Fractional AI Architect: when it pays off

Most CTOs I talk to are stuck between two bad options.

Either they are being pushed by leadership to "do something with AI" and have no idea where to start. Or they started somewhere and made a few expensive wrong turns. The third group, the rare ones, hired fast and are now carrying a senior ML engineer who spends much of their time waiting for someone to decide what to build.

There is a gap in the market for regulated-industry companies with 5 to 30 people: law firms, medical practices, tax advisors, small regulated SaaS shops. Not because larger firms do not need help, but because at that size the owner is the decider and an audit from PLN 4,900 net fits in their head without a procurement committee.

This post names it.

In short. A fractional AI architect is one senior person who spends one to two days a week co-deciding AI architecture in your company. The model fits regulated firms of 5 to 30 people that have a clear AI use case but no architecture decision.

The hiring trap

When boards hear "we need AI," they think the answer is headcount. Hire an ML engineer. Maybe a data scientist. Build a team.

That instinct is understandable and almost always wrong for an SME in a regulated industry.

The hiring market for senior AI engineers is broken in Europe. Candidates with genuine production experience in RAG, inference optimization, and compliance-aware system design are rare. Most are employed by hyperscalers or large banks. What the market offers at your price point is usually a graduate with two years of notebook experiments and strong opinions about PyTorch.

I have been through this myself. In 2023 I was involved in hiring for an AI architect role on a team building language-model-assisted systems. Among the candidates, only a few had production deployment experience. The rest stopped at Kaggle and Hugging Face tutorials. That is not the candidates' fault. The market simply has not produced enough people who have built real AI systems in a regulated context.

Even if you hire well, you face the second problem: the field is moving faster than any individual can track. The architecture that was right for an internal knowledge assistant in January 2026 is already suboptimal in May 2026.

A full-time hire anchors you to their existing mental model. A fractional engagement by design rotates across clients and problems, which means current patterns arrive at your door, not six months later.

The third problem is scope mismatch. Most regulated SMEs do not need implementation capacity right now. They need to answer three questions first:

• What is safe to build given our regulatory obligations?
• What is the right architecture before we write a line of code?
• What does "done" look like for a compliance auditor?

A senior ML engineer is not the right person to answer those questions. A Solutions Architect with regulatory context is.

The consultant trap

So you skip the hire and call an agency. They send a deck with a roadmap. It includes "AI integration," "model fine-tuning," and a reference to a specific cloud API model. The proposal is a high figure for three months.

The problem is not the price. The problem is that most generalist IT consultancies lack two things that matter in regulated AI work.

First, EU regulatory fluency. AI Act (Regulation 2024/1689) Article 6 defines high-risk AI systems. Article 10 sets data governance requirements for training data. Article 12 requires automatic logging of system operation. Article 13 requires transparency and information for deployers sufficient for a human to understand and audit the output.

An agency that cannot walk you through those articles in concrete engineering terms is not equipped to make architectural decisions on your behalf. They will hand you a system that works in demo and fails a conformity assessment.

Second, domain-specific AI judgment. Last month I spoke with a CTO in the insurance sector who received an agency proposal for a RAG pipeline to analyse policy contracts. The solution assumed sending full policy content to an external cloud model API. For an insurance company that is not an option: client personal data, claims history, contract terms. All of it is subject to professional secrecy and requires an architecture where the model runs in an environment controlled by the firm, not by a vendor based outside the EU.

Generalist consultants pattern-match to their last project. That is how you end up with an expensive solution that fits no one.

What a Fractional AI Architect actually does

The fractional model is straightforward. One person, senior level, embedded into your decision-making process for one to two days per week, for a defined period or on a retainer basis.

The deliverables are concrete, not advisory theater. In a typical engagement I produce:

• An architectural decision record for your AI stack (what to build, what to buy, what to avoid and why)
• A data flow map showing exactly which data touches which system and where your regulatory exposure is
• A build-vs-buy analysis for each AI component you are considering
• A compliance readiness checklist mapped to the specific regulations that apply to your product (AI Act, DORA if you are in financial services, NIS2 if you are in critical infrastructure)

Hands-on implementation guidance is part of the role. I review the code your team writes. I prototype the tricky parts, typically the MCP orchestration layer, the RAG pipeline, or the inference configuration. I am on the call when your CTO needs to explain the architecture to the board or the regulator.

What I do not do is replace your engineering team. The fractional model assumes you have developers who can build. My job is to make sure they build the right thing, in the right order, with the right constraints in place from day one.

The cost structure is fundamentally different from a hire or an agency. No employer taxes, no notice period, no six-month ramp. You get a senior architectural opinion from the first week. If the engagement produces what it should, you will have made several decisions that save more than the retainer cost, before the retainer cost compounds.

Who needs one now

Three indicators tell me a company is ready for this model.

First: you have a clear AI use case but no architectural decision. You know you want to build an internal assistant, automate a document review step, or add AI-powered anomaly detection. You do not know whether to use a cloud API or run something on-premises, which model family, how to handle the audit logging, or whether your use case is high-risk under the AI Act. That combination, clear problem plus unresolved architecture, is the exact entry point.

Second: you have already started but something is wrong. You built a prototype that works in isolation but you cannot figure out how to take it to production while staying compliant. Or your team is building in circles. Or the original approach worked but does not scale to the data volume you actually have. A fractional engagement is often faster and cheaper than continuing down a path that needs correction anyway.

Third: your timeline is being driven by external pressure, not internal readiness. A large enterprise client is asking about your AI governance. Your regulator has issued guidance. Your board has set a deadline. These are not problems a six-month hiring process solves.

The clock

EU AI Act obligations for high-risk AI systems listed in Annex III start to apply on 2 December 2027, and for AI embedded in regulated products (Annex I) on 2 August 2028. For Annex III systems the earlier date was 2 August 2026; Regulation (EU) 2026/1744, in force since 27 July 2026, moved it. The dates are set in the regulation.

If you are running an AI-assisted credit scoring tool, an automated underwriting assistant, or an AI-powered document classifier in a regulated context, you need to have made your architectural and compliance decisions already. Not by the deadline, before the deadline, because the work to implement them takes time.

Article 12 (logging), Article 13 (transparency), Article 10 data governance, and the incident reporting obligations under Article 73 are not checkbox items. Each one requires engineering decisions that affect your database schema, your inference pipeline, and your audit trail architecture. None of that happens in a sprint if you have not made the foundational decisions first.

Companies that wait until autumn 2027 will not make the deadline. The geometry of the problem is that simple.

Next step

If any of the above maps to your situation, the lowest-friction way to find out whether a fractional engagement makes sense is the “Does AI make sense for your process?” consultation.

It is 60 minutes online with me. We look at one process and establish what is worth testing next. Within 3 business days you get a short written summary of conclusions and the next step. It costs PLN 299 total, and I credit the fee against an audit ordered within 30 days of the call.

Choose a time: https://artcode.software/#consultation

Prices: pricing annex. Scope: description of the offer.