How to choose an AI agency as a small or mid-size company: 12 questions before you sign

More companies are facing the question. In 2026, 34.7% of Lithuanian companies with 10 or more employees used at least one AI technology, against 21.3% in 2025. Among small companies (10–49 employees) the share was 30.8%, and among mid-size ones (50–249 employees) 46.7%.¹ Across the EU, companies that considered AI and then did not adopt it most often said they lacked the expertise: 70.9% gave that reason in 2025.² So most small and mid-size companies in Lithuania, the Baltics and the rest of the EU end up buying the expertise, from a market where "AI agency" covers very different kinds of provider.

In this guide a small or mid-size company (SME) means what EU and Lithuanian law say it means: fewer than 250 employees, and annual turnover up to €50 million or a balance sheet up to €43 million.³⁴ Such companies often have no separate IT department, and the AI decision falls to an owner or managing director who needs clear criteria more than another slide deck.

This guide is written by Vertara, an AI agency in Lithuania, so the conflict of interest is stated up front. The table and the questions work for choosing any provider.

The problem decides the type of provider. A bot that answers the phone, a chat assistant for staff and a system that posts supplier invoices into the accounting software need different skills, different pricing and different contract terms. The table lists the eight types Lithuanian SMEs meet most often.

If your problem falls into two rows, choose by its hardest part. A chatbot that tells customers where their order is turns out to be an integration project, because the answer has to come from the ERP. In that case ask the chatbot specialist exactly how they will connect to your system, or go to a custom builder. Vertara itself is a custom builder.

Where to find AI providers in Lithuania

paslaugos.lt has a category called "Dirbtinio intelekto sprendimai" (AI solutions), which listed 43 providers on 23 October 2026, mostly freelancers and small firms.⁵ Clutch and GoodFirms both have Lithuania AI categories. Clutch shows each firm's hourly rate and minimum project size, and states that it may earn a fee for some placements.⁶ Rekvizitai, the Lithuanian company register directory, has no AI category, but it lets you check any candidate's legal details, and its "Kompiuterių programinės įrangos kūrimas" (software development) and "Konsultacijų paslaugos" (consulting) categories turn up new names.⁷ LinkedIn shows who actually works at a firm and what they built before.

The European Digital Innovation Hubs (EDIHs) are worth knowing about too. The EU catalogue lists ten of them in Lithuania, and some offer SMEs free or state-aid-funded advice and "test before invest" services.⁸ A hub does not replace the company that builds your system, but it lets you test an idea at no cost before you buy.

Send the same 12 questions to every candidate in writing and compare the answers. Under each question you will find what a good answer sounds like and what should worry you. The more answers come back specific, the lower your risk.

Which single process do we start with, and what result in euros?

The first project needs a name, such as "supplier invoices into the accounting system", where a pitch would say "AI strategy". Today's cost of the process is simple to work out: time per case × cases per month × loaded hourly cost. That number is what you compare every quote against.

• Good answer: names one process, asks for your real volumes and gives you a calculation you can check yourself.
• Warning sign: proposes a "transformation" or several processes at once before seeing a single one of your documents.

Can I call two of your clients?

Logos on a website and unnamed testimonials prove nothing. Ten minutes on the phone with a company of similar size that automated a similar process tells you more than any presentation: how long the rollout really took, what broke and how the agency responded.

• Good answer: gives two contacts and lets you call. A new agency says openly that it is new and lowers your risk another way, with a short paid stage at a fixed price that you can stop after, or a test on your own documents.
• Warning sign: every client is hidden behind a confidentiality agreement and there is nothing you can verify.

Will it connect to our accounting software or ERP, and who builds that connection?

Lithuanian SMEs run Rivilė, Centas, Skaita, Business Central, Navision or another accounting or ERP system. The connection to it usually decides whether an AI system saves time or just adds one more place from which data has to be copied by hand.

• Good answer: explains how it will connect (API, database or file import), asks for a test environment or the API documentation, and says who maintains the connection when the system is updated.
• Warning sign: "we'll export to Excel", or "your ERP partner will handle it" without having asked them.

Who owns the code, the model and cloud accounts, and the data?

If the system runs on the agency's accounts, the agency controls it in practice. Agree in writing that the code lives in your repository, that the accounts with the model and cloud providers are opened in your company's name, and that you can export the data at any time.

• Good answer: code and accounts in your name from day one, model usage billed to your company, and the economic rights to the code transferred to you in the contract.
• Warning sign: "it's our platform", accounts in the agency's name, and handing over the code is something to "discuss later".

Where is the data processed, and will we sign a data processing agreement?

If the agency or the system it builds processes personal data about your customers or staff, GDPR Article 28 requires a written contract with the processor. It must state what data is processed and why, that the processor acts only on your documented instructions, which security measures it applies, which other processors it uses and what happens to the data when the service ends.⁹ If any of those processors is a US company, ask whether it is certified under the EU–US Data Privacy Framework. The framework still stands: the EU General Court dismissed a challenge to it on 3 September 2025, although that judgment has been appealed to the Court of Justice.¹⁰

• Good answer: names the processing region, lists every other processor (model API, hosting, email service) and signs the agreement before work starts.
• Warning sign: "it's all secure", with no region and no list.

Fixed price or time and materials, and what is not included?

Both models can be fair. A fixed price suits a clearly defined process; time and materials suits work whose scope is still unclear, as long as it has a cap. Most disputes come from what the price leaves out: licences, model usage fees, work by your ERP partner, data clean-up and changes after the scope is signed.

• Good answer: the price is tied to a named scope, exclusions are written down and extra work is priced in advance.
• Warning sign: a price "per AI agent" with no scope, or an hourly rate with no limit.

How will we measure accuracy, and against what?

A demo on sample data shows that the technology works in general and says nothing about how it will handle your documents. Accuracy has to be measured on a sample of your own cases whose right answers you already know, such as last month's invoices.

• Good answer: proposes shadow mode, in which the system runs on real data and its output is only compared with your staff's work; accuracy is measured field by field and the acceptance threshold is agreed in advance.
• Warning sign: "the model is 99% accurate", before it has seen one of your documents.

Which steps does a person approve?

An AI system that emails customers, posts invoices or changes prices needs points where a person signs off. Our guide to human in the loop approval tiers shows how to sort actions into tiers by whether they can be undone, how much money they move and whether anyone outside the company sees the result.

• Good answer: lists the system's actions, says for each whether it runs on its own or waits for approval, and explains what the system does when a check fails.
• Warning sign: "fully autonomous from day one".

What does support cost after launch, and what does it cover?

Costs do not stop on launch day. Separate two things: model usage and hosting, billed to your own accounts, and the agency's support, meaning monitoring, fixes and changes when document layouts change or the ERP is upgraded.

• Good answer: a monthly price with included hours and response times, a separate estimate of usage costs, and a clear statement of whether the system keeps running without a support contract.
• Warning sign: support left undefined, or a mandatory subscription you cannot cancel.

What happens if the person who built it leaves?

In a small agency or with a freelancer, one person often knows the system. That is fine as long as the knowledge is written down and another developer could take over the code.

• Good answer: documentation and a runbook, automated tests, code in your repository and a second person who knows the system.
• Warning sign: "it's all in my head", and no documentation.

Can we switch to another provider, and what do we get when we leave?

Exit terms are easiest to agree at the start, while both sides are still optimistic. For platform subscriptions, the EU Data Act has applied since 12 September 2025: cloud and other data processing service providers must let customers switch to another provider, and from 12 January 2027 they may not charge for switching at all.¹¹ Some of these rules do not apply to services that are mostly custom-built, so for a bespoke system the exit terms have to be written into the contract.

• Good answer: a handover package defined in advance: code, credentials, documentation, data in a standard format and a handover period.
• Warning sign: a fee to export your data, a closed format, or a platform the system cannot run without.

What do we have to do under the EU AI Act?

The EU AI Act applies to small companies too. Since 2 February 2025, Article 4 has required every provider and deployer of AI systems to attend to their staff's AI literacy. Regulation (EU) 2026/1744, in force since 27 July 2026, softened that duty: a company must take measures to support its staff's AI literacy, but no longer has to guarantee a set level of it.¹²¹³ Article 50 has applied since 2 August 2026: a chatbot or voice agent must be designed so that people know they are dealing with an AI system, unless that is obvious.¹²¹⁴

The obligations for high-risk systems, such as AI used to screen job applicants or score individuals' creditworthiness, have been postponed and now apply from 2 December 2027.¹³ Most back-office systems that read invoices, enter orders or draft quotes are not high-risk.

• Good answer: asks what the system decides and about whom, says whether the use could be high-risk, explains who counts as the provider and who as the deployer in your case, and plans how people will be told they are talking to an AI.
• Warning sign: "the AI Act doesn't apply to small companies", or the opposite, an expensive "compliance package" for an invoice-reading tool.

One of these on its own does not make an agency bad. If you see two or three, stop and ask again before you sign.

• An "AI strategy" with no named process. The proposal contains an audit, a roadmap and workshops, and not a single process that will run differently afterwards.
• No client you can call, and no other way to check. A new agency may have no references yet, but then it should offer a short paid stage you can stop after.
• A platform subscription you cannot leave. The system runs only on the agency's platform, and when the contract ends all you are left with is a paid data export.
• A price "per AI agent" with no scope. Nothing says what the agent does, which systems it works with or when the work counts as accepted.
• Borrowed certificates. The agency proves its security with a cloud or model provider's ISO 27001 certificate or SOC 2 report. Those documents cover the provider's data centres and services, not how the agency handles your access rights and data. Ask whose name is on the document.
• A demo on sample data only. If the agency will not try the solution on at least a few of your documents, you do not know how it will work in your company.

The contract should answer the same questions you asked during selection, in writing. If an answer sounded good on the call and is missing from the contract, treat it as missing.

Few prices are published, but there are enough to set expectations. AI providers on paslaugos.lt mostly list an hourly rate, with a category average of €46, and most Lithuanian AI firms on Clutch list USD 50–99 an hour. Scope decides the final price, so compare any quote with what the process costs you today.

On 23 October 2026 the summary for the "Dirbtinio intelekto sprendimai" category on paslaugos.lt showed prices "from €30–65", with an average of €46, and most listings quoted an hourly rate.⁵ Clutch's Lithuania AI page listed 50 firms the same day. Of the 45 that showed a rate, 35 listed USD 50–99 an hour and seven USD 25–49. Minimum project sizes ran from USD 1,000 to USD 100,000 and above: 18 of the 46 firms that showed one take projects from USD 5,000 or less, and 14 only from USD 25,000.⁶ The list includes international software houses with offices in Lithuania, so read it as a cross-section of the market more than a price list for AI agencies.

Clutch's global AI development pricing guide, updated on 21 September 2026, puts a typical AI development project at USD 10,000–49,999 and the average project at about USD 120,600.¹⁶ Those are worldwide figures from client reviews, not a Lithuanian average.

When you set a budget, keep three parts apart: the build, usage costs (model API and hosting, billed to your own accounts) and support. A useful check on the build price is the 60% rule that Vertara applies to its own quotes: the build price must fit within 60% of the conservative scenario's first-year return. You can hold any provider's quote to the same bar. The AI agency in Lithuania page explains how we model that return.

EU funding for AI projects in Lithuania

Grants can cover part of the cost, though not for every company. The Innovation Agency (Inovacijų agentūra) call "Dirbtinio intelekto sprendimų diegimas" (Implementing AI solutions, No. 02-126-K) is for industrial SMEs whose main activity falls under NACE section B or C, that have operated for at least two years and that had sales of at least €145,000 in each of the last two financial years. The grant is €15,000–70,000 and covers up to 50% of eligible costs, the project must be completed within 12 months, and applications close on 8 December 2026 at 17:00.¹⁵ The call funds an AI product or solution that is already on the market, plus its implementation, so a fully custom build may not qualify; settle that with the provider before you apply. An earlier call, "Skaitmeniniai MVĮ čekiai" (digital SME vouchers), which could pay for AI consulting up to €6,000, closed on 12 February 2025.¹⁷

Start from your own process, not from the providers' proposals. That way every candidate's answers are measured against the same baseline.

1. Describe one process and what it costs

   Who does it, how many times a month, how long one case takes and the loaded cost of an hour. That is the baseline for every quote.

2. Pick the provider type

   Use the table to decide which type you need. If the problem spans two types, choose by its hardest part.

3. Shortlist three candidates

   Search the directories and marketplaces, and check each candidate's company details on Rekvizitai.

4. Send everyone the same brief and the 12 questions

   A one-page description of the process and the same questions for all, answered in writing.

5. Start with a paid first stage

   An analysis or a pilot on one process, with a fixed price, an acceptance criterion and the option to stop afterwards.

6. Check references and the contract

   Call the clients and check the contract against the table above: scope, price, acceptance, ownership, data, support, exit.

Aurelijus Jakas

Founder · Vertara

Designs and builds custom AI systems for mid-sized manufacturing, logistics, export and professional-services companies in Lithuania and across Europe. Researched with AI tools; written and checked by the author. All articles