Interior Optimum (containment rate)
The Interior Optimum is the single most counterintuitive operating result in the Value-Based Planning Model: maximum containment is not the cost optimum. Plotting total cost (AI cost + human cost + cascade cost + rebound cost) as a function of AI containment rate produces a U-shape, and the bottom of the U typically sits between 40% and 60% containment — well below the 80-90% targets vendor business cases routinely set.
Interior Optimum is a Level 4 — Advanced (The Ecosystem Emerges) concept on the WFM Labs Maturity Model™ because finding it requires modeling all four cost components simultaneously, which depends on the escalation tax, the demand rebound model, and the three-pool partition all being instrumented.
The result is documented in Lango (2026), Value-Based Models for Customer Operations.[1]
Containment as a planning variable, not a target
In Level 2-3 operations, containment rate is treated as a target — usually pushed toward 80-90%, occasionally higher. The implicit assumption is that more containment is always cheaper.
The assumption holds for marginal AI cost in isolation. It fails for total cost. Three other components rise as containment rises:
- Cascade cost (Escalation Tax) grows because higher containment thresholds admit lower-confidence interactions to Pool AA, which raises p_esc.
- Human cost on the remainder grows per-unit because the Complexity Premium raises AHT.
- Rebound cost grows because higher containment generates more induced demand (R_d, R_i, R_s) that lands back in the system.
The sum is U-shaped. Cost falls as containment rises from low values, reaches a minimum at an interior point, then rises again toward 100%.
The four cost components
Total cost as a function of containment rate c:
Total(c) = AI(c) + Human(c) + Cascade(c) + Rebound(c)
| Component | Behavior | Driver |
|---|---|---|
| AI(c) | Falls monotonically with c | More automation = more c_AI volume, but per-unit AI cost is small |
| Human(c) | Falls per-volume but rises per-unit | Volume drops, but Complexity Premium raises AHT on the remainder |
| Cascade(c) | Rises with c, often super-linearly above ~60% | Higher c admits lower-confidence interactions, raising p_esc |
| Rebound(c) | Rises with c, with delay | Induced demand (R_d, R_i, R_s) materializes over months |
The first two are decreasing-then-flattening. The last two are accelerating. The crossover is the Interior Optimum.
Why the optimum is interior
Two things must be true at the same time:
- At low containment, AI(c) and Human(c) per-volume both fall fast as c rises. Cascade(c) and Rebound(c) are still small. Total cost falls.
- At high containment, AI(c) is nearly zero in marginal terms. Human(c) per-unit rises as the Complexity Premium accelerates. Cascade(c) rises super-linearly because the routing has admitted lower-confidence cases. Rebound(c) is fully matured. Total cost rises.
Between the two regimes, total cost has a minimum. Calculus does the rest.
Numerical example
The white paper's reference operation, swept across containment levels:
| Containment rate | AI cost | Human cost | Cascade cost | Rebound cost | Total |
|---|---|---|---|---|---|
| 0% (no AI) | $0 | $43,000,000 | $0 | $0 | $43.0M |
| 25% | $500,000 | $33,000,000 | $1,500,000 | $1,500,000 | $36.5M |
| 50% | $1,000,000 | $24,500,000 | $4,000,000 | $4,000,000 | $33.5M ← optimum |
| 70% | $1,400,000 | $18,500,000 | $9,000,000 | $7,500,000 | $36.4M |
| 90% | $1,800,000 | $11,000,000 | $19,000,000 | $11,000,000 | $42.8M |
The vendor business case at 90% containment shows AI cost ($1.8M) and "saved" human cost ($32M of the $43M baseline) — an apparent savings of $30M+. The honest total cost at 90% is $42.8M — almost the entire baseline gone, plus $10M in cascade and rebound. Net savings: $0.2M.
The honest total cost at 50% is $33.5M. Net savings: $9.5M.
The interior optimum delivers 50× the realized savings of the maximum-containment strategy, precisely because it stops climbing the curve before cascade and rebound dominate.
How to find your operation's optimum
- Build the cost-curve sweep. Compute Total(c) at c ∈ {0, 0.1, 0.2, ..., 0.9, 0.95}. Use distributional inputs where possible — the curve is not deterministic.
- Use type-level cascade probabilities, not aggregate. Aggregate p_esc smooths over the heavy tail; type-level p_esc captures it.
- Use the Service Demand Rebound Model for Rebound(c). Apply R_d (15-35%), R_i (10-20%), R_s (5-15%) as planning inputs; tighten as longitudinal data accumulates.
- Use post-deflection AHT for Human(c). Apply the 5-8% per 10pp Complexity Premium.
- Plot the curve. Find the bottom. Set containment policy there. Re-run quarterly.
The bottom of the U is the operating point. It is not maximum savings — it is the savings that are actually achievable without the rising terms eating them.
Strategic implications
The Interior Optimum reframes containment from a target to a portfolio decision:
- More containment is not always cheaper. Beyond the optimum it is more expensive. The marginal AI cost saved is exceeded by the cascade and rebound cost added.
- The optimum shifts. AI capability improvements push the optimum higher. Rebound elasticity changes (e.g., a new digital channel) push it lower. Periodic re-sweeping is required.
- The optimum is operation-specific. Two operations with the same baseline volume and AHT can have different optima depending on Value Score distribution, customer-segment churn sensitivity, and cascade probabilities.
- Some operations have multiple local optima. If the cost function is multi-modal — typically when one segment behaves very differently from the rest — segmenting and optimizing per-segment is more honest than averaging.
The general rule: do not target maximum containment. Target the operating point where total cost is minimized.
Live visualization
The U-curve is rendered live by the Value Routing Model interactive tool at valuerouting.wfmlabs.com. The "AI Containment Optimization" view sweeps containment from 0% to 100% and plots net benefit (cost saved minus value destroyed) at each level. The orange dashed line marks the optimum. Adjust weights or per-type value scores in the tool to see how the peak shifts; the drop-off after the peak is the empirical demonstration that 100% containment is never the answer when some interaction types create real human value.
Limitations
- The curve is approximate. All four component functions have parameter uncertainty; the U is a band, not a line. Decisions should account for the band width.
- The optimum is not stable. It moves with AI capability, customer behavior, regulation, and product change. Treat it as a recurring planning question, not a one-time setting.
- Multi-objective surfaces complicate it. If CX or EX constraints bind, the cost-only optimum is not the operating optimum. The full Pareto frontier is the right surface, with cost-only U as a 1-D slice.
Maturity Model Position
- Level 1 — Initial (Emerging Operations) — Containment is not a planning concept.
- Level 2 — Foundational (Traditional WFM Excellence) — Containment is sometimes a vendor-set target, sometimes ignored. The U-curve is invisible.
- Level 3 — Progressive (Breaking the Monolith) — Containment is targeted, typically high (80%+). Cascade and rebound are observed but not modeled. The U is suspected but not measured.
- Level 4 — Advanced (The Ecosystem Emerges) — The Interior Optimum is the operating principle. Containment is set at the U-curve minimum, swept quarterly.
- Level 5 — Pioneering (Enterprise-Wide Intelligence) — Closed-loop governance recalibrates the optimum continuously. Drift in any of the four cost components triggers automatic re-sweeping and threshold adjustment.
The Level 3 → Level 4 transition is the moment containment stops being a target and starts being an output of a cost optimization.
See Also
- Value-Based Planning Model — the framework the U-curve sits inside
- The Escalation Tax — the cascade cost component of the curve
- Service Demand Rebound Model — the rebound cost component of the curve
- Three-Pool Architecture — the architecture that defines what containment means operationally
- Cognitive Portfolio Model (N*) — Pool Collab's staffing math contributes to Human(c) on the rising side
- Multi-Objective Optimization in Contact Center — the Pareto frontier the U-curve is a slice of
- Value Routing Model — type-level Value Scores feed the per-type cost components
References
- ↑ Lango, T. (2026). Value-Based Models for Customer Operations — From Traditional Queuing to Bottom-Up Value Planning. WFM Labs white paper.
