An EU AI Act Consultant has to do two things at once: read the regulation precisely and turn it into systems that run in production and survive an audit. I do this for banks, insurers and public bodies across Europe, in English and German, remotely from Frankfurt and Berlin with on-site phases wherever the client sits. This page explains how the articles map to a governance layer and what to do before December 2027.
Most organisations I meet have already read the AI Act. They have a memo from legal, a slide from risk and a list of use cases from IT. What they do not have is a system in production from which a supervisor can pull a decision and see its history. That gap is where I work. I am an engineer, not a lawyer; this is not legal advice. The legal basics are in my EU AI Act guide, the article-by-article series in the EU AI Act hub.
Who an EU AI Act Consultant is for: regulated organisations across the EU
The AI Act applies in the same wording in Vienna, Amsterdam, Madrid and Frankfurt, so this is not a page for Germany only. Companies outside the EU whose output is used in the Union have their own page: EU AI Act for UAE and GCC companies.
- Banks and other lenders. Creditworthiness assessment of natural persons is high-risk under Annex III point 5(b). Fraud models and service assistants usually are not, but still need a registry entry, a log and an owner.
- Insurers. Risk assessment and pricing in life and health insurance is high-risk under Annex III point 5(c). Claims routing and intake assistants are mostly lower risk and deserve proportionate controls.
- Public bodies. Systems for benefits, permits and case handling fall under several Annex III points, and authorities carry the fundamental rights impact assessment under Art. 27 as deployers.
- HR in all three. Screening, ranking or assessing people in employment decisions is high-risk under Annex III point 4, whatever the industry.
What an EU AI Act Consultant must bring: law reading plus production engineering
This is the point most consultancies miss. Law firms read the regulation well and stop at the memo. Technology consultancies build well and treat the regulation as a checkbox at the end. The AI Act does not work like that, because its high-risk obligations are engineering requirements written in legal language: logs that let a decision be reconstructed, oversight a human can actually exercise, accuracy that is tested, not asserted. An EU AI Act Consultant has to argue with the general counsel in the morning about whether a modification is substantial under Art. 25, and decide with the platform team in the afternoon which fields the evidence log needs so that Art. 12 holds at the next audit.
The method: every article maps to a component of the governance layer
We never start a project with the model. We start with four questions: Who owns this system? What is it allowed to see and do? How do we know afterwards what it did? Who can stop it, and how fast? The answers become a governance layer between the business applications and the models, described in my post on the governance layer. For the AI Act, its useful property is that each obligation lands on a component, not on a policy document.
| Obligation | Article | Governance layer component |
|---|---|---|
| Inventory and classification | Art. 6 with Annex III, Art. 25 | System registry: purpose, risk class, role, owner, expiry |
| Risk management | Art. 9 | Risk register per model, enforced by gateway policies and quality gates |
| Data governance | Art. 10 | Data zones red, gray, green; pseudonymisation with IDs; data sheets per data set |
| Technical documentation | Art. 11 with Annex IV | Dossier assembled at every release from registry, model card, data sheets, tests |
| Record-keeping | Art. 12 | Evidence log written by the gateway, append-only, per decision |
| Human oversight | Art. 14 | Tiered review: full, sampled, monitoring only; blind re-checks; escalation queue |
| Accuracy, robustness, security | Art. 15 | Quality gates on a business-owned evaluation set, adversarial tests, rollback plan |
| Conformity assessment | Art. 43 with Annex VI | Internal control run as a pipeline over the artefacts above |
| Post-market monitoring | Art. 72 | Monitoring on gateway and log: drift, out-of-scope attempts, disagreement signals |
Inventory and classification first
The first weeks go into finding every AI system, including the ones IT does not know about, and classifying each one: prohibited, high-risk under Annex III, subject to Art. 50 transparency duties, or none of these. The second question is the role. A bank that buys a scoring model and fine-tunes it on its own portfolio may have become a provider under Art. 25. Role, risk class and owner go into the registry, and no system gets a route to a model without a record.
Art. 9 to Art. 15: the high-risk core
Two choices in this block matter most. The Annex IV documentation under Art. 11 is assembled at every release from the registry, the model card, the data sheets and the gate results, because a dossier written once is outdated at the second deployment. And oversight under Art. 14 is tiered, full review for high-impact decisions and sampled review with blind re-checks for the rest, because the promise that a human checks every output does not survive two weeks at volume.
Art. 43 and Art. 72: before and after go-live
For most Annex III systems, conformity assessment under Art. 43 means internal control under Annex VI, which we run as a pipeline over the artefacts that already exist. After go-live, Art. 72 requires post-market monitoring under a plan; ours is monitoring configuration on the gateway and the log, with alerts to the business owner. For banks and insurers, Art. 17(4) lets the quality management obligation be met through the governance rules of financial services law, except for risk management, post-market monitoring and incident reporting, which is exactly what the layer covers. How the same components satisfy DORA is on my page on DORA and the EU AI Act together.
Timeline: what to do before December 2027
Prohibitions and AI literacy apply since 2 February 2025, GPAI obligations since 2 August 2025, transparency duties under Art. 50 since 2 August 2026. The Digital Omnibus on AI (Regulation (EU) 2026/1744) moved the high-risk obligations: Annex III systems from 2 December 2027, Annex I products from 2 August 2028. The Omnibus moved deadlines only; it removed no obligation.
Two duties already apply to almost every client: AI literacy under Art. 4 and the transparency duties under Art. 50. The high-risk date is two budget cycles away, and a credit or pricing model needs most of that time. The order I recommend:
- Now: inventory and classification of every system, including vendor products and shadow deployments; role decision per system; named owners.
- Next quarter: the governance layer at the scale of the first high-risk use cases: registry, gateway, evidence log, oversight tiers, kill switch.
- First half of 2027: risk register, data sheets and Annex IV dossier per high-risk system; quality gates defined with the business; vendor contracts updated so providers deliver what Art. 13 expects.
- Second half of 2027: internal control under Annex VI end to end, declaration of conformity, registration, post-market monitoring switched on, Art. 27 impact assessment where it applies.
- Throughout: AI literacy for everyone who builds, reviews or owns these systems.
Harmonised standards: where things stand as of November 2026
Many organisations wait for the harmonised standards from CEN-CENELEC JTC 21, because a standard applied in full gives a presumption of conformity under Art. 40. As of November 2026, to my knowledge, most of those standards are still in drafting, and waiting is not a strategy. The regulation anticipated this: Annex IV point 7 asks for the harmonised standards applied and, where none were applied, a detailed description of the solutions adopted to meet the requirements of Chapter III Section 2. Documented own solutions, traceable to each requirement, are the practical path today, and the table above is the start of that description.
How an engagement runs and what it costs in effort
I work as an AI consultant with a team of consultants, architects and developers who also build the systems. A typical engagement has five steps, and the order is part of the method.
- 1. Assessment. Inventory, classification, role decision and a gap analysis per system against Art. 9 to Art. 15, Art. 26 and Art. 50. A few weeks of a small team, workshops on site, analysis remote.
- 2. Target picture and plan. Which systems need the full high-risk programme, which need registration and light controls, what the layer must look like in your landscape, who owns what.
- 3. Governance layer. Registry, gateway, evidence log, oversight tiers, kill switch, integrated with your platform. A few months of an engineering team; it pays back with every further system.
- 4. First high-risk system through all gates. One system, real data, risk register, data sheets, Annex IV dossier, gates, internal control, monitoring. Its artefacts become the template.
- 5. Handover and scale. Your teams in risk, business and IT run the layer. The knowledge stays with you.
What it costs depends on the number of systems, the state of your data and how much of the layer already exists; the honest number comes out of the assessment, so I do not publish prices. Where the effort goes is predictable: the model is the smallest part, data governance and integration the largest.
Typical mistakes
- Compliance as a document. The risk management system is a Word file, the monitoring plan a slide. The auditor will ask to see the log, not the policy.
- Classifying only what IT knows about. The vendor tool in HR and the chatbot a business unit bought are in scope whether or not they were on the inventory.
- Missing the role change. Fine-tuning, rebranding or repurposing a bought system can make the deployer a provider under Art. 25.
- Boundaries in the prompt. “Do not use personal data” is an instruction, not a boundary. Boundaries belong in the gateway, the data zones and the permissions.
- Treating the Omnibus as a pause. It moved the high-risk dates and removed no obligation. Art. 4 and Art. 50 apply today, and late standards do not move the deadline.
About Umair Zaffar
I am the founder and CEO of sifamo GmbH, an AI consultancy for regulated industries with offices in Berlin and Frankfurt am Main. Founded in 2019, self-financed and without loans, 28 people today. I come from enterprise architecture and built large platforms for banks, insurers and the public sector for more than ten years before we focused on AI in production in 2023. Our clients include Commerzbank, VW Bank, Bosch, DHL, Zurich, VGH and the German federal administration. More on my AI expert and founder pages, under AI Governance, and for the Gulf on my page as an AI governance consultant in Dubai and the UAE.
Frequently asked questions
Do we need an EU AI Act consultant if we already have legal counsel?
Keep your counsel; you need them for the role questions and the contracts. What counsel does not do is design the evidence log, define the quality gates or integrate a registry with your platform. The best projects have both at the table from the first workshop.
We are a deployer, not a provider. Does this still apply to us?
Less, but more than most deployers expect. Art. 26 requires deployers of high-risk systems to follow the instructions for use, assign oversight to competent people, keep the logs, monitor operation and report risks and incidents. And if you modify the system substantially or put your own name on it, Art. 25 makes you the provider.
Can this run remotely across the EU?
Yes. Engineering, documentation and most of the analysis run remotely from Frankfurt and Berlin. Classification workshops, gate reviews and sign-offs need a room, and we plan those as on-site phases wherever you are in the EU.
Do we have to replace our models or vendors?
No. The layer is provider-agnostic: the registry names the provider and version, the gateway talks to whatever sits behind it, hosted or on premise. Switching a model is a registration change plus a run through the gates.
If you want to know where your systems stand under the AI Act and what has to be in place before December 2027, reach me via the contact page, by e-mail at contact@sifamo.com or on LinkedIn. If you prefer to read first, start with the EU AI Act guide.