00 The problem
Routing plans were becoming too expensive to maintain by hand.
Contact-center routing sits at the intersection of customer intent, staffing reality, quality standards, and operational policy. Manual plan authoring can work for a small surface area, but it breaks down as queues, intents, and staffing conditions keep changing.
Before
Human-authored plans
slow to create and revise
hard to keep aligned with quality signals
fallback behavior depended on operator judgment
After
AI-assisted planning
recommendations built from evidence
reviewable before production use
live routing stays deterministic
01 Constraints
The plan is configuration, product behavior, and operational policy.
Evidence
Ground it in operating signals
Structured fields were only part of the picture. Quality-control signals and routing history helped keep plans current.
Supply
Account for availability
The best specialist is not always free. A good plan needs an ordered fallback path.
Control
Make trust incremental
Operators needed to inspect and promote recommendations before trusting more automation.
Core tension A plan that is too narrow creates wait time. A plan that is too broad loses the benefit of specialization.
02 System shape
Separate learning from serving.
The architecture kept expensive intelligence away from the live customer path. Analysis and planning happened before routing. Runtime only consumed already-approved configuration.
- Offline analysis. Prior interaction patterns become bounded evidence about routing demand and quality-control outcomes.
- Candidate planning. Evidence feeds a recommendation workflow that proposes specialist-first routing plus fallbacks.
- Deterministic runtime. The live path reads promoted configuration instead of waiting on model calls.
Read historyAnalyze prior routing patterns, demand, and quality-control outcomes.
Build evidenceConvert operating history into constrained, reviewable signals.
Generate planRecommend routing behavior and fallback steps.
PromoteMove only reviewed output into production configuration.
03 Key decisions
The system design mattered more than the prompt.
Decision 1
Batch analysis over runtime inference
Model calls are easier to retry, inspect, sample, and control when they are not blocking a live decision.
Decision 2
Constrained labels over free-form output
The model mapped evidence into bounded routing concepts instead of inventing arbitrary labels.
Decision 3
Fallback plans over best-match only
The useful artifact was the ordered relaxation path when the ideal expert was unavailable.
Decision 4
Human in the loop first, then remove it
The first version made generated plans staged artifacts. Later, the same boundary became the path to more automated promotion.
04 Operating model
The human gate was a temporary trust boundary.
The first safe version kept people in the loop. That was the right initial product choice: generated plans were staged, reviewed, and promoted explicitly. Once the evidence, checks, and operational confidence improved, the same boundary could be reversed into an automation path.
Phase 1
Human in the loop
Operators inspected recommendations before they affected routing behavior.
Bridge
Earn confidence
Evidence, diffs, checks, and rollback made generated plans easier to trust.
Phase 2
Automate the gate
The review boundary could shift from manual promotion toward automated checks and promotion.
05 Lessons
Applied AI succeeds or fails around the model.
- Review is a product feature. AI systems earn trust faster when the first version makes review cheap instead of demanding full autonomy.
- Operators care about behavior. Predictability, clear failure modes, and rollback matter more than model sophistication.
- The workflow is the product. Evidence quality, promotion gates, observability, and ownership boundaries decide whether the system can run.
Through-line The important move was separating offline intelligence from online execution, then wrapping generated output in the same controls as any other production configuration.