Raymond James launches "Rai" - an internal AI operations agent with human oversight
Raymond James rolled out a proprietary digital operations agent called Rai to give advisors and staff faster answers to day-to-day questions, while keeping people in control. Rai pulls guidance from internal systems and policy documents, adapts to each user's role and entitlements, and will expand firmwide after an initial rollout to select business units.
Leadership is clear on intent. "Rai reflects our strategy of applying artificial intelligence to enhanced service models and secure, scalable applications that empower professionals and financial advisors across the firm," CEO Paul Shoukry said in a statement Tuesday. Human-in-the-loop remains the standard.
What Rai means for operations
Think of Rai as a single chat interface that reduces swivel-chair work. It answers policy, process, and "how do I" questions using curated internal sources, then routes staff to the right next step. Faster resolution. Fewer tickets. Cleaner handoffs.
The agent learns within guardrails: role-based access, entitlement-aware responses, and human review over recommendations and actions.
Why this matters right now
Agentic AI is moving from proof-of-concept to production inside broker-dealers. Regulators are watching. FINRA has highlighted both the momentum and the risks: agents acting beyond a user's intended authority, opaque decision paths, and potential exposure of sensitive data. Those are operations problems first, technology problems second.
Useful context: FINRA's Artificial Intelligence key topic page summarizes the supervisory lens on AI use in financial services. See FINRA: Artificial Intelligence. For control frameworks, the NIST AI Risk Management Framework is a practical reference.
Inside the Raymond James AI stack
- Rai: internal chat agent with human-in-the-loop oversight
- Adoption signals: 10,000+ associates using conversational AI
- Developer copilot output: ~3.2 million lines of code per month (under oversight)
- Ecosystem: CRM note assistant, Zoom-to-CRM summaries, internal generative search, secure access to ChatGPT Enterprise and Microsoft Copilot
- Investment: ~$1.1B annually in technology
"As an intelligent agent, Rai's impact will only grow to deliver a bespoke experience for each individual user," Chief AI Officer Stuart Feld said, noting the current rollout is the first of several phases.
What operations leaders should do in the next 90 days
- Define high-volume use cases: policy Q&A, doc lookups, CRM note hygiene, new-account workflows, exception triage.
- Map entitlements: tie Rai responses to roles, data domains, and approval levels; block out-of-scope sources.
- Instrument everything: log prompts, sources, recommendations, user actions, and final outcomes for audit.
- Set guardrails: pre-approved source list, PII masking, rate limits, dual approval for higher-risk actions.
- Create a review loop: weekly error review, false-positive/false-negative tracking, rapid model guidance updates.
- Pilot with constraints: small group, tight scope, clear success metrics, and an opt-out path.
- Train for usage and refusal: how to ask, how to verify, when to escalate, when to say "no" to the agent.
Operational controls that reduce risk
- Authority boundaries: enforce least-privilege by default; require explicit approvals for trades, money movement, or client communications.
- Traceability: show sources alongside answers; store decision trails that supervisors can replay.
- Content governance: curate policies and SOPs; mark stale docs; auto-notify owners to update or retire content.
- Data protection: redact PII in prompts, block outbound calls to unapproved systems, and monitor for data exfiltration patterns.
- Adversarial testing: red-team prompts, jailbreak attempts, and prompt-injection scenarios before every release.
- Fail-safes: clear fallbacks when confidence is low-route to a human queue, not a guess.
- Incident playbook: sever access quickly, roll back model configs, notify compliance, and perform root-cause analysis.
KPIs to track from day one
- Time-to-answer and queue deflection rate for internal queries
- First-contact resolution and escalation rate
- Error rate by category (policy misreads, stale source, entitlement block)
- Compliance exceptions linked to agent use
- Coverage of approved sources and percent of stale documents
- User adoption by role and weekly active users
- Supervisor review cycle time and rework rate
Rollout blueprint you can copy
- Phase 1 (4-6 weeks): Choose two use cases, set guardrails, build logs/telemetry, train 50 users, measure baseline.
- Phase 2 (6-8 weeks): Expand to 200-300 users, add approvals for medium-risk actions, publish transparency dashboards.
- Phase 3 (ongoing): Broaden sources, integrate with CRM and ticketing, quarterly red-team testing, and policy refresh cycles.
Regulatory pulse check
FINRA's Greg Ruppert called out three core risks: agents stepping beyond user intent, opaque multi-step decision paths, and accidental exposure of sensitive data. Translate that into controls: entitlement checks, full traceability, and strict data boundaries. If you can explain how an answer was produced, who approved it, and where the data came from, you're on the right track.
Bottom line for ops
Rai aims to reduce friction across internal workflows without removing human judgment. The opportunity is speed and consistency; the work is governance and measurement. Set the rules, publish the metrics, and keep people accountable.
Further resources
- FINRA: Artificial Intelligence
- NIST AI Risk Management Framework
- Curated AI tools for finance - implementation ideas
Your membership also unlocks:
