We assess AI products against the regimes they operate under. We write the governance record a regulator or auditor will ask for and design the evaluations that prove the system does what you say it does.
Next obligationAI Act · Art. 5(1)(ba)–(bb) · Art. 111(4)
11days · 2 Dec 2026
Art. 50 transparencyIn force · 2 Aug 2026
Instruments we assess against
EU AI Act
EN 18286
ISO/IEC 42001
ISO/IEC 42005
NIST AI RMF
OSFI E-23
Québec Law 25
PIPEDA
NIST AI 600-1
ISO/IEC 23894
EDSTA 2024
EU AI Act
EN 18286
ISO/IEC 42001
ISO/IEC 42005
NIST AI RMF
OSFI E-23
Québec Law 25
PIPEDA
NIST AI 600-1
ISO/IEC 23894
EDSTA 2024
Start now
Compliance for every AI Act you answer to.
The EU’s Digital Omnibus pushed the high-risk rules back to December 2027 and August 2028. What you have to prove did not change. Records of where training data came from and how it was checked must be kept while the work happens. Test results have to build up across releases. Neither can be recreated in the month before a deadline.
Some rules already apply. The Article 50 transparency duties took effect on 2 August 2026. Machine-readable marking of generated content follows on 2 December 2026. Treating the delay as breathing room is how teams miss them.
2 Aug 2026
Art. 50 transparency
Applied on schedule
2 Dec 2026
Machine-readable marking
Of generated output
2 Dec 2027
Annex III high-risk
Deferred from Aug 2026
Practice
We do three things.
Assessment, documentation and assurance. We have no software to license.
01
Compliance assessment and product remediation
We look at the application as built: its models, data flows, human review steps and the points where it affects a person. We work out which regime applies in each market you sell into. Then we tell you what has to change in the product as well as in the policies.
The regulator, the notified body, the enterprise buyer and the auditor all ask for the written record in different words. We write it in the structure the instrument specifies. We build it to survive the next model version without a rewrite from scratch.
Evaluation design, metrics and the audit programme
Most AI risk assessments stop before anyone proves the system behaves as claimed. We design the evaluation: metrics, datasets, thresholds and cadence. Then we set up an audit process that produces evidence on a schedule instead of during a fire drill.
These are the obligations that change what a product team has to build. Each day count is computed from today's date.
In forceEuropean Union
657days ago
Prohibited practices and AI literacy
AI Act · Art. 5 · Art. 4
The eight prohibited practices became directly applicable. The AI literacy duty for providers and deployers took effect on the same date. The Omnibus has since recast it as a duty to take measures that support AI literacy. It no longer requires any specific level.
In forceEuropean Union
476days ago
General-purpose AI model obligations
AI Act · Ch. V · Ch. XII
Technical documentation, training-data summaries and a copyright policy. Models presenting systemic risk add evaluation and incident reporting. Governance structures and most penalties took effect at the same time. Fines on providers of general-purpose AI models under Article 101 apply from 2 August 2026.
In forceEuropean Union
117days ago
Omnibus enters into force
Regulation (EU) 2026/1744 · Digital Omnibus on AI
Published in the Official Journal on 24 July 2026. It defers the high-risk regimes and makes targeted simplifications. It also adds two prohibited practices to Article 5. They cover non-consensual intimate imagery and child sexual abuse material and apply from 2 December 2026.
In forceEuropean Union
111days ago
Transparency duties apply on schedule
AI Act · Art. 50
The duties cover deepfake disclosure, disclosure that a person is interacting with an AI system and the marking of synthetic audio, image, video and text. The Omnibus did not move this date. Generative systems placed on the market from 2 August 2026 must mark their output at once.
UpcomingEuropean Union
11days
New prohibitions and the marking deadline for existing systems
AI Act · Art. 5(1)(ba)–(bb) · Art. 111(4)
Two prohibited practices added by the Omnibus start to apply. They ban AI that generates intimate imagery of an identifiable person without consent and AI that generates child sexual abuse material. Generative systems placed on the market before 2 August 2026 must also mark their output in machine-readable form by this date.
UpcomingUnited States — Colorado
41days
Colorado's replacement AI law
SB 26-189 · C.R.S. 6-1-1701 et seq.
Signed 14 May 2026 to repeal and replace the Colorado AI Act. It drops the duty of care and deployer risk-programme obligations in favour of a narrower disclosure regime for consequential automated decisions.
· Moved from 30 June 2026
UpcomingCanada
161days
Model risk management for federally regulated institutions
OSFI Guideline E-23 · Model Risk Management (2027)
Scope widens from deposit-taking institutions to federally regulated insurers and branches. The model definition expressly includes AI and machine-learning methods. OSFI recognises that not every model it covers carries model risk.
UpcomingEuropean Union
376days
Standalone high-risk obligations
AI Act · Annex III · Art. 6(2)
The full Chapter III regime for Annex III systems: risk management, data governance, technical documentation, logging, human oversight, accuracy and robustness, conformity assessment and registration.
· Deferred from 2 August 2026
UpcomingEuropean Union
620days
High-risk AI embedded in regulated products
AI Act · Annex I · Art. 6(1)
Systems that are safety components of products already covered by EU product-safety legislation. Assessment follows the conformity route of the underlying sectoral instrument.
Certain deployers of high-risk AI systems will use this template for their Article 27 fundamental rights impact assessments. Feedback on the draft is open until 15 January 2027.
The regulation sets common rules for setting up and supervising AI regulatory sandboxes under the AI Act. SMEs and start-ups get priority access and free regulatory guidance and supervision.
The EU AI Act applies in Northern Ireland only where AI systems are closely integrated with physical products. Standalone software and online AI tools are generally out of scope.
Department for Business, Innovation, Science and Trade
New York law S.9051B/A.10379C bars companies from offering listed unsafe AI chatbot features to users under 18. Banned features include pretending to be a real person and promoting self-harm.
The Secretary of Homeland Security ordered DeepSeek products excluded from federal civilian agency procurements. Agencies must also remove DeepSeek products and third-party products developed using DeepSeek from their IT.
The decision publishes the reference of EN 18286:2026 on quality management systems. High-risk AI systems that conform with it are presumed to meet the AI Act requirements it covers.
California's SB 947 bars employers from relying solely on automated decision-making systems to fire or discipline workers. It also requires human oversight when such systems assist those decisions.
California State Senate
Method
How an engagement runs.
It runs in five phases. The first often makes the other four smaller.
01
Scope
We inventory the AI in the product, establish your role in each value chain and confirm which markets and regimes apply. Most engagements narrow here. The scope you feared is usually larger than the one you have.
1–2 weeks
02
Assess
Classification, gap analysis against the applicable provisions and a read of the controls you already run. We work from the system as built and with the engineers who built it.
3–5 weeks
03
Document
We draft the artefacts and you review them. Then we fix the places where the record and reality disagree. Finding those disagreements is the useful output. The document follows from it.
4–8 weeks
04
Evaluate
Evaluation design, the first full run and the thresholds that decide what ships. If you have no measurement capability yet we build the harness with your team instead of handing over a specification.
3–6 weeks
05
Sustain
The audit calendar, the monitoring plan and the review that keeps documentation current through model changes. Governance programmes usually fail at this stage. It is the part we hand over most carefully.
Ongoing
Deliverables
What you receive.
Each artefact is tied to the provision that requires it. None of them is a slide deck.
Regulatory applicability determination
AI Act Art. 6 · Annex IIIAssessment
AI system inventory and register
ISO/IEC 42001 §8Assessment
Gap register and remediation roadmap
AI Act Art. 9–15Assessment
Annex IV technical documentation file
AI Act Art. 11Documentation
AI system impact assessment
ISO/IEC 42005:2025Documentation
Risk management system record
AI Act Art. 9 · ISO/IEC 23894Documentation
EU declaration of conformity
AI Act Art. 47 · Annex VDocumentation
Post-market monitoring plan
AI Act Art. 72Documentation
Evaluation specification and metric register
AI Act Art. 15Assurance
Bias and subgroup analysis
AI Act Art. 10(2)(f)Assurance
Red-team and robustness protocol
NIST AI 600-1Assurance
Internal audit programme
ISO/IEC 42001 §9.2Assurance
Model validation report
OSFI E-23Assurance
Position notes
Recent changes and what they mean.
Our reading of the changes that matter most for the work in front of you.
EU AI Act
What 2 December actually changes
Two new prohibited practices start to apply. They target non-consensual intimate imagery and child sexual abuse material. The same date ends a grace period for machine-readable marking. That grace period only covers generative systems already on the market before 2 August 2026. Anything placed on the market since then has had to mark its output from the start. Check which side of that line each of your systems sits on.
Harmonised standards
The first harmonised standard is about quality management
EN 18286:2026 was cited in the Official Journal in October. It is the first standard that gives a presumption of conformity under the AI Act. It covers the quality management system that Article 17 requires of providers. ISO/IEC 42001 does not give that presumption on its own. A team with a 42001 programme should map it to EN 18286 before relying on it.
Canada
Still no AI act. Two bills now carry AI duties.
Bill C-36 would replace the privacy part of PIPEDA and adds duties for automated decision systems. Bill C-34 brings AI chatbots into a new online safety regime. Both were tabled in June 2026 and neither is law yet. Until they pass the binding instruments are PIPEDA and Law 25 section 12.1. OSFI E-23 joins them on 1 May 2027.
Questions
The ones we are asked first.
The deferral moved the deadline. It did not shrink the work. A conformity file describes a system that already exists. Data governance you did not record cannot be reconstructed afterwards. Evaluation evidence has to accumulate across releases. Teams that treated the original August 2026 date as distant found that the eighteen months they saw as buffer was the realistic build time. Meanwhile Article 50 transparency applied on schedule on 2 August 2026. Generative systems already on the market before then must mark their output by 2 December 2026. So you have live obligations regardless.
Possibly. Where you are incorporated is not the test. The AI Act reaches providers placing systems on the EU market. Under Article 2(1)(c) it also reaches providers and deployers outside the EU where the output produced by the system is used in the EU. For a Canadian company the binding constraints are more often domestic. They are PIPEDA and Québec's Law 25 automated-decision rights. Ontario's AI directive reaches you through contract terms if you sell to Ontario ministries or agencies. From 1 May 2027 OSFI E-23 matters if your customers are federally regulated financial institutions.
There is no sign of it. AIDA was part of Bill C-27. The bill died on prorogation in early 2025 and has not been reintroduced. The federal position is now the National AI Strategy released on 4 June 2026. AI-related duties are arriving through other bills instead. Bill C-36 would add automated-decision duties to federal privacy law. Bill C-34 would regulate AI chatbots. Neither is law yet. We would not build a compliance programme around an anticipated Canadian AI act. We would look closely at any programme built that way.
It depends what you need it for. Certification does not make you compliant with the AI Act. The Act has its own conformity route. Certification gives you an auditable management system and a credential that replaces bespoke buyer questionnaires. That credential is usually the commercial reason to pursue it. For enterprise procurement it is often worth it. With a statutory deadline the technical file matters more and the certificate can follow.
More of the process than most teams expect and less of the substance. Risk registers, audit calendars, control libraries and management review all transfer. The technical content does not. Classification arguments, data governance evidence, evaluation design and subgroup analysis need people who can read the model and the statute at once. We are usually engaged to supply that layer and to leave it with your team.
A description of the systems in scope, the markets you sell into and access to the people who built them. We keep the first fortnight conversational on purpose. We would rather narrow the scope early than bill for assessing systems that turn out to be out of scope. If we think you need less work than you asked for we will tell you.
Next step
Settle the scope before any programme.
One conversation is usually enough to tell whether you have a four-week problem or a six-month one. We would rather tell you it is the shorter one.