and if it "needs a week of research," the rest will be the same.

10. "What's the first paid stage and what do I get in hand?"

A healthy format is a pilot in 2-3 weeks with limited scope: one channel, 10-15 scenarios, real data. If you're offered only "the full project for $16,000 and 4 months" with no intermediate checkpoints, all the risk is yours.

11. "Who maintains the system after release and how much does it cost?"

Models get updated, APIs change, the business gets new products and rules. Ask about the support format: SLA, response time, what's included in the fixed rate. The answer "nothing breaks there" is untrue.

12. "What metrics will you consider success?"

For a support bot, the share of requests closed without a human (60-80% is normal). For an AI salesperson, the dialogue-to-lead conversion. For a content system, the time you stopped spending. If the contractor doesn't propose a metric themselves, they're not thinking about your result.

Red flags: when it's better not to start

• They promise "the bot will replace the whole sales department", AI amplifies the team, it doesn't replace it in the first month.
• They don't ask about your processes and data but name a price right away.
• All their expertise is "we took a prompting course."
• There's no contract, or the contract doesn't spell out rights to the code.
• They propose starting with the biggest and most expensive module instead of a pilot.
• To the question about token costs they answer "that's pennies" with no numbers.

A separate marker is the reaction to the word "no." Say that you want to postpone part of the work to a second stage or do it in-house. A healthy contractor will propose phasing and honestly say what can be cut without risk to the result. A contractor who pushes "nothing will work without the full package" is most likely selling you a budget, not a solution. The same goes for deadlines: "we'll do it in three days" for a system with integrations is either a template or a promise no one plans to keep.

Price benchmarks for the Ukrainian market in 2026

So that choosing an AI developer doesn't turn into a lottery, keep these ranges we see on the market:

Type of solution	Freelancer	Agency (Ukraine)	MaxICo Labs benchmark
Support/sales chatbot	$600-1,600	$1,600-6,000	from $1,000
AI agent with integrations (CRM, payments)	$1,000-3,000	$3,000-12,000	from $1,000 for the base setup
Parser / data collection	$300-1,000	$1,000-3,000	from $600
CRM with AI modules	rarely taken on	$6,000-30,000	from $3,000

A lower price isn't always worse, a higher one isn't always better. The question is whether the contractor passes the 12 questions above. The detailed logic of our pricing is on the pricing page.

What a healthy collaboration process looks like

In short, the normal sequence of working with an AI agency looks like this:

1. Audit (3-5 days). The contractor looks at your processes, data, and volumes, and honestly says where AI pays off and where a script is enough.
2. Pilot (2-3 weeks). A limited working setup on real data. You see the quality before you've paid for everything.
3. Production (2-6 weeks). Integrations, edge cases, eval testing, handover of access.
4. Support. Monthly monitoring of metrics and token costs.

This is exactly the scheme by which we build AI solutions for business, from bots to autonomous agents, and exactly why a pilot always precedes a big contract for us.

If you're currently choosing who to commission a chatbot or AI system from, start with a free 30-minute AI audit. We'll review your task, honestly say what should be AI here and what should be ordinary code, and give a budget estimate with numbers rather than "from this to infinity." Book here: maxicolabs.com/contact.