Ask a compliance officer what worries them about AI agents and the answer is rarely accuracy. They want to know what the agent can do without asking anyone.
“Human in the loop” is the standard reply, and on its own it doesn’t say much. In the loop for what? Every message? Every withdrawal? Every withdrawal above $10,000? An agent that needs approval for everything saves nobody any time. An agent that needs approval for nothing is a risk the firm has accepted without writing it down.
This guide gives a specific answer: a matrix of AI agent permissions that sorts the actions an agent can take in a forex brokerage into three tiers, function by function, with the reasoning behind each placement. It builds on our overview of AI and automation in forex brokerage operations, which set out the principle that the machine advises and the human decides. Here we get down to individual actions.
What regulators already expect
For investment firms in the EU, the clearest reference is ESMA’s public statement of May 30, 2024 on the use of AI in retail investment services. Four points from it shape any permission design:
- Firms’ decisions remain the responsibility of the management body, whether a person or an AI tool takes them. An agent’s action is the firm’s action.
- ESMA expects firms to keep comprehensive records of how they use AI and of related client complaints.
- The statement covers AI tools the firm adopts officially and third-party AI tools staff use, with or without management’s knowledge.
- When the AI comes from a third party, MiFID II outsourcing requirements apply.
Data protection law adds another layer. Article 22 of the GDPR gives people the right not to be subject to a decision based solely on automated processing that has legal or similarly significant effects on them, and where such decisions are allowed it requires safeguards that include human intervention. Rejecting a withdrawal or closing an account is the kind of decision you want a person to own.
If your entity is licensed outside the EU, none of this binds you directly. It’s still the most specific public description of what a financial regulator expects from firms using AI, and it’s a sensible standard to design against if you might license in the EU later. This article is operational guidance, not legal advice, so run your final matrix past your compliance counsel.
The three tiers
Kenmore’s Decision Engine sorts every agent action into three tiers, set in a plain-text rulebook that operations and compliance own.
| Tier | What the agent does | Kenmore’s published examples |
|---|---|---|
| May do | Acts on its own and logs the action | Pre-check KYC documents |
| Ask first | Prepares the action with evidence; a person approves or rejects it | Release a withdrawal, change leverage, open a demo account |
| Never | Refuses, whatever the request or channel | Give investment advice, promise returns, move funds, override compliance |
The middle tier does most of the work. “Ask first” doesn’t mean “the agent does nothing.” The agent gathers the data, runs the checks, drafts the action and hands a person a decision that takes seconds instead of the ten minutes of digging it used to take.
Five questions that place any action in a tier
You’ll meet actions this guide doesn’t list. Run each one through these questions.
- Does it move money or change financial exposure? If it does, the action starts in “ask first” at minimum. If it would move funds directly, the agent never does it; it prepares the transfer and a person executes.
- Can it be undone without the client noticing? Internal, reversible actions such as drafting a message, adding a note or tagging a record are usually “may do.”
- Does it change the client’s regulatory or contractual position? KYC status, client classification, leverage and account type all qualify. These are “ask first” at minimum.
- Could a reasonable client read the output as a personal recommendation? If so, it’s “never,” or you redesign the output so it’s clearly analysis of the client’s own past activity.
- Does it override a control someone else set? Compliance holds, risk restrictions and rule exceptions are “never” for an agent.
An action that clears all five is a candidate for “may do.” Start it in “ask first” anyway and promote it once you have approval data, as described further down.
The permission matrix, function by function
The tables below are a starting point. Your licenses, products and risk appetite will move some lines.
Trader support
| Action | Tier | Reasoning |
|---|---|---|
| Answer deposit, withdrawal and KYC status questions from live data | May do | Read-only, limited to the trader’s own account |
| Explain fees, swap and account settings | May do | Factual, based on published terms |
| Open a support ticket or send a document upload link | May do | Reversible, low risk |
| Open a demo account | Ask first | Kenmore’s examples place it here; some brokers move it to “may do” once they’ve reviewed the approval history |
| Change leverage | Ask first | Changes exposure and may be restricted by client classification |
| Release a withdrawal | Ask first | Money moves |
| Close an account | Ask first | Changes the client’s contractual position |
| Recommend instruments, entry points or position sizes | Never | Investment advice |
| Promise returns or outcomes | Never | Misleading by definition |
| Transfer funds between accounts | Never | The agent prepares; the trader or a staff member executes |
The last line surprises some brokers, because internal transfers are a normal self-service feature. The trader can still make the transfer themselves in the Trader’s Room. The agent just doesn’t do it for them.
Sales desk
| Action | Tier | Reasoning |
|---|---|---|
| Qualify and route new leads by language, geography and question | May do | Internal, follows your routing rules |
| Read notes and email threads to propose the next action | May do | Advisory only |
| Log a touch and schedule a follow-up | May do | Internal, reversible |
| Draft an outreach message | May do | Draft only, nothing is sent |
| Send a message to a lead or client | Ask first | A rep reviews, edits and sends |
| Offer a bonus or promotion | Ask first | Bonus terms carry compliance and abuse risk |
| Reassign a lead outside the routing rules | Ask first | Affects rep compensation and accountability |
| Contact someone who opted out | Never | Consent |
| Claim guaranteed or expected returns | Never | Misleading, and a conduct breach in most regulated markets |
Kenmore’s Sales Assistant works this way. In the example on the Agentic Forex CRM page, a Spanish-language registration from Mexico asking about swap-free accounts is qualified and routed to a rep with a drafted message in Spanish. The rep edits one line and sends it, and the Next-Step Engine logs the touch and schedules the follow-up. For the human side of the desk, see our guide to forex sales and retention desks. Bonus abuse is covered in fraud prevention for forex brokers.
Introducing brokers
| Action | Tier | Reasoning |
|---|---|---|
| Explain commissions level by level | May do | Read-only, the IB’s own data |
| Generate referral links and landing pages | May do | Uses approved templates |
| Show conversion data for the IB’s own traffic | May do | Read-only |
| Adjust a commission or pay a manual bonus | Ask first | Money moves |
| Move a client from one IB tree to another | Ask first | Changes who gets paid, often disputed |
| Approve an IB payout | Ask first | Money moves |
| Edit the commission plan itself | Never | Plan changes go through your normal approval process |
Commission questions suit an agent well, because the answer is arithmetic over data the CRM already holds. In Kenmore’s example, an IB asks why the commission was $312 instead of $400, and the IB Agent explains that a level-2 trader moved to a cent account, showing the rate and the lots. The multi-level IB module holds the data. Our guide on how to pay your IBs covers the payout side.
Payments and back office
| Action | Tier | Reasoning |
|---|---|---|
| Pre-check a withdrawal: PSP status, KYC, name match, open positions and margin, bonus terms | May do | Read-only checks |
| Flag a withdrawal with reasons | May do | Advisory |
| Release a withdrawal | Ask first | Money moves |
| Reroute a payment to another PSP | Ask first | Affects cost, timing and reconciliation |
| Change a client’s payout details | Never | A common account-takeover route |
| Execute any transfer of funds | Never | The agent prepares; a person executes |
This is where the “ask first” tier pays for itself. In Kenmore’s example, a back-office manager starts the day with 11 withdrawals that were pre-checked overnight. Nine are clean and approved in about a minute. Two are flagged, with reasons, and get her full attention. The approval stays with a person, and the digging is already done. For the rails behind it, see forex payment solutions.
KYC and compliance
| Action | Tier | Reasoning |
|---|---|---|
| Pre-check KYC documents for legibility, expiry and name and address match | May do | Kenmore’s own example of a “may do” action |
| Tell the client which documents are still missing | May do | Factual, own account |
| Approve a KYC application | Ask first | Unless your compliance policy explicitly allows auto-approval of clean cases, with thresholds you’ve documented |
| Change a client’s classification | Ask first | Changes protections and permitted products |
| Clear a sanctions or PEP match | Never | A person must review and record the decision |
| Override a compliance hold | Never | Overrides a control someone else set |
Whether clean KYC applications can be approved automatically is a compliance decision, not a technology one. Make it explicitly and write the thresholds into the rulebook. Our guide to KYC and AML workflows in forex CRM covers the queues this plugs into.
Risk and dealing
| Action | Tier | Reasoning |
|---|---|---|
| Flag deposit-withdrawal churn without trading, multi-accounting, bonus abuse, IB self-referral loops and unusual exposure, with evidence | May do | Advisory |
| Restrict an account or cancel a bonus | Ask first | Affects the client and is often disputed |
| Penalize a client automatically on the strength of a flag | Never | Flags produce false positives; a person weighs the evidence |
Kenmore’s risk monitoring flags each of the patterns in the first row and attaches the evidence. The flag goes to a person; the agent doesn’t act on it. Automated flagging is safe. Automated punishment isn’t.
Trader-facing analysis and agentic trading
Two features look alike and sit in different places in the matrix.

Trading History Analysis gives a trader a plain-language breakdown of their own history: what worked, what cost them money and where risk concentrated. That’s “may do,” because it’s analysis of the trader’s past activity. In Kenmore’s example, a trader asks for a review of the previous month and learns that most losses came from three Friday sessions and that hold time doubled after losses. The line is crossed the moment the output tells the trader what to trade next.
Agentic Trading, where a trader connects their own strategy through MCP and the system executes it, doesn’t involve the broker’s agent making a decision at all. The trader acts through their own scoped access, so the questions are about the trader’s account limits and the permissions on their MCP session. Our guide to MCP for forex brokers and prop firms covers how to scope that access.
What an approval request should contain
An approval screen that only says “Approve withdrawal #48213?” turns the approver into a rubber stamp. Each request should show the proposed action, the checks the agent ran, the data it used and the rule it applied, which is the evidence set Kenmore attaches to every one-click approval.
A few design rules make approvals hold up under audit:
- The approver can’t be the person who requested the action, and the agent can never approve its own proposal.
- Above a set amount, require a second approver.
- Rejections need a reason code, and the codes feed back into the rulebook review.
- Clean and flagged cases arrive separately, so attention goes to the flagged ones.
The last point protects against approval fatigue. A reviewer who sees 200 identical-looking requests a day stops reading them. A reviewer who sees 180 marked clean and 20 marked flagged, with reasons, reads the 20.
The audit trail
| Field | Why it matters |
|---|---|
| Requester (client or staff ID) | Who asked |
| Channel (trader portal, Telegram, Slack, back office) | Where the request came from |
| Agent and model version | What produced the answer |
| Rulebook version | Which rules applied |
| Data read | What the agent saw |
| Tool calls and inputs | What the agent did |
| Output shown to the client | What the client saw |
| Approver, decision and timestamp | Who decided, and when |
This is the record ESMA’s statement points toward, and it’s the one you’ll need the first time a client says “your chatbot told me.” Pair it with a kill switch that lets operations stop one action, one agent or the whole AI layer instantly. Kenmore’s agentic layer includes both a full audit log and a kill switch.
Test before you promote
Run new agents in shadow mode first. The agent proposes, your team works exactly as before, and you compare the agent’s proposals with what people actually did. Disagreements show you where the rulebook or the data descriptions need work.
Then try to break it. Ask it for a trade recommendation. Ask it to raise leverage “just this once.” Paste an instruction into a support ticket telling it to approve something. Ask about another client’s account. Every one of these should end in a refusal or an “ask first” proposal, never in an action.
Promote actions from “ask first” to “may do” one at a time. A sensible bar is a sustained run of approval history in which reviewers accepted the agent’s proposals unchanged, plus a compliance sign-off recorded as a new rulebook version. Demotion should be immediate and needs no approval at all.
Mistakes that show up in the first month
The first is permissions written only in the prompt. The model is told “never approve withdrawals” and also has a tool that approves withdrawals. A cleverly worded ticket or a model error is then all it takes. Remove the tool instead.
The second is one shared admin API key for the agent. Every action the agent takes then looks like an admin action in your logs, and the agent can reach every account. Give it scoped credentials per role and act as the user it’s serving.
Rules without versions make every past decision impossible to explain. When a client disputes something from three months ago, you need to know which rule applied that day.
The last one is easy to miss: shadow AI. While you design a careful permission matrix, staff paste client data into personal chatbot accounts to draft emails faster. ESMA’s statement explicitly covers staff use of third-party AI tools. Giving staff a sanctioned assistant that follows the same rulebook does more to stop that than a policy memo.
Frequently asked questions
Can an AI agent approve withdrawals on its own?
Technically, yes. It shouldn’t. Keep withdrawal release in “ask first” and let the agent do the pre-check, so the human decision takes seconds and stays a human decision.
Is a chatbot that answers account questions giving investment advice?
Answering factual questions about a client’s own account, fees and request status generally isn’t investment advice. Recommending instruments, timing or position sizes is. Draw the line in the rulebook, test it with adversarial prompts and confirm it with your compliance counsel.
Who should own the rulebook?
Operations and compliance, with engineering implementing it. Kenmore’s Decision Engine follows that model: a plain-text rulebook owned by the broker’s operations and compliance teams.
Does the matrix change for prop firms?
The structure stays the same and the actions change. Explaining a breach is “may do,” resetting a challenge is “ask first” and changing the rules for an individual trader is “never.” Our guide to prop firm breach disputes applies this to the prop side.
How many actions should start in “may do”?
Few. Read-only answers about the client’s own account, ticket creation, document links, drafts and internal notes. Everything else starts in “ask first” and earns promotion with data.
Request a Consultation on Designing AI Agent Permissions for Your Brokerage
Get expert guidance on deciding what your AI agents may do on their own, what they must ask first and what they must never do. We’ll help you build a rulebook your operations and compliance teams own, with evidence-backed approvals, a full audit trail and a kill switch.
Together, we’ll map your support, sales, IB, payments and compliance workflows into a permission matrix you can roll out step by step and defend in an audit.