Three functions, three clocks

A conversation comes up whenever we consider a new market, and it rarely resolves cleanly.

Product asks whether the market is worth entering. Legal asks what we would be on the hook for. Engineering asks what it would take to build. Each function gives an answer that is correct on its own and incomplete next to the other two. Product can want a market that legal is still scoping. Legal can describe an obligation that engineering would need a year to satisfy. Engineering can build for a regime that never arrives.

One way to settle this is to pick one perspective and let it lead. That works when the ground is stable. Right now the ground is moving.

What changed this summer

Two things happened within a few weeks of each other, in opposite directions.

The EU adopted a simplification package that entered into force on 27 July 2026. It deferred the high-risk obligations in Annex III of the AI Act from 2 August 2026 to 2 December 2027, and product-embedded systems under Annex I to 2 August 2028. That is the expensive tier: risk management systems, conformity assessment, CE marking, quality management, registration, post-market monitoring. The stated reason was timing. The European standardisation bodies were still drafting the harmonised standards, and certification needs a published standard to certify against.

At the same time, Article 50 became enforceable on 2 August 2026, on its original date. It applies to any system that interacts directly with a person, whatever its risk tier, and it requires that the person be told they are dealing with AI. Generated content has to be marked. The ceiling is 15 million euros or three percent of worldwide turnover.

Canada moved the other way. AIDA died with Bill C-27 when Parliament was prorogued in January 2025. The responsible minister has said any future framework should be light. The live federal vehicles are privacy and digital safety legislation.

The binding obligations in Canada arrive through provincial privacy law. Quebec's Law 25 has been fully in force since September 2024, with penalties to 25 million dollars, and its automated decision provisions read a lot like the European ones: inform the person, explain the logic and the criteria, allow them to submit observations, provide human review on request.

"Are we compliant" has a shelf life

Set the two together and the useful lesson sits above either jurisdiction.

An organization that built the full high-risk apparatus for August 2026 spent it sixteen months early, against standards still in draft. An organization that built nothing missed a transparency duty that became enforceable with a real penalty attached. Both made the same mistake. Each treated a moving requirement as a fixed one and committed early in one direction.

"Are we compliant" has an answer with a shelf life measured in months.

The question that lasts

The question that survives the churn is a delivery question:

When the next substantial change lands, are we in a position to meet it, or do we leave that market?

Both answers are legitimate. Deciding that a market is worth leaving is a real strategic choice, and it is best made deliberately, early, and on the record. The outcome to avoid is discovering that a change is out of reach because of a decision made years earlier for reasons that had nothing to do with regulation.

Answering that question takes knowing where the load lands when the law changes, and keeping those specific points loose.

Where the load lands

For an AI feature inside an operational system, we watch a short list. Each item is a place where a rule change would put weight.

The model provider. Ours is a setting an administrator picks. It is stored in the database, and the rest of the pipeline works with any configured provider.

Where data sits, and which law governs it. These are two different questions. One instance and one database per customer, in the jurisdiction the customer chooses, keeps them answerable separately.

Whether a person is in the loop. The assistant proposes, and a person applies or discards. Human oversight appears in nearly every regime we looked at, and it is far easier to design in from the start.

What leaves the instance. Names, emails and phone numbers are replaced with deterministic tokens before a request goes out. The reverse map stays in the customer's own database.

What you can prove afterwards. An audit log records who did what and when. Every regime we read wants records. They differ on retention and format, and retention and format are easy to change once the records exist.

Disclosure. This is the one enforceable in Europe today, and it is also the cheapest to satisfy. It takes a string in each supported language and a marking on generated output. Our assistant labels every response with the name of the AI provider that produced it.

The organizational half

The list above is an architecture list, and architecture is the easier half. The organizational twin is where teams tend to be more locked down than they realise.

Does any customer agreement name a specific model provider? Does a signed data processing agreement pin a country you would then have to renegotiate to leave? If a regulator asked for documentation, would you generate it from the record, or would producing it be a project you had to staff? Who watches for the next Article 50, and how would they find out about it? And when a market is no longer worth it, who owns that call?

The hard part here is organizational.

The artifacts that prove you take governance seriously are often the same artifacts that reduce your ability to change. A certification has a scope. A contractual commitment has a term. A residency guarantee names a place. Each felt like rigour on the day it was signed, and each narrows what you can do next year. That tension resolves in a conversation where legal and engineering account for both sides together.

The last switch belongs to the customer

We sell operational software. The AI makes some of that work easier.

That has a design consequence we think is right in any regulatory climate. The assistant ships disabled on a new instance and stays off until an administrator turns it on. The customer holds that switch. If their own obligations change, their council takes a position, or they decide they would rather go without it, they turn it off themselves, with no release from us and no request.

What agile means here

Agility is usually discussed as a property of how a team plans its work. In an environment like this one it is closer to a comparison of two rates: how fast the ground moves, and how fast you can respond.

When the ground moves faster, the process has to change. Walking away from a market is a legitimate output of that comparison. The mistake is arriving at the decision without having noticed that it was being made.