Ninth Wave turns bank API onboarding into a governed multi-agent workflow

Ninth Wave has built Compass, an AI onboarding assistant for banks and fintech partners, using Amazon Bedrock AgentCore. The interesting bit is the architecture: specialist agents handle distinct jobs such as documentation search, field mapping and readiness analysis, while tenant-scoped grounding keeps one bank's information out of another's session.

AWS AI Watch analysis

What happened

According to AWS's technical walkthrough, Compass routes each request through an orchestrator to one of seven specialists. The system helps validate bank APIs, map fields to the Financial Data Exchange standard and score readiness for production connectivity.

The readiness score is calculated in application code from required-field coverage, rather than estimated by a language model. Ninth Wave also describes per-tenant OpenSearch indices and S3 prefixes, MFA, least-privilege access, encryption and audit logging as part of the deployment. The company reports that beta clients went live on 1 March 2026, production was ready on 15 May and general availability followed on 1 June.

Key findings

• Routing beats one giant prompt
  An intent classifier sends each request to a focused specialist, keeping task context narrower and outputs more predictable.
• Each bank gets its own context
  Compass assembles tenant-specific documentation and configuration before invoking an agent, designed to prevent cross-bank data leakage.
• The score is deterministic
  FDX readiness comes from required-field mapping coverage in application code, not a model's best guess.
• Seven jobs share one workspace
  Search, Q&A, classification, mapping, analysis, interactive workflows and readiness analysis sit behind the same portal.

Why it matters

This is a useful example of enterprise AI becoming less about asking one enormous model to do everything and more about assigning bounded jobs to separate workers. The pattern gives teams clearer places to apply permissions, observe costs and investigate failures.

It also puts a sensible line around model discretion. A language model can help interpret documents and guide work, but the final readiness score comes from a reproducible calculation. In regulated workflows, that distinction is less glamorous than an autonomous demo and considerably more useful.

Our read

Compass is a credible reference architecture for AI-assisted integration work: narrow agents, tenant-specific grounding and deterministic outputs where the number matters. The evidence is AWS's account of Ninth Wave's system, not an independent performance audit, so readers should treat the reported production milestones and security design as claims from the builder.

For teams considering a similar build, the practical lesson is to decide which steps need model judgement and which should remain ordinary application logic before choosing the agents.

What to watch

• Whether Ninth Wave publishes measured reductions in onboarding time, mapping errors or manual intervention.
• How Compass handles ambiguous field mappings and conflicting bank documentation.
• Whether the tenant-isolation design receives independent testing or certification evidence.
• How per-agent cost and latency behave as the number of specialists grows.

Discussion spark: Where should an enterprise AI workflow draw the line between model judgement and deterministic application logic?

not affiliated with or endorsed by Amazon Web Services (AWS)