The challenge
A customer wanted something simple: set up email on a newly purchased domain, forward two older domains, and understand why one had been cancelled. The intent was clear from the first message. Resolution still took 24 minutes.
The front-line bot detected every signal correctly. Guides followed procedure. The customer got the outcome they asked for. And yet 91% of the interaction time went to wait, re-explanation, and compensating for missing system awareness.
Constraints
- Data compliance. Fraud signals cannot leak to front-line bots or customers, boolean-only surfacing required.
- Legacy system sprawl. CRM, RegManager, and fraud tooling don't share a canonical state; integration layer had to reconcile, not rebuild.
- Cross-team ownership. L2 (second-line support), ATS, fraud, and billing each owned partial context. No single team could fix this unilaterally.
- Guide cognitive load. Any proposed system had to reduce, not add to, what guides juggle during live interactions.
My approach
I mapped the interaction minute-by-minute against the data that existed but wasn't flowing. Then I designed a four-layer orchestration architecture where AI does the cross-system work the humans were doing by hand:
- System Orchestration Layer, real-time reads across CRM, Domain Registry, and fraud systems.
- Reasoning Layer, interprets conflicting status codes and resolves them into a single authoritative narrative.
- Conversation Layer, adapts language to the system's certainty (compliant, calibrated, never overclaims).
- Escalation Intelligence, decides if, when, and to whom to hand off, with a structured payload.
Four specific manual behaviors were identified and converted into orchestration primitives: cross-system correlation, fraud awareness without data leakage, structured bot-to-human handoff, and elimination of redundant re-explanation through shared conversation state.
Four narrow layers, not one larger agent. The tempting build was a single smarter model that read every system, reasoned, talked to the customer, and decided when to escalate, all in one call. I split it instead. AI does the mechanical cross-system work that humans were doing by hand, and the escalation logic stays structured and auditable. The trade-off I accepted was more moving parts, in exchange for layers I can test and govern one at a time. A regulated support flow cannot do without that.
Artifacts
The full case study includes the detailed orchestration diagram, the handoff payload schema, the fraud-signal boolean pattern with example copy, and a minute-by-minute before/after timeline of the same interaction resolved by the proposed system.
Want the full version? Emailchristi [at] christi.iowith your role and company and I'll send it over, usually within two business days.
Projected impact
Wait, re-explanation, manual system bridging Actual resolution work
This isn't an edge case. The same orchestration gap shows up across thousands of similar interactions. The architecture generalizes.
About these projections
The figures on this page are drawn from internal program reporting I authored or co-authored as the practitioner on the engagement. They are reproduced here in rounded form. They were not produced by an independent third party, and proprietary detail has been omitted where required by the engagement.
Lift figures (CSAT, accuracy, handle time, hallucination rate) reflect pre/post comparisons against a matched baseline using the cohort, time window, and measurement instrument noted in the case study. Volume and adoption figures come from production analytics dashboards. Cost figures reflect either avoided spend or unlocked budget in the named fiscal period.
- 24-minute baseline and 91% wait/re-explanation share: derived from a minute-by-minute audit of a single representative interaction, cross-referenced against guide tooling logs and CRM timestamps.
- <2 min resolution and 90% handling-time reduction: projected impact based on the proposed orchestration architecture, not a measured post-implementation result. The full case study contains the before/after timeline used to derive the projection.
- All identifying details about the customer, the guide, and the underlying systems have been anonymized in this public teaser.
What I'd do differently
Start the cross-system correlation layer before the reasoning layer. In the original scope I had them co-equal, but reasoning quality degrades fast when upstream data is stale. Lock the orchestration first. Everything else compounds off of it.
Collaborators
Analysis reviewed with L2 leadership, fraud operations, and the registrar engineering team responsible for CRM/RegManager integration. All identifying details are anonymized in this public teaser.
Skills demonstrated
- Multi-agent orchestration
- Handoff payload design
- Fraud-signal compliance patterns
- Conversation state architecture
- Cross-system reconciliation
- ROI modeling
- Stakeholder alignment (L2, Fraud, Eng)