An honest guide to AI consulting: why most AI projects never deliver a result, how AI implementation actually works in practice, and how to choose a consultant.
According to McKinsey, roughly 88% of organizations use AI in at least some part of their operations. Yet only 1% describe their deployment as fully mature. Most of the AI consulting market lives in the gap between those two numbers.
The most common reason an AI project fails to deliver isn't the technology. ChatGPT, Claude and Copilot already work well today. The problem is almost always the same: nobody in the company clearly defined which specific problem AI was supposed to solve.
A company buys licenses, runs a training session, and a month later discovers employees are back to their old habits. Or the opposite happens: one enthusiastic employee builds an impressive prototype, it gets shown to leadership, earns some applause, and then quietly disappears into a drawer because nobody knew how to integrate it into real work.
This article isn't written to convince you that you need AI consulting. It's written to show what AI implementation for business actually looks like when it works, and where it most often breaks down when it doesn't. If you finish it and decide you don't need help, or that you can do this yourselves, that's a good outcome too.
The term "AI consulting" has become so widely used over the past few years that it has largely lost its meaning. It's worth separating three things that constantly get confused with one another.
Tools. A ChatGPT subscription for an employee is not an AI strategy. It's simply a license. It can raise individual productivity, but it doesn't by itself change how an organization operates. Most companies stop at this stage: they have tools, but no plan. A few months in, it becomes clear that every employee uses the tool differently, nobody shares what actually works, and the organization as a whole hasn't moved any closer to systematic AI adoption.
Agents. An AI agent is something else entirely: a system you give a goal to, which then plans its own steps, uses your data and tools, and carries the task through to completion. This is more powerful than a simple tool, but it requires a clearly described process to work correctly. The difference from a chatbot is simple: a chatbot answers a question according to a predefined script, while an agent decides for itself what actions to take to reach the stated goal.
Strategy. This is the level where genuine AI consulting for business actually operates: identifying specifically where in your organization AI would create real value, in what order to pursue it, which risks to avoid, and how to measure whether it's actually working. Without this level, tools and agents remain isolated experiments rather than a systematic shift.
Confusion between these three levels is the most common reason AI project budgets and expectations don't match up. Leadership, when talking about "AI strategy," often really means buying tools. The IT team, when talking about "AI implementation," often means a single specific agent, not a change in how the whole organization operates. When these two understandings don't align, a project looks like a failure even when it technically works, simply because a different outcome was expected than the one that was actually planned.
Good AI consulting rarely starts with technology. It starts with a question: where in your organization is time and money currently draining into manual, repetitive work that could be automated without introducing additional risk. Only after answering that question is it worth considering whether you need a tool, an agent, or a broader strategy.
Over the past few years working with businesses, four scenarios keep recurring. They rarely sound like "we want to implement AI." More often, they sound like this.
"We don't actually know what's going on." Leadership suspects employees are already using AI tools at work but has no clear picture of how much or how. This is called shadow AI: each employee individually decides how to use AI, and the organization learns nothing from it and doesn't understand where the real risks lie. As our co-founder Gediminas Buivydas wrote in a Delfi column, the biggest risk today isn't that employees use AI, it's that leadership doesn't know how they're doing it. In practice, this means a marketing specialist processes customer data through a personal ChatGPT account, an accountant feeds invoice information into an AI tool, and a lawyer asks AI to summarize a contract containing confidential terms. None of them are doing anything wrong on purpose; nobody ever told them where the line is.
"We don't know where to start." Leadership understands that AI could deliver real value but has dozens of ideas and no clear priority among them. The result: either nothing gets done because there are too many options, or the first idea that happens to come up gets pursued, not necessarily the one that would deliver the greatest return. This often plays out as one meeting deciding to try a chatbot for customer service, another meeting a week later deciding that automating reports matters more, and a month later both projects are still "in the planning stage" because nobody ranked them by actual return and implementation difficulty.
"We had a great prototype, but it went nowhere." One of the most common scenarios. A team enthusiast or an IT specialist builds a working AI solution over a weekend. It gets demonstrated, receives approval, and then stalls, because nobody calculated what it would take to make it work with real data, real load, and real employees every day. A prototype that works with ten test records and a system that works with ten thousand real customer records every day are two completely different engineering problems. This is exactly where internal experiments most often stall: the idea was right, but there was no plan for moving it from demo to daily work.
"The EU AI Act is already here, and we have no policy at all." Since February 2025, there's been an obligation to ensure AI literacy among employees, and enforcement in Lithuania takes effect in August 2026. Companies that already give employees AI tools but have no clear AI usage policy for business find themselves in a position where risk is growing faster than their understanding of it. This is often only noticed when a client or partner asks how the company ensures data security when using AI, and there simply isn't an answer.
These four problems share one common trait: none of them gets solved by buying another AI tool. They need a systematic approach.
The foundation of our work is the S.E.A. method: Strategy, Experimentation, Adaptation. This isn't a marketing term, it's a real 30-day AI plan where every phase has a specific goal and a specific risk if it isn't done well.
Strategy (week one). In this phase, we identify where in your organization AI would create the greatest value, and, just as importantly, where it's better not used at all. We review actual processes, data and team capabilities, not general market statistics. We talk to the people who actually do the work, not just leadership, because they're the ones who know where time is really being lost. The most common mistake at this stage: skipping it and jumping straight to a solution that looks interesting but doesn't solve a real problem. The output: a handful of concrete automation candidates, ranked by potential return and implementation difficulty.
Experimentation (weeks two and three). Priority ideas are tested against your real data, not a demo environment. This is where it becomes clear whether an idea that sounded good in a meeting actually works in practice: is the data clean enough, is the process stable enough, is the result accurate enough to be trusted. The most common mistake at this stage: testing too many ideas at once and never getting a clear answer on any of them, or testing with "clean" sample data that doesn't show how the solution behaves with real, messy information.
Adaptation (week four). The validated idea is implemented as a working solution, and the team is prepared to use it daily. This phase is frequently undervalued: the technology can work perfectly, but if the team doesn't understand how and when to use it, the solution goes unused even though it's technically fully functional. The most common mistake at this stage: considering the project finished once the system is launched, rather than once it becomes a normal part of daily work, with a clear owner and an understood way to report problems if they arise.
The outcome after 30 days isn't a promise that everything will be automated. It's a validated action plan grounded in your own data, not general assumptions about what AI "should" be capable of. In some cases, the honest outcome is this: a specific idea that looked promising turns out, after testing, not to deliver enough return relative to its implementation complexity. That's also a valuable outcome, because it saves months that would otherwise have been invested without a clear answer.
The honest answer to "how much does this cost" is: it depends. Not because we're avoiding the question, but because two seemingly similar projects can differ several times over based on four factors.
Data quality. If your data is scattered across five different systems that don't talk to each other, the first job isn't AI, it's data cleanup. This often takes longer than the AI solution itself. A company whose customer data is neatly consolidated in one CRM system starts from a completely different point than a company where the same information is spread across email inboxes, spreadsheets and four different applications.
Process clarity. An AI agent can only automate what you yourselves are able to clearly describe step by step. If a process today exists only in one experienced employee's head, it first needs to be distilled into clear steps, which sometimes reveals that the process itself has inconsistencies worth fixing before automating it.
Integration depth. A solution that works standalone gets built faster than one that has to connect to your CRM, accounting system and email all at once. Every additional integration is additional work and an additional place where something can break, so the value between a "standalone tool" and a "fully integrated agent" can differ several times over.
Compliance burden. A project that processes customer personal data or produces decisions with legal consequences requires more caution than an internal productivity tool. EU AI Act and GDPR requirements here aren't a formality, they're a real driver of scope: you need to plan how the solution will be explained, how actions will be logged, and how a person will be able to review or override the AI's decision when needed.
It's precisely because of these four variables that you won't find a price list on our website. Any fixed price published without context would either be irresponsibly low or artificially inflated, depending on which scenario actually applies to you. Instead, we start with a free 30-minute conversation, during which we review these four factors together, and you walk away with a clear understanding of what your specific case would actually require, even if you decide at that point to work with a different partner, or on your own.
Whether you work with consultants or handle this yourselves, several obligations already apply to you now, and they don't only apply to large corporations.
Article 4 of the EU AI Act requires organizations to ensure AI literacy among employees if they use AI tools at work. This obligation has applied since February 2025, and its enforcement in Lithuania is assigned to the Communications Regulatory Authority (RRT), taking real effect from August 2026. In practice, this means that if your team uses any AI tool at all, from Copilot to plain ChatGPT, you already need a clear policy on how that's done safely.
If personal data is processed, GDPR applies, and most business AI use cases fall into the limited or minimal risk category, which carries transparency requirements rather than prohibitions. This means, for example, that a customer must know when they're interacting with AI rather than a person, and that certain decisions affecting an individual must be capable of being reviewed by a human.
On the national context, it's worth watching the Ministry of Economy and Innovation's (EIMIN) initiatives and Lithuania's national AI strategy, which shape both requirements and funding opportunities. This regulatory layer changes faster than most companies can track, so it's worth periodically revisiting and updating your AI policy rather than writing it once and forgetting about it.
A practical first step you can take today, for free: our AI policy generator (currently available in Lithuanian) produces a company-specific AI usage policy in minutes, aligned with EU AI Act and GDPR requirements. It doesn't replace a lawyer for more complex cases, but it gives you a solid foundation to start from, and often makes clear exactly where your current practice already has gaps.
The market is full of people calling themselves AI experts after one weekend with ChatGPT. Over the past few years, the number of consultants offering AI services in Lithuania has grown faster than the number of specialists with real experience. A few practical guidelines for telling them apart.
Do they show concrete cases, not general promises? Anyone can say AI will "transform your business." Look for consultants who can show specific, verifiable examples: what the problem was, what the solution was, what real risk was assessed. If the answer to "show me a concrete case" is a general phrase about "innovation" and "digital transformation," that's a signal.
Do they publish under their own name? A consultant who writes publicly about AI topics under their real name puts their professional reputation on the line. That's different from an anonymous agency nobody can verify. It's worth checking whether a consultant has a verifiable professional history, not just a well-designed website.
Do they talk about risk, not just upside? Every AI project has both an upside and a risky side: data security, legal compliance, dependency on a single vendor, the possibility that the solution produces a wrong result. A consultant who doesn't bring this up first either doesn't understand the field or doesn't want to stop you before starting a bad project, because that wouldn't be good for their business.
Can they say no? The best signal: a consultant willing to tell you that your idea won't deliver a return, or that you don't need an AI solution at all for a particular problem that could be solved more simply. If every proposal ends with "yes, that fits perfectly," it's worth questioning whether you're actually being advised, or just being sold hours.
These four criteria don't point to one specific name. They simply help you understand, in that first conversation, whether you're talking to someone who genuinely does this work, or someone who's just following a market trend.
If you want to start on your own, here are three steps you can take this week:
If you'd like to review your specific situation with someone who does this every day, book a free 30-minute call. We'll go through your processes together and give you an honest answer on whether and where AI would actually create value, even if that answer is that now isn't the right time.