Written by: Karol Sobieraj, Founder & CEO, Digital Colliers
Binance flipping the switch on agent trading for 300M+ registered users is the shot every retail broker and neobank has been waiting for, whether they wanted it or not. The interesting bit isn't the headline. It's that supervision now sits largely with the user, which means the venue still gets the subpoena when something goes wrong. If you run a trading product, your order table and your audit log have a new column to design, and you have less time than you think to design it well.
The three actors your schema now has to tell apart
Up to about six months ago, every order in your system had one meaningful actor: a human, authenticated, clicking a button. That's over. From here on out, a single account can produce orders from at least three distinct origins, and they behave differently under scrutiny.
- Human, authenticated, clicking the UI or hitting the API themselves.
- Agent, running under a first-party integration you built and supervise.
- Agent, running under a third-party the customer authorised, where you don't control the prompt, the model, or the guardrails.
The third one is the new problem. It looks like the customer to your auth layer, because the customer handed it a key or an OAuth grant. It does not behave like the customer at all. Your order table needs, at minimum, an actor_type, an agent_id, an agent_principal (which human authorised it), the authorisation scope, the authorisation timestamp, and the model or vendor identifier where you can get it. Your audit log needs the same fields on every state transition, not just at submission.
If you can't answer "was this order placed by a human, by our agent, or by a third-party agent the customer authorised, and under what scope" in a single SQL query, you are not ready for the questions coming.
Why the regulators will ask, and when
DORA has been in force across the EU since 17 January 2025, and it already expects you to trace incidents end to end across your ICT supply chain. An agent your customer authorised is now part of that chain from the regulator's point of view, even though you didn't procure it. EU AI Act Article 50 transparency obligations kick in on 2 August 2026, and the high-risk obligations follow on 2 December 2027. Trading recommendations delivered by an agent to a retail customer sit uncomfortably close to the high-risk line, and the fines reach up to €15M or 3% of global turnover for the serious breaches.
On top of that, the SCHUFA ruling from the ECJ in December 2023 already established that automated decisions with significant effect on individuals fall under GDPR Article 22. Trading on someone's behalf clears that bar without much argument. GDPR fines reach up to €20M or 4% of global turnover, and the supervisory authorities have shown they will use the automated-decision route when they want to.
The practical read: if a customer disputes a loss caused by an agent they authorised, and you can't produce a clean record of who placed the order, under what scope, at what time, using which model, you are the deep pocket in the room.
The cost of waiting six months
Retrofitting an actor model into an order and audit schema after the fact is genuinely painful. You have to backfill historical rows with best-guess values, you have to rewrite every downstream report that assumed one-actor-per-account, and you have to explain to your auditor why the first eighteen months of agent traffic are labelled "unknown." That last one is the expensive conversation.
Operators who move on this in the next two quarters get three things the laggards won't:
- A clean cutover date, so agent traffic is labelled from day one and the "unknown" bucket stays small.
- Real telemetry on how much of their flow is already agent-driven, which you probably underestimate today.
- A supervision story they can hand to the FCA, BaFin, or the SEC without a nervous pause.
Where to actually start this quarter
You don't need a committee. You need a schema change, a token-issuance flow that captures agent metadata at authorisation time, and a monitoring rule that flags orders where the actor fields are null. That is a two-sprint job for a competent platform team, and it buys you the option to say yes to agent traffic on your own terms rather than someone else's timeline.
The operators I'd bet on in 2026 are the ones treating the audit-log column as a product decision, not a compliance task. The ones treating it as a compliance task will ship it in 2027, after the first enforcement action makes the news.