The FCA's Mills Review, published on July 6, 2026, is one of the clearest official signals yet that financial AI is moving from assistance toward delegated action. The review says retail financial services are shifting from human-led, episodic activity toward AI-enabled, continuous, and delegated services.
That matters for FX and macro trading because the same control problem appears as soon as an agent can monitor data, recommend action, or initiate a workflow. A model that comments on EUR/USD after US CPI is one thing. A model that keeps monitoring accounts, schedules follow-up tasks, reads files, and recommends portfolio changes is a different operating model.
Fit
Who should read this
FX teams, fintech builders, risk leads, and agent developers deciding how far a finance agent should be allowed to go.
Best first workflow
Read-only research: release monitoring, macro evidence packets, scenario notes, and post-event review.
Not the starting point
Unreviewed investment advice, account changes, live execution, or hidden model-to-broker loops.
What Happened
The FCA published the Mills Review as a forward-looking assessment of how AI could reshape retail financial services by 2030 and beyond. The regulator's press release identifies four large shifts: firm operations, consumer journeys, competition and market power, and fraud and cyber risk.
The most important phrase for agent builders is delegated services. The FCA publication page says future systems will not only assist people, but may recommend actions, initiate transactions, and execute decisions within agreed parameters. That is the point where a chatbot becomes part of a financial control plane.
The timing is not isolated. OpenAI's July 9, 2026 release notes introduced ChatGPT Work as an agent for longer tasks across connected apps, files, reports, spreadsheets, presentations, and scheduled tasks. OpenAI's Help Center also describes scheduled tasks that can monitor for changes and notify users. These are productivity features, not trading systems, but they show why financial-services regulators are focusing on autonomy, monitoring, and delegated action now.
Why It Matters for Trading Agents
Most early finance agents are framed as research tools: summarize a Federal Reserve statement, compare a policy-rate decision with prior guidance, prepare a pre-event checklist for USD/JPY, or explain why a surprise in Non-Farm Payrolls changed the rates narrative.
Those workflows still need controls, but the risk profile is manageable when the agent is read-only and the output is reviewable. The risk changes when the same agent can keep a standing mandate, monitor accounts, recommend financial products, trigger transactions, or route tasks into other systems.
The Mills Review is useful because it does not treat this only as a model-quality question. It frames the issue as a system question: who authorised the agent, what data did it use, what authority was delegated, what action was taken, and how can the chain be audited if the result is wrong?
1. Explain
The agent retrieves data and writes an interpretation. Human judgment remains central.
2. Recommend
The agent proposes next actions, scenarios, watchlists, or alerts based on current evidence.
3. Initiate
The agent drafts an action, starts a workflow, or prepares an order that still requires approval.
4. Execute
The agent acts within a mandate. This requires the strongest identity, authority, logging, and recourse controls.
The Control Stack
The Mills Review highlights foundations for agentic finance such as data, identity, authorisation, delegation, execution, accountability, and supervision. For a trading-agent workflow, those foundations translate into a practical control stack.
| Control | Question to answer | Trading-agent implication |
|---|---|---|
| Trusted data | What facts did the model use, and were they current? | Use source-linked release calendars, announcement rows, FX context, COT, commodities, and sessions instead of model memory. |
| Identity | Who authorised the agent and under which account or role? | Keep agent activity tied to a user, desk, environment, and tool permission set. |
| Mandate | What may the agent do without asking again? | Start with retrieve, summarize, and monitor. Treat execution, allocation, and customer-facing advice as separate permissions. |
| Approval | Which actions require a human or deterministic gate? | Require approval before orders, risk-limit changes, account actions, or regulated-advice-like recommendations. |
| Audit trail | Can the firm reconstruct the decision later? | Log tool calls, source URLs, timestamps, expected values, actual values, model output, approvals, and final action. |
The FXMacroData Angle
The strongest position for FXMacroData is not "let the model decide." It is "give the model a trusted macro evidence layer." That difference matters. In a delegated finance workflow, the model should not be the source of market facts. It should reason over market facts retrieved through controlled tools.
Trusted data workflow
1. Ask
What changed in macro conditions for a pair, central bank, or release window?
2. Retrieve
Use FXMacroData REST or MCP tools for calendars, announcements, FX, commodities, COT, and sessions.
3. Reason
Let the model explain the scenario while preserving source paths, timestamps, and known data gaps.
4. Approve
Keep execution, allocation, alerts, and account actions behind explicit gates.
That architecture stays useful even if model products change. A team may use OpenAI, Claude, Gemini, Grok, an open model, or an internal model. The macro evidence layer should remain stable: same calendars, same release histories, same dashboard links, same source-aware product surface.
This is especially important around event-driven FX. If an agent is writing about Federal Reserve policy-rate decisions, release-calendar risk, or cross-asset confirmation from gold and oil, it should be able to show what it fetched before it writes the recommendation.
Practical Readiness Checks
Before a finance or trading agent moves beyond research support, run a simple readiness review:
- Data: Does every market claim come from a current, traceable source rather than model memory?
- Scope: Is the agent limited to a narrow task, currency set, pair set, or account role?
- Mandate: Is the agent's authority explicit, bounded, revocable, and visible to the user or firm?
- Approval: Which actions are always human-reviewed before they affect money, risk, or customer outcomes?
- Logging: Can the team replay the data, prompt, tool calls, output, and approval state later?
- Fallback: Can another model or interface use the same macro data contract if the preferred agent product changes?
Sources
Source context includes the FCA's July 6, 2026 press release, the FCA's Mills Review publication page, the full Mills Review PDF, OpenAI's July 9, 2026 ChatGPT Work release notes, and OpenAI's Help Center note on scheduled tasks in ChatGPT.
