Agents that earn the automation.

Agents propose changes to your routing and operations, prove them against your own history, and wait for a person. On your infrastructure, over your rails.

Rails, gateways, correspondents, the core, the ledgerConnector FabricWhat each payment actually cost, and how long it actually tookSettlement TruthDeterministic feasibility, learned scoring, deterministic selectionPolicy EngineBacktest, shadow, canary, and honest impact measurementProving GroundAgents that propose, with evidence, and never release moneyAgent HubRoutingComplianceReconciliationOperationsOnboarding
31 bpsBest price list, worst settlement
3Gates before live money
100%Decisions on a hash-chained log

Propose. Prove. Promote.

In that order, every time.

Propose

Agents read your settlement history and open a change with the evidence attached.

Prove

Replay it against your own payments. Each change alone, then in combination.

Promote

Shadow, canary, live. Live needs a second approver.

Built by the people who ran the rails.

Production systems our team built and ran. None is a Switchboard deployment, and we will not pretend otherwise.

$130M+
cross-border settlement
100,000+
transactions settled
$3.5B+
OTC and FX volume
3 weeks
concept to production
~5 sec
settlement, was 2 to 3 days
6
systems live in production

Figures as of August 2026. Client names available on request, under NDA.

Fourteen agents, and a hard line on what they may do.

Each one shows what it can read, and what it can never do without a person.

Objectives in plain English

You say what to optimise in a sentence. The agent repeats it on every proposal.

Evidence attached

Every finding carries its observation count. Too few, and it says so.

Proved before promoted

Backtest, then shadow, then canary. Then a person.

Automation earned

Promotion needs measured agreement with your team, not confidence.

Routing and liquidity4 agents
Route OptimiserRoutes on realised settlement, not published terms.
Rail Health MonitorCatches a rail degrading before the SLA sheet does.
Liquidity PlannerPrefunding from realised timing, not nominal cutoffs.
Retry StrategistFailover built from your actual failure taxonomy.
Operations4 agents
Break ResolverWorks breaks and proposes a disposition with evidence.
Automation FrontierFinds which decisions are now safe to automate.
Exception TriageRoutes each exception to the owner who can close it.
Alert TunerFinds the rules your team closes without acting.
Compliance3 agents
Screening TunerThreshold changes with the tradeoff priced.
Case NarrativeDrafts the write-up. Your analyst keeps the pen.
Sanctions ReviewerDisposition recommendation, with the gaps named.
Onboarding and analysis3 agents
Counterparty OnboarderReads documents, scores against your policy.
Limit RecommenderLimits from observed behaviour, exposure stated.
Analytics AssistantAnswers questions about your data. No SQL.

One layer, every rail.

The bottom two need your ledger. That is why this runs inside your walls.

Deterministic where it must be

Rules are rules

Limits, cutoffs, jurisdiction and screening are rules with named reasons. A model never decides whether a payment is allowed, only which allowed rail to use.

Explainable by construction

Every decision, on the record

Every rail considered, the rule that excluded each one, and one sentence a person can read. Hash-chained, so an altered record is caught at the exact entry.

Where your data already lives

Your account, your region

Bedrock in your account, a local model on your hardware, or no model at all. One config line, same product.

Four processes, one layer.

Routing is the money. The other three are the hours. One engine, one audit trail.

Route on what settlement cost.

A price list gives you the fee. Not the slippage, the latency tail, the repair rate, or the hours your team spends on the ones that go wrong.

01Scored on realised cost, time, breaks and manual load.
02Confidence beside every estimate, and a mark when it is still a price list.
03Limits, cutoffs, jurisdiction and screening stay rules, never probabilities.
04A guard that refuses a route materially worse than the cheapest feasible one.
05If scoring cannot run, the payment takes the default and is flagged, not held.

Three ways to deploy it.

Where the data goes, in writing, before you ask.

Inside your VPC

Runs in your cloud account. Inference goes to Bedrock in your region, over a VPC endpoint.

No payload leaves your account. Nothing retained for training.

On your own hardware

For institutions that will not send a payment instruction to any cloud.

Nothing leaves the building. The provider is one config line.

On top of GSOS

The agentic layer over the same ledger and connectors. No integration work.

One deployment, one audit trail, connectors already in place.

Why this is different.

Routing systems decide on published terms: the fee schedule, the quoted time, the stated success rate. Those are a partner's marketing, and they are optimistic in exactly the places that cost you money.

The difference between quoted and realised is visible in one place: your ledger. A product outside your walls cannot see it. One that can has to run inside them.

In the reference book, the rail with the best price list is the worst rail to use, by 31 basis points. An operator choosing on price picks it. So does the engine, for the first few payments, and then it stops.

See it on your corridors.

Bring ninety days of history for two corridors. We will show you the gap in your own numbers.