Dear Reader,
Eurostat’s most recent enterprise digitalisation survey shows that 8.36% of Polish enterprises have deployed an AI-based solution by 2025. The PARP/UJ adoption study gives a higher figure for businesses of all sizes — 23%. By any measure, deployed AI in the Polish market is a minority of it. This concerns official deployments. The actual share of employees using AI at work is probably closer to 90% — but outside company control. That is a different topic.
The supply side does not look like a minority of the market. A scan of Polish AI consulting offers visible in 2026 on LinkedIn and in the digital space shows a delivery model that has become near-template: “we’ll build agents to improve your daily work” — an automation layer in n8n, Make.com or Zapier, three to five “agents” wired to commercial LLMs, two-day workshops for the staff. Pricing on the smallest engagements starts at 1,499 PLN per month. Asseco’s Academy and PFR’s Strefa Wiedzy now run public courses on the same template.
The gap between the volume of consulting on offer and the volume of AI actually deployed in Polish enterprises is the subject of this issue. The argument is not that the delivery model is fraudulent. It is that the model delivers what its practitioners learned to deliver — and skips the upstream questions that determine whether enterprise AI builds value.
This issue opens a six-issue series on what real AI transformation in non-AI-native enterprises actually looks like.
The delivery model#
What the delivery model sells is consistent across the catalogues: “let’s build an agent to automate your daily work”. The artefacts produced are workflows running on a low-code orchestrator, agents calling commercial LLM APIs, a Confluence or Notion page with usage instructions, training delivered to a small group of users described as champions. Engagement length is typically a few weeks.
This shape is calibrated to the simplest deployment contexts: one stakeholder owns the process, one user accepts the outputs, failure cost is bounded by one person’s time, and no integration is required with any system the consultant did not deploy themselves. In those contexts the shape is appropriate. Workflows ship immediately. The consultant maintains them while present. The user-as-owner can test and accept the outputs in real time.
The shape stops fitting when the same engagement is sold into a company of thirty to fifty or more people, an operations director or COO sponsors the work, an internal IT manager — sometimes none — owns whatever is left after the consultant rotates out, and the workflows touch systems and stakeholders the consultant never met during discovery. The issue is not that the company hired a solo consultant or small boutique instead of Big4. On the contrary — in my view, small, agile organisations have the greatest future in AI consulting. The gap is in experience: the large category of AI consultants who entered the market after 2023 has, by virtue of how recent the category is, mostly not been inside a large organisation through a full vendor migration or a regulated change-management programme. They deliver what they have seen work.
What the delivery model skips#
There is no discovery — no mapping of what processes exist, how they actually run, what variation they handle. There is no requirements engineering — no specification of inputs, outputs, exception cases, success criteria. There is no data review — no inventory of what data the workflow needs, what quality it has, where it lives. There is no measurement design — no defined way to know, in six months, whether the deployment worked. There is no enterprise architecture review — no assessment of how the proposed solution fits the company’s long-term infrastructure and obligations. The questions that are not being answered are:
Which processes to automate first. The choice of first use case determines what data assets exist for the second, which governance patterns are established, which integration challenges are solved. The argument was the subject of last week’s issue and will not be repeated here. The delivery model does not perform this selection. It automates whichever process the operations director points at, in the order the operations director thinks of them.
How to redesign the process for AI-native operation. Adding an LLM call to an existing workflow accelerates the workflow, including the parts of it that were broken before. The escalations the process was producing before AI now arrive faster. The exception cases the process never handled well are now mishandled at scale. The redesigned version of the process — the version that reorganises decision rights, removes intermediate handoffs, and uses AI where AI fits rather than as a wrapper around what was already there — is the version that captures the value.
Which architecture fits the company. Architecture is whether the workflow runs on infrastructure with single sign-on, secrets management, observability, audit logging, and ownership clearly assigned to a role rather than a person. A one-person consultancy does not need this. A five-hundred-person company with a Microsoft 365 estate, SAP installation, and three sectoral compliance regimes needs the architecture choice to determine whether the workflow survives the next system upgrade or the next employee departure. The “market” delivery model defaults to whatever the consultant deployed for their last client.
The downstream consequences of skipping these questions are the failures the adjacent literature documents. RAND’s 2024 study of 65 AI projects identified the leading root cause of failure as management misunderstanding how to set the project on a pathway to success — selection failure. Forrester’s tracking of robotic process automation has shown for over a decade that maintenance accounts for up to 60% of total programme cost; EY puts the failure rate of initial RPA programmes at 30-50% — architecture failure, in a technology category close enough to agent workflows to be the strongest published analogue available. The redesign failure is the one with the least published quantification, because the data is observational and the failures take months to surface. These are patterns known for years. The delivery model currently being sold does not address them.
What the delivery model produces#
The output of skipping the upstream work is reproducible across the engagements visible in mid-market consulting practice. The patterns surface after the consultant has rotated out.
I have heard of “vibe coding” companies claiming to CEOs that they can build an ERP system in 3 days, 3 weeks at most. Maybe the spec did not say it has to actually operate.
GenAI tools allow for very quick and effective prototyping and for delivering a magic-like impression in a demo. They also allow for legacy code refactoring, architecture and process discovery, and many other substantive advancements in the software development lifecycle. What they do not allow for is a “magic” implementation of a new system with a couple of prompts, without deep analysis of the processes, the data, and the architecture. If we do not redesign the processes to use AI properly, to ensure proper human control, to manage data in compliance with the law, we end up with a nice-looking prototype that does not deliver anything that moves the needle in business.
You may have seen the memes saying: “Claude, build a $1B company for me. Make no mistakes”. The vibe-coded ERP is just one level below that.
The AI technology vendors push for AI transformation defined as “buy 100 licences for our best-in-class copilot and you’re AI-native”.
Both examples are real.
I define three levels of AI automation:
- L1 — What competent generic training delivers: people prompt ChatGPT well, get +10-20% on individual tasks. Same processes underneath, same governance gaps. Where most of the “LinkedIn experts” and AI training programmes live.
- L2 — AI tools are implemented in the context of company processes and knowledge. RAG, graph, or ontology retrieval is in place, access control is in place, and people are able to use AI as a tool rooted in the context of their work rather than as a generic interface. This may give a 30-40% gain.
- L3 — We actually redesign processes, data structures and the organisation itself to use AI effectively. Processes run automatically where possible and ask for human intervention where we cannot rely on AI, and the organisation becomes AI-first or AI-native. This can give a three- to five-fold throughput increase for some processes.
The historical evidence is consistent with what we see in the field — RAND’s 2024 analysis on selection, Forrester and EY on the RPA architecture analogue. The failure narrative for this delivery model in mid-market enterprises is being written right now, in the engagements that started in late 2024 and early 2025.
Why the gap exists#
The supply side has filled rapidly since 2023. The demand side has continued to under-price the work.
The supply-side fill is straightforward: a category of consultants exists because a category of buyers wanted services priced at a fraction of what large consultancies charge. The buyers got the price they were paying for. What they did not appreciate is that the service they bought was a delivery model calibrated to the simplest deployment environments, sold into companies whose complexity the consultants had not previously navigated. Above a structural threshold — measured not by headcount but by process variation, multi-stakeholder ownership and integration complexity — the delivery model stops fitting. The consultants delivering it may well be competent. Just not in what their clients need.
Briefing#
Digital Omnibus on AI: provisional agreement reached 7 May 2026
EU Council and Parliament negotiators reached a provisional agreement on the Digital Omnibus on AI in the early hours of 7 May 2026. The agreement defers the application of the AI Act’s high-risk obligations: standalone high-risk AI systems classified under Annex III move from 2 August 2026 to 2 December 2027, and embedded high-risk systems under Annex I move to 2 August 2028. Article 4 (AI literacy) is being restructured — the Commission and Council proposals soften the original mandatory obligation on providers and deployers into an encouragement framework led by the Commission and Member States; the Parliament’s compromise retains a mandatory obligation but lowers the standard from “ensuring sufficient AI literacy” to “supporting improvement of AI literacy.” The obligation to train staff for human oversight in high-risk deployments remains.
The provisional agreement still requires formal endorsement by Council and Parliament, with adoption targeted before 2 August 2026 — the date on which the Annex III obligations would otherwise have applied. For Polish mid-market deployers, the new operating dates for Annex III high-risk obligations are 2 December 2027 and, for embedded high-risk systems, 2 August 2028 (Council press release, 7 May 2026; Bird & Bird analysis).
OpenAI and Anthropic both launched PE-backed enterprise AI services ventures on 4 May 2026
On 4 May 2026, OpenAI announced a $10 billion joint venture — provisionally named “The Deployment Company” — with TPG, Brookfield, Advent, Bain and fifteen other private equity investors. OpenAI raised $4 billion at a $10 billion pre-money valuation and retains majority ownership and governance control. The vehicle has built-in access to more than 2,000 PE portfolio companies and is in active discussions to acquire AI services firms, absorbing hundreds of engineers and consultants to help mid-sized companies deploy AI.
The same day, Anthropic announced a $1.5 billion joint venture with Blackstone, Hellman & Friedman and Goldman Sachs. Anthropic’s vehicle adopts Palantir’s forward-deployed engineer operating model, with the explicit target being mid-market portfolio companies in healthcare, manufacturing, financial services, retail and real estate.
The structural implication for the mid-market AI services market is direct. The model providers are entering distribution themselves, with PE backing, with portfolio-company access built in, and with engineers attached. Eighteen months from now, a Polish mid-market enterprise asking who should help us deploy AI has a third option to evaluate alongside the BigCo consultancy and the post-2023 individual operator: the model provider’s own PE-vehicle services arm, priced as a loss-leader for model adoption (Bloomberg, 4 May 2026; TechCrunch coverage). My aim is to build a fourth option — a small, efficient, flexible firm that understands both AI and the complexity of processes and procedures inside a large organisation.
Questions for your leadership team#
For each AI deployment currently running in your organisation: did anyone make a deliberate selection between this use case and others, with reasoning recorded? Or was the use case chosen because the operations director or a department head asked for it?
For the same deployments: was the underlying process redesigned for AI-native operation, or was an AI layer added to the process as it existed? If the process was producing escalations, exceptions or quality issues before AI, has the AI accelerated those, slowed those, or hidden them?
What is the architecture each deployment runs on? Specifically: does it access corporate data and how, how is that data secured, how does it integrate with other systems, and is it connected to monitoring and audit logging?
For each deployment: is it recorded in your register of processing activities, and has a data protection impact assessment been completed where personal data is processed by the model? Who has named responsibility for the literacy of staff working with the system?
Summary#
ING’s wholesale banking operation runs Katana, Katana Lens and Domino as ING-owned products. The Wholesale Banking Advanced Analytics team was built internally over years. McKinsey’s 2024 case study describes a seven-week joint build of a customer-service generative AI assistant; what the case makes clear is that the seven weeks were possible because ING had spent the previous decade building a Model Factory that democratised model-building, scaled across more than fifty support functions, and assigned named ownership of every model to roles inside the bank rather than to vendor engagements.
The seven weeks would not be possible without the work that preceded them.
The market model of many “AI consultants” sells the seven weeks. The work that preceded them — the selection, the redesign, the architecture, the named ownership — is the work these consultants cannot do. This approach harms clients, because it leads to dangerous and unstable solutions that do not deliver expected results — and it corrupts the market, because a non-technical CEO will come to believe that “we’ll build a new ERP in 3 days”. And how much can 3 days of work cost? Free is a fair price.
For the Polish mid-market enterprise considering its next AI engagement, the question to ask the consultant before signing is which of those four upstream activities is in scope. If the answer is a blank stare, the engagement is the delivery model this issue is about.
Stay balanced,
Krzysztof Goworek