AI controls
Controls for AI features, so you can say what it told your clients, and why
Wording checks, evaluation suites, approval steps, logging and kill switches for AI that sits close to clients and markets. Added to features you already run, or built in from the first sprint.
Illustrative screen built on synthetic data
Who it is for
Built for controls for ai features
Brokers with AI in the app
Chart summaries, assistants or alerts already live, with no clear record of what they have said.
Education platforms
AI explanations for students that must stay educational and never become tips.
Product and compliance teams
Teams that need to sign off an AI feature and want evidence, not assurances.
What you get
Features for both sides of the screen
01
Controls before display
- Wording checkBlocks advice-like language, price targets, forecasts and guarantees, with a safe fallback.
- Approved vocabularyThe phrases a feature may use, agreed with your compliance lead.
- Fact groundingThe model may only state facts it was given, and every number is checked against its source.
- Human reviewHigher-risk outputs routed to a person before release.
02
Controls after release
- LoggingInputs, outputs, prompt and model versions for every response, with retention you set.
- Evaluation suitesRegression tests run on every change to a prompt, model or rule.
- MonitoringRates of blocked outputs and fallbacks tracked, with alerts when they move.
- Kill switchAny AI feature turned off in one step, without a release.
Signed-off analysis
The method behind every control
- 1
Code detects
Deterministic code finds the facts: levels, ranges, account data, document contents.
- 2
A model explains
A language model writes up only those facts, in approved wording.
- 3
A wording check runs
A separate check blocks anything that reads as advice or prediction.
- 4
A person reviews
Every output is logged, and the riskier ones are reviewed before release.
What can go wrong
The four failure modes we design against
- Advice-like wording
- A summary that says consider buying. Controlled by the wording check and approved vocabulary.
- Invented numbers
- A model that rounds, guesses or misremembers a price. Controlled by fact grounding and number checks.
- Drift
- A provider updates a model and the tone changes overnight. Controlled by pinned versions and evaluation suites.
- Data exposure
- Client data in prompts or logs where it should not be. Controlled by redaction, retention rules and access controls.
Why AlchmAI
Why trading firms choose a specialist
Shown, not claimed
Try the wording check on our home page.
One method
The same controls across every feature.
Evidence for sign-off
Logs and test results your compliance lead can read.
Works on existing features
No rebuild required.
Guardrails and evals
Built for a bank's markets division.
Clear about our role
We build controls; your advisers set policy.
Selected work, 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.
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
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. 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. 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. Day 90
Results check
What changed, measured against the baseline you signed at the review.
Questions
Controls for AI features: common questions
Do these controls make an AI feature compliant?
No control does that on its own. Compliance depends on your permissions, your clients and how the feature is presented. The controls give your compliance team the evidence and the levers to make that judgement.
Can you add controls to a feature someone else built?
Usually, yes. The wording check, logging and kill switch can sit between the existing feature and the client, without rewriting the feature itself.
How do you test that the wording check works?
With a growing suite of advisory and descriptive examples, run on every change, plus a sample of real outputs reviewed by a person each week.
What happens when an output is blocked?
The client sees a plain, pre-approved fallback, and the blocked output is logged for review.
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 moreProp trading firms
Evaluation dashboards, tick-level rule engines, breach evidence and payout workflows.
Learn moreTrading education platforms
Chart lessons, practice on historical data and AI explanations that teach without tipping.
Learn morePaste one of your AI feature's sentences into our wording check.
If it is blocked, we should talk. If it passes, you still need the logs.