Skip to content

AI agents

AI agents that prepare the work, and a person who approves it

Narrow agents for research, operations and client support, each with a fixed set of tools, a spending limit and an approval step wherever money or clients are involved. Built so your team can see what every agent did, and stop any of them in one click.

Book a 30-minute callHow it works
ops.yourbroker.com/agentsLiveAgent approvals3 waitingKill switchReconciliation agent09:143 breaks found, 3 fixes proposedNeeds approvalOnboarding check agent09:02File 4471 complete, 1 document expiringNeeds approvalSupport draft agent08:57Reply drafted for ticket 8812Needs approvalDaily report agent08:30Management pack preparedApprovedResearch agent08:106 filings summarised, 14 sourcesApprovedReconciliation agentrun 2041Checked 1,284 trades against drop copy and back office. Prepared 3 fixes for approval.TRADEBREAKPROPOSED FIXEVIDENCET-88412Quantity 50,000 vs 5,000Amend booking to 50,000Drop copy, confirmT-88430Missing in back officeCreate bookingDrop copyT-88457Price 1.08791 vs 1.08719Amend priceExecution reportTOOL CALLSmcp.trades.read(date=2026-10-09)1,284 rowsmcp.dropcopy.read(date=2026-10-09)1,284 rowsmcp.backoffice.read(date=2026-10-09)1,283 rowsmcp.backoffice.propose_fix() x3proposal onlyLIMITSScoperead, propose onlySpend today£0.42 of £5.00Modelv3.2, pinnedApprove 3 fixesRejectEditApprover: head of operations

Illustrative screen built on synthetic data

Who it is for

Built for ai agents for trading teams

Operations teams

Reconciliations, breaks, onboarding checks and daily reports that take hours of careful, repetitive work.

Client support

First replies drafted from your own documents and the client's account, reviewed before they are sent.

Research teams

Summaries of filings, news and internal notes, with every statement linked to its source.

What you get

Features for both sides of the screen

01

What agents do

  • PrepareGather data, check it, and draft the reconciliation, the reply or the summary.
  • ExplainShow the evidence behind every proposed action.
  • EscalateRoute anything unusual to a person rather than guessing.
  • RecordLog every tool call, input and output.

02

What agents never do alone

  • Move moneyPayments, refunds and payouts always need a person.
  • Place or change ordersAgents have no trading permissions.
  • Give adviceClient-facing text passes the same wording check as our chart analysis.
  • Change systemsConfiguration changes are proposed, not applied.

How an agent is built

Tools, limits, approvals, logs

Every agent we build follows the same four-part design, which your compliance team can review once and recognise everywhere.
  1. 1

    Fixed tools

    MCP servers expose exactly the data and actions an agent needs, read-only by default.

  2. 2

    Limits

    Spending, rate and scope limits per agent, enforced outside the model.

  3. 3

    Approval queue

    Proposed actions wait for a named person, with the evidence beside them.

  4. 4

    Audit and kill switch

    Everything is logged, and any agent can be stopped immediately.

Choosing the first agent

What makes a good first workflow

The first agent should be boring, frequent and easy to check. These are the questions we ask in the Platform Review.
Is it repetitive?
Work done the same way every day, by someone who would rather be doing something else.
Is it checkable?
A person can tell whether the result is right in a minute or two.
Is the data reachable?
The systems involved have APIs, exports or a database the agent can read safely.
Is a mistake contained?
A wrong draft is caught at approval, not discovered by a client.

Why AlchmAI

Why trading firms choose a specialist

Approval by design

Agents prepare, people decide.

MCP built in

Governed access to your data.

Evaluated

Test suites run before every release.

Model-agnostic

Change models without rebuilding.

Logged

Every tool call on record.

Bank experience

Agents with guardrails and evals in a markets division.

Case study, uk investment bank, markets division

Six front-office trade capture and pricing applications on one configuration-driven ticket library

A prop firm dashboard, a broker's order ticket and an RFQ blotter are the same problem: fast, validated input that stays in sync with everything else. We have built it where the stakes were highest.

Read the case

6

applications on one shared ticket library

0

code changes to onboard a new product or field

15+

engineers working to the standards set at inception

How we engage

It starts with a two-week Platform Review

Every engagement starts with a two-week Platform Review. We take read-only access to your charts, data feeds and any AI feature, test them against what your clients actually see, and hand you a findings report, a costed plan and, where it helps, a working prototype. Fixed price, agreed before we start, and credited in full against a build that begins within 60 days. The report is yours whichever way you go.

  1. 1. Weeks 1 to 2

    Platform Review

    Read-only access. Findings report, costed plan and a prototype where it helps. You sign off the baseline we will measure against.

  2. 2. Weeks 3 to 8

    Build, in fortnightly releases

    Working software every two weeks on an environment you can use. Any AI wording is agreed with your compliance lead before a client sees it.

  3. 3. Go-live

    Evidenced release

    A runbook, an evidence file your compliance lead can read in one sitting, and an off switch for every AI feature.

  4. 4. Day 90

    Results check

    What changed, measured against the baseline you signed at the review.

Questions

AI agents for trading teams: common questions

Which models do you use?

Whichever suits the task and your data policy. We have built with Claude and OpenAI models, and design agents so the model can change without rebuilding the workflow.

Will our data be used to train models?

We configure providers so your data is not used for training where they offer that option, and agents can run in your own cloud account. Your data policy decides.

How do you know an agent is working?

Against a baseline taken before it starts: time per case, error rate and the share of cases a person had to correct. You sign off the baseline in the Platform Review.

Can agents message clients directly?

We recommend that client-facing messages are reviewed by a person, especially at first. Where you choose otherwise, the wording check and logging still apply.

Related

Often built alongside

Brokers and CFD providers

Web terminals, mobile apps and back-office screens in your brand, on the back end you already run.

Learn more

Prop trading firms

Evaluation dashboards, tick-level rule engines, breach evidence and payout workflows.

Learn more

Trading education platforms

Chart lessons, practice on historical data and AI explanations that teach without tipping.

Learn more

Bring us the task nobody wants to do.

A 30-minute call is enough to tell whether it would make a good first agent.

Book a 30-minute callSend a message