Questions · for product, design and engineering leaders

You can’t build compliance after the fact.

EU digital regulation has changed how compliance works. It is no longer a tick in a box at the end. It’s procedural: it lives in how your teams design, build and run products, or it doesn’t exist.

Compliance in digital products

What is regulatory operations (RegOps)?

Compliance is not a layer added at the end. It’s a property of how the work is organised. RegOps is the work of organising it.

It sits between Legal and your product teams. Legal interprets the rules. RegOps builds them into the roadmap, the backlog, the design system, the release process and the evidence trail, so every release is compliant by default instead of checked at the end.

It grows out of design, research and product operations: the same discipline of designing how teams work, now applied to regulation.

What has changed in how EU regulation works?

The rules now regulate the process, not just the result.

The AI Act, GDPR and the European Accessibility Act ask you to prove how risks were found, how decisions were made and recorded, who could intervene, and how the product is monitored after launch. By my count, a high-risk AI system can need up to 49 documents, and many of them start in Figma files and product requirements.

  • Done used to mean it works. Now it means it works and we can prove it’s controlled.
  • Success includes override rates, complaints and fairness, not just conversion and speed.
  • Friction can be a safeguard, not a bug.
  • Every change is assessed before release.

I’ve written about this: The AI Act is not a compliance problem. It is a design problem.

Why can’t we fix it later?

Because it’s like trying to take the salt out of a cake. Once it’s in, you can’t remove it. You throw the cake away and start again.

  • You can’t add an override to a flow that never let a human in.
  • You can’t explain a decision the system never recorded.
  • You can’t prove a design choice nobody wrote down.
  • An inaccessible component in your design system spreads into every screen built on it.

Either it’s in the design, or it’s a rebuild.

Legal owns compliance. Why does product need to be involved?

Because most of what the law asks for is built by product, design and engineering.

A high-risk AI system has to do four things. Design and Product own three of them:

  • Protect people from harm: find and reduce the risks · Product, Data
  • Keep a record: document how the AI works and log its decisions · Engineering, Product
  • Keep humans in control: people can understand, override and stop it · Design, Product
  • Be reliable: accurate, fair and secure · Engineering, Data

Legal tells you what an obligation means. Your teams decide whether it exists in the product.

Won’t building compliance in slow our teams down?

No. Late compliance is what slows teams down.

Rework, delayed releases and last-minute exceptions all come from finding the rules too late. When controls live in the components, templates, checklists and pipelines teams already use, doing it right becomes the quickest way to ship.

Example: when accessibility became part of the development workflow at a global SaaS company, testing costs fell by 65%.

How I work

The goal: compliance becomes part of how your teams work, so everything you deliver is compliant by default, with the EAA, GDPR and the AI Act at least.

  1. Review your product roadmap against the AI Act, GDPR and the European Accessibility Act, and the Data Act and DORA where they apply.
  2. Review how you work today across design, product and engineering.
  3. Assess issues and risks in your products and your process.
  4. Report your current exposure: what is at risk, where, and how urgently.
  5. Define a remediation plan with owners, priorities and quick wins.
  6. Execute it with your teams, not for them.
  7. Embed it through design and product operations, so the controls live in the way teams already work and compliance is the path of least resistance.

Most work starts with step 1. Ask me for a readiness review →

The AI Act in practice

Is the AI Act relevant to us?

If you answer yes to any of these, it is.

  • Do you serve EU users?
  • Does your AI affect people’s rights, for example in hiring, credit, health or education?
  • Do you use general-purpose models, such as GPT or Claude, to deliver your services?
  • Do you use biometrics or emotion recognition?

It’s not about the AI you use at work. It’s about the AI you ship. Once a chatbot talks to your customers, or a model helps decide who gets credit, your screens, flows and requirements become the controls a regulator checks. Read the AI Act or the European Commission’s overview.

Does the AI Act apply if we’re not based in the EU?

Yes, if your AI reaches the EU. Where your company is based doesn’t matter.

  • It applies to providers that put AI on the EU market, wherever they are established.
  • It applies to providers and deployers outside the EU when the AI’s output is used in the EU (Article 2).
  • Providers outside the EU of high-risk systems or general-purpose models need an authorised representative in the EU (Articles 22 and 54).

The same logic applies to the European Accessibility Act if you sell covered products or services to EU consumers, and to GDPR if you offer services to people in the EU.

When do the AI Act rules apply?

In stages. Most of the rules for chatbots and generative AI already apply, and the next deadline is 2 December 2026.

  • Since February 2025: prohibited practices are banned, and staff who work with AI need AI literacy.
  • Since August 2025: rules for general-purpose AI models.
  • Since 2 August 2026: transparency rules for AI that talks to people or creates content (Article 50).
  • By 2 December 2026: AI-generated content from systems already on the market must carry a machine-readable mark (Article 50(2)). A new ban also applies to AI that generates sexual deepfakes of real people or child sexual abuse material.
  • 2 December 2027: high-risk rules for stand-alone systems, such as credit scoring or CV screening (Annex III).
  • 2 August 2028: high-risk rules for AI built into regulated products (Annex I).

Dates as amended by the Digital Omnibus on AI in 2026.

The high-risk deadline moved. Can we wait?

No. More time is not less work.

The Digital Omnibus on AI moved the high-risk dates to December 2027 and August 2028. But the transparency rules already apply, and classification, supplier evidence and documentation take longer than most plans allow. Systems you design today are the ones that will be checked.

Penalties: up to €35 million or 7% of global turnover for prohibited practices; up to €15 million or 3% for other breaches, including transparency.

What do chatbots, assistants and generative features need to do?

This already applies to every AI system that talks to people or creates content, not just high-risk ones.

  • Tell people it’s AI from the first message.
  • Agents too: an AI agent acting for a customer must say it’s AI, including to the other side.
  • Show when a person takes over, and when the AI comes back. Not required, but it builds trust.
  • Label realistic AI images, audio and video, and mark generated content in a machine-readable way.
  • Keep every notice clear, visible and accessible. Say it in voice channels, and never bury it in the terms and conditions.
How do we know if a system is high-risk?

Ask: does the AI output materially influence a decision about an identifiable person?

The uses listed in Annex III include:

  • Credit: the creditworthiness of individuals, not companies, and not fraud detection
  • Insurance: life and health risk pricing
  • Employment: CV screening, promotion, task allocation, monitoring
  • Education: admissions, grading, proctoring
  • Public services: benefit eligibility, emergency triage
  • Biometrics and critical infrastructure

Example from financial services: a company card limit set by AI is not high-risk. If the employee repays the card personally, the AI is now judging a person, and the same product becomes high-risk. It depends on who is being judged.

Are we a provider or a deployer?
  • Provider: you build or commission the AI and put it on the market under your name. You prove the system is safe and controlled: technical documentation, data governance, risk management, quality management, conformity assessment, CE marking, registration and post-market monitoring.
  • Deployer: you use someone else’s AI in your business. You prove you use it correctly and protect the people affected: impact assessments, human oversight, notices to workers and customers, logs and incident escalation.
  • Both: you build and use your own system, and carry both sets of obligations.

In my analysis of a high-risk system, that’s about 46 documents for a provider, about 21 for a deployer, and up to 49 for both.

Watch out: rebranding, substantially changing or repurposing a vendor’s AI for a high-risk use makes you the provider (Article 25).

What should we ask our AI vendors?

Ask early, and get the answers in writing. Vendors of high-risk AI must give you the information you need to comply (Article 25(4)).

  • What is the system for, and what is it not for?
  • Is it high-risk under the AI Act, and who is the provider?
  • Can we see the technical documentation, including for any general-purpose model underneath?
  • What are its accuracy, testing and bias results, and its known limitations?
  • What does it log, and can we access the logs for at least six months?
  • How do we oversee, override or stop it?
  • How will you tell us about changes, new versions and incidents?
  • If it processes personal data: do we have a data processing agreement, or a joint-controller arrangement?

Five vendor documents matter most: an AI and third-party asset register, supplier due diligence, the supplier agreement, the model’s technical documentation and the data protection agreement. Procurement usually leads, with Product, Engineering, Legal, Security and Privacy.

What does meaningful human oversight look like?

A human who rubber-stamps the output is not oversight. A reviewer must be able to:

  • understand the context and the limits of the AI output
  • challenge or override it, in practice
  • have the time, authority, training and information to judge independently
  • record why they accepted or rejected a recommendation
  • escalate error, bias, unsafe performance or an incident
  • recognise automation bias, the tendency to over-trust the machine

All of this is designed: review, override, stop and escalate screens, decision points in the flow, and evidence that reviewers really can intervene (Article 14).

I’ve written about this: When machines make outputs, humans must own outcomes · Decision theatre or decision design? · AI: incidents by design

What do designers and product managers actually produce?

For a high-risk system, most of it goes into the technical file and is kept for 10 years.

Product managers

  • Intended purpose, and what the AI is not for
  • Risk check with Compliance: is it high-risk, and are we the provider?
  • Misuse and harm scenarios
  • Oversight model: when the AI acts alone, and when a human decides
  • Acceptance criteria, including accuracy thresholds
  • Monitoring metrics: overrides, complaints, errors, drift
  • A change impact check on every release

Designers

  • AI touchpoint map: where AI talks to, or decides about, a person
  • Oversight screens: review, override, stop, escalate
  • AI label and notice components in the design system
  • Explanation screens for customers and staff
  • Challenge and redress flow
  • Accessibility check of every AI screen
  • Design change log: what changed, why, and what it means for oversight

Turn it into a plan: EU AI Act checklist for product teams →

How long do we keep the evidence?
  • 10 years after the system is placed on the market: technical documentation, quality records and the EU declaration of conformity (Article 18)
  • At least 6 months: automatically generated logs (Articles 19 and 26), or longer where sector rules say so
  • 15 days at most to report a serious incident, and less in the gravest cases (Article 73)

Every decision creates evidence. If you can’t retrieve it, you can’t demonstrate control.

The European Accessibility Act

Does the European Accessibility Act apply to us?

If you sell to consumers in the EU, probably. It has applied since 28 June 2025.

  • In scope: e-commerce, banking services, e-books, electronic communications and parts of passenger transport
  • Exempt: micro-enterprises providing services
  • The standard: EN 301 549, which points to WCAG 2.1 level AA
  • Also required: a public explanation of how your service meets the requirements
  • Timing: some existing services have transition periods, but every new release must comply now

Read the European Accessibility Act.

What does accessibility look like when it’s built in?

At a global SaaS company, accessibility went from no owner and no budget to a funded programme.

  • Risk assessed with Marketing, Product, Sales and Legal in the same room
  • Checks at each stage of development, and sign-off before release
  • Training across product, design, engineering, sales and marketing
  • Bids and marketing claims reviewed against what we shipped
  • Disabled users involved routinely, not once
  • A registry tracking conformance across products

Result: 75% conformance across the main products in 14 months, 65% lower testing costs and a 162% return in year one. Read the case study →

Does accessibility apply to our AI features too?

Yes, twice over.

  • AI interfaces in consumer services fall under the European Accessibility Act like any other screen.
  • Providers of high-risk AI systems must meet EU accessibility requirements under the AI Act itself (Article 16).

Oversight screens, AI labels and explanations have to work with a screen reader, a keyboard and plain language, or they don’t work at all.

GDPR, the Data Act and DORA

How do GDPR and the AI Act overlap?

Most high-risk AI uses personal data, so both apply.

  • GDPR protects individuals when personal data is used in consequential automated decisions (Article 22).
  • The AI Act governs the safety, transparency and lifecycle controls of AI systems that pose significant risks to people.
  • Assessments overlap: a data protection impact assessment and a fundamental rights impact assessment cover much of the same ground. Run them as one piece of work.

Watch out: meaningful human review can reduce your exposure under GDPR Article 22, but it doesn’t remove AI Act high-risk obligations.

Where do the Data Act and DORA come in?

I flag them when I review a roadmap, where they apply.

  • Data Act, since September 2025: users of connected products and related services must be able to access and share the data they generate, and customers must be able to switch cloud and data processing services.
  • DORA, since January 2025: financial entities and their critical technology providers need ICT risk management, incident reporting, resilience testing and third-party risk management.

Both change what you build and how you run it, not only your contracts.

Working with me

Where should we start with AI Act compliance?
  • Find the AI in your products, including features that come from vendors.
  • For each one, ask: does it talk to customers, or help decide about them?
  • Add oversight, labels and explanations to your next release.
  • Keep the evidence as you design, not after.
  • Bring Product, Design, Engineering and Legal together around one map of what applies to which product.

That last step is usually where I come in.

How can you help our organisation put responsible AI into practice?

Not with another policy. By building it into the way your teams already work.

  • Map where AI sits in your products and internal tools, including vendor features.
  • Classify each use, and agree who owns it.
  • Put the checkpoints, controls and evidence into your product lifecycle, from discovery to release and monitoring.
  • Give product, design and engineering the templates, components and training to run it themselves.

Governance that lives in a separate document gets skipped. Governance that lives in the backlog, the design system and the release checklist gets done.

How can product, design and engineering share responsibility for AI compliance?

By agreeing who owns which control, and making the hand-offs explicit.

  • Product: intended purpose, risk classification with Compliance, acceptance criteria and monitoring
  • Design: oversight, transparency and explanation screens, and accessibility
  • Engineering: logging, testing, accuracy and security, and the evidence trail
  • Legal and Compliance: what the obligation means, and whether the evidence holds up

I usually start with one shared map of obligations against the product lifecycle, so each team can see what’s theirs and what they hand on.

Can you help us turn GDPR and European Accessibility Act requirements into practical processes?

Yes. I’ve done both inside product organisations.

  • Accessibility: at a global SaaS company, checks at every stage of development, sign-off before release and training across teams.
  • GDPR: at a global fintech, a research participant service compliant with GDPR, PIPEDA and CCPA across seven markets. Read the case study →

The pattern is the same: requirements in the backlog, checkpoints in the workflow, evidence kept as you go.

What outcomes and deliverables can we expect?

Documents you can act on, and teams that can carry on without me.

  • A map of which rules apply to which products, and who owns them
  • A report of your current exposure, with risk ratings
  • A prioritised remediation plan with owners, and a business case where you need one
  • Controls built into your workflow: checklists, templates, design system components, release gates
  • Training for the people who will run it
Do you work alongside our legal, compliance and data protection teams?

Yes, always. I’m not a lawyer and I don’t give legal advice.

I work alongside Legal, Compliance, Risk and your Data Protection Officer, turning their interpretation into how the product organisation actually works.

How long does it take?
  • Advisory, 2 to 6 weeks: a roadmap review, an exposure report or a leadership briefing
  • Enablement, 3 to 12 months: remediation and embedding, because it changes habits, not just documents
  • Transformation, 12 months or more: governance and delivery at scale, often in a fractional or interim role

I prefer to start with one product, prove it works, then scale.

How do you price your work?

It depends on the shape of the work, and it’s agreed before anything starts.

  • Advisory: a fixed price for a defined scope, such as a roadmap review and exposure report
  • Enablement and fractional roles: a day rate, for an agreed number of days a month
  • Training: per session or per programme

You’ll find the price in the proposal, together with the scope and timing.

What is the difference between working with you as a consultant and in a fractional role?

As a consultant, I deliver a defined piece of work. In a fractional role, I’m part of your team and own the outcome.

  • Consulting: a fixed scope, such as a roadmap review, an exposure report or a remediation plan
  • Fractional or interim: a senior role for part of the week, such as AI governance lead, regulatory operations lead or DesignOps lead, working inside your teams

Fractional works best when the change needs someone inside the organisation to keep it moving.

What training do you offer on the EU AI Act, responsible AI and accessibility?

I teach the people who build products, using your products as the examples.

  • AI Act training for product, design and tech teams, most recently in-house at a Tier 1 global bank and as a Henry Stewart tutorial
  • Responsible AI workshops
  • GDPR and accessibility training

Where I can, I bring product, design, legal and risk together, because they read the same rule differently.

The EU AI Act workshop for product teams →

How does your DesignOps and R&D operations experience shape your approach?

Operations is how intent becomes repeatable practice. Regulation needs exactly that.

For over a decade I’ve built and run design, research and product operations, the whole of R&D operations: workflows, tooling, budgets, training, and the checkpoints that make good work the default.

Those disciplines now converge into regulatory operations. Regulation is one more requirement the operating model has to carry, and the one you have to prove.

Read the DesignOps case study →

How do we get started working together?

Email me a few lines about your organisation, what you’re building and what’s worrying you. I read every message myself and usually reply within two working days. We then have a short call to check the fit, and I send a proposal with scope, timing and price. Get in touch →

These answers are general information, not legal advice. Dates reflect the AI Act as amended by the Digital Omnibus in 2026. Last reviewed November 2026.

Is your product compliant? You’ll need to prove it.

Start with a review of your roadmap and the way your teams work today.