In 2024 the Council of Europe opened the Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law for signature, the first international legally binding treaty on AI. It was signed by the European Union, the United Kingdom, the United States and Israel among others, and reported at the time as the moment the world acquired binding AI rules.
Two years later the honest status is uncomfortable: the treaty has not entered into force. In the meantime something else took over the job of governing AI inside companies, and it was not a treaty at all. This piece sets out where ratification actually stands, what is binding on businesses right now, and why the operative deadline for most companies arrives from a customer rather than a regulator.
What the Convention actually is
The Framework Convention is not a technical rulebook like the EU AI Act. It is a human-rights instrument: signatories commit to ensuring that activities within the lifecycle of AI systems are consistent with human rights, democracy and the rule of law. It sets principles and obligations at the level of the state, leaving each party to implement them through its own law.
Its significance was always as much diplomatic as legal. The signature list crosses blocs that agree on very little else in technology policy, alongside European states such as Andorra, Georgia, Iceland, Norway, Moldova and San Marino sit the United Kingdom, Israel and the United States, with the European Union signing as a bloc. Getting that set of names onto one binding AI text was the achievement.
Where ratification actually stands
Signature is the easy part. A treaty binds when it is ratified, and here the picture is thinner than the 2024 headlines implied:
| Milestone | Status |
|---|---|
| Opened for signature | 2024 (Council of Europe) |
| EU ratification | 15 May 2026 |
| Recent signatures | North Macedonia (8 May 2026), Albania (15 June 2026) |
| Required for entry into force | 5 ratifications, including 3 Council of Europe member states |
| In force? | Not yet, reporting indicates the threshold has not been reached |
The mechanism is standard treaty practice: once five ratifications including three Council of Europe members are deposited, the Convention enters into force on the first day of the month following a further three months. So even after the threshold is crossed, there is a built-in quarter-year delay before anything binds.
Two years of signatures and one bloc-level ratification is not failure, treaties routinely take longer, but it does mean that anyone who described this in September 2024 as "the world now has binding AI rules" was describing an intention rather than a legal fact. Signing a convention and being bound by one are different events, often years apart.
What is actually governing companies right now
While the treaty waits, the operative rules arrived from somewhere else entirely. For any business selling into Europe, the binding instrument is the EU AI Act, and its enforcement machinery switched on this year:
| Obligation | Applies from |
| Prohibited practices, AI literacy duties | 2 February 2025 |
| General-purpose AI obligations | 2 August 2025 |
| Transparency rules (Article 50) | August 2026 |
| Annex III high-risk systems (standalone, e.g. hiring, credit, education) | 2 August 2026 |
| Certain high-risk categories moved by the Omnibus | 2 December 2027 |
| Annex I high-risk (AI embedded in regulated products) | 2 August 2028 under the Omnibus timetable |
The penalties give it teeth: up to β¬35 million or 7% of global turnover for the most serious breaches. The EU AI Office has begun high-risk audits, and first enforcement actions are expected as national regulators build inspection capacity.
Note the last row, because it is the part most coverage still gets wrong. The Digital Omnibus political agreement in May 2026 moved certain high-risk obligations to 2 December 2027, the softening I covered in how the EU quietly delayed its own AI Act. Delay is not repeal, and the 2 August 2026 date has already passed for the Annex III systems it covers. The two dates are not in conflict: 2 August 2026 bound standalone high-risk systems, the Omnibus pushed a subset of those to December 2027, and AI built into products already regulated under sectoral law sits further out again, at 2 August 2028 on the Omnibus timetable. Which row applies to you depends entirely on which annex your system falls under, and that is the first question to answer before anything else.
The actual enforcement mechanism turned out to be procurement
Here is the development that matters more than any treaty for a working business, and it did not come from a legislature. ISO/IEC 42001, the certifiable standard for an AI management system, has moved, in the space of about a year, from an interesting credential to a line item in enterprise procurement questionnaires.
The logic is straightforward. The EU AI Act demands structured evidence that AI is governed: policies, roles, monitoring records, an audit trail. ISO 42001 packages exactly that into something auditable and certifiable. So large buyers, unwilling to assess each supplier's AI practices from scratch, began asking for the certificate instead. That converts governance from a compliance cost into a condition of sale.
This is a familiar pattern, it is how ISO 27001 became a de facto requirement in software procurement long before any law demanded it. The practical consequence is that a company can be entirely outside the EU AI Act's high-risk scope and still need documented AI governance, because its customers require it. That is the same dynamic I described for smaller firms in the four-part governance framework: the buyer's questionnaire arrives long before the regulator does.
The money: governance became a job market
The clearest evidence that this stopped being theoretical is payroll. According to the IAPP's 2025 report, dedicated AI governance roles grew 156% year over year, with Chief AI Officer titles becoming routine. Organisations do not create expensive new senior functions for problems they consider hypothetical.
Put the three money flows side by side, because they point the same way:
| Pressure | What it costs |
| Regulatory penalty | Up to β¬35M or 7% of global turnover (EU AI Act, most serious breaches) |
| Lost deals | Failing a procurement AI review, the contract, not a fine |
| Certification | ISO 42001 audit and preparation: real but bounded, and reusable across every customer |
| Staffing | Governance roles +156% YoY; a growing salary premium |
The asymmetry is what should drive the decision. The fine is a tail risk that most companies will never face. The procurement gate is a certainty for anyone selling upmarket, and it arrives on a specific date with a specific customer attached. Governance spending is best justified by the second row, not the first, it is closer to sales enablement than to insurance, which is also why the roles pay what they do, part of the wider skills premium reshaping AI careers.
What would change if it entered into force
It is worth being clear about the stakes, because "not yet in force" can read as "irrelevant," and that would be the wrong conclusion.
The Convention binds states, not companies directly. Entry into force would not create new obligations for a business overnight the way the EU AI Act did. What it would do is commit each ratifying party to bring its own domestic law into line with the human-rights, democracy and rule-of-law standards the text sets, and to maintain oversight and remedy mechanisms. The effect on companies arrives second-hand, through national legislation written to satisfy it, which is exactly why the timeline stretches into years rather than months.
Its more immediate value is as a common reference point across jurisdictions that otherwise regulate very differently. A treaty signed by the EU, the UK, the US and Israel is one of the few things that could pull fragmented national rules toward a shared floor. Whether that potential is realised depends entirely on ratifications that have not yet happened.
The evidence pack buyers actually ask for
Since procurement is the binding constraint for most companies, it helps to know what is being asked for. Enterprise AI reviews converge on a short list, and none of it requires a treaty:
| What buyers ask for | What satisfies it |
| Inventory of AI systems | A maintained register of every model, tool and embedded AI feature in use |
| Data flows | What each system ingests, from where, and where output goes |
| Human oversight | Documented points where a person reviews or approves AI output |
| Named accountability | One person responsible, with a review cadence |
| Incident and monitoring records | Evidence you would notice and could reconstruct a failure |
| Certification (increasingly) | ISO/IEC 42001, or a credible plan to obtain it |
The first four cost time rather than money and can be assembled by an existing team in a fortnight. Certification is the expensive item, and the honest test for whether it is worth it is empirical: count how many deals in the last two quarters asked for it. If the answer is none, build the documentation and wait. If the answer is more than one, the certificate has already paid for itself in deals you would otherwise re-fight from scratch each time. This is the same enterprise-adoption gate that governs every other trust-dependent purchase.
The rest of the 2026 calendar
The EU is not the only jurisdiction moving, and September 2026 is unusually busy. California faces a signature deadline on SB 1047, Brazil is voting on Bill 2338, and India has presented an AI liability framework. None of these is settled, and all of them would apply to companies far outside their borders through the usual extraterritorial reach.
The strategic reading is that the world is not converging on one AI rulebook. It is producing several partially overlapping ones, on different timetables, which is precisely the environment in which a single certifiable management system becomes attractive, you would rather implement one governance framework and map it to four regimes than build four compliance programmes. That fragmentation is also why the debate I covered in whether AI regulation helps or blocks has not resolved: the burden is now real, and so is the market access it buys.
What this means for you
If you sell to enterprises, treat AI governance as a revenue prerequisite rather than a legal one. Get the inventory, the data map, the written rules and the named owner in place, and decide whether ISO 42001 certification is worth it based on how often your buyers ask, that question is answerable by talking to your own sales team. The free starting materials in the AI transparency kit cover the documentation most questionnaires probe.
If you are already in scope of the EU AI Act, the August 2026 high-risk deadline has passed and the AI Office has begun audits. The relevant question is no longer whether obligations apply but whether you can produce evidence on request. Note also that transparency duties under Article 50 are live.
If you deploy AI agents, governance now has an operational security dimension that treaties do not address. Agents hold credentials and take actions, and the failure modes are concrete: I documented models reaching real companies' production systems through weak passwords, and the newest frontier models ship with explicit cyber-capability classifications and staged access. Inventory your agents and cap their permissions.
If you are choosing a career direction, the 156% growth figure is the signal. AI governance is one of the few functions where demand is documented, the skills are learnable without a research background, and the work sits at the intersection of legal, technical and commercial, the profile that pays.
The honest take
This is a useful corrective to how AI policy gets reported. The 2024 signing produced headlines about the world's first binding AI treaty. Two years later it is not in force, and the instruments actually changing corporate behaviour are a regional regulation with a shifting timetable and a private standard that customers demand in procurement. The gap between the announcement and the enforceable reality was roughly two years, and counting.
That is not an argument against the Convention, which may well matter a great deal once ratified. It is an argument about where to look. Treaties set direction; procurement sets deadlines. For a business deciding what to do about AI governance this quarter, the customer questionnaire is a more reliable guide to obligations than any diplomatic milestone, and unlike a treaty, it does not need five ratifications before it starts affecting your revenue.
Status of the Framework Convention reflects reporting available at publication; ratification counts change and should be checked against the Council of Europe treaty office before being relied upon.
Sources: Council of Europe: The Framework Convention on Artificial Intelligence; Council of Europe: EU ratifies the Framework Convention.
