Interior Optimum (containment rate)

From WFM Labs
(Redirected from Interior Optimum)
The intuition: people cost falls as automation takes contacts, automation cost plus the cost of fixing its mistakes rises, and the interior optimum is the share where the two balance. Illustrative shapes; the full curve below adds escalation and rebound.

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 for the reference mix — 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)

Cost components and their behavior in containment
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:

  1. 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.
  2. 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 reference case drawn: total cost against containment is a U whose bottom sits near half, while the vendor business case points at ninety percent. The shape is the finding; the numbers are illustrative.

The white paper's reference operation, swept across containment levels:

Total cost by containment rate (illustrative; based on white paper's reference scenario)
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

  1. 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.
  2. Use type-level cascade probabilities, not aggregate. Aggregate p_esc smooths over the heavy tail; type-level p_esc captures it.
  3. 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.
  4. Use post-deflection AHT for Human(c). Apply the 5-8% per 10pp Complexity Premium.
  5. 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.


Where the optimum sits: the interaction's share of value

The optimum moves left as the interaction carries more of the value the customer is buying. Four illustrative archetypes: a commodity service, a service with upside in the interaction, a service where the interaction is part of the experience, and a service where the experience is the product. Positions are illustrative, not measured.

The four-term curve above prices what automation costs and saves and what its failures cost. It does not price the value an interaction creates for the business when a person handles it. The "Strategic implications" section notes that the optimum is operation-specific and moves with the Value Score distribution and churn sensitivity of the customer base, and the live visualization below already nets value destroyed against cost saved. This section supplies the variable behind both: the share of the customer's value that is created in the interaction itself.

Call that share s, and call the value forgone at containment rate c for an operation with share s the value-lost term, V(c; s). The term extends the four published terms of the model; it is this page's addition, not the white paper's. The value-adjusted total is

Total(c; s) = AI(c) + Human(c) + Cascade(c) + Rebound(c) + V(c; s)

and the value-adjusted cost optimum is the bottom of that curve. It is a single-objective scalarization of the multi-objective surface the Limitations section describes, not a replacement for it: where customer- or employee-experience constraints bind, the Pareto frontier remains the operating surface, and V is the cost-axis projection of one of them.

V has two properties. Over most of the range it rises with c, because each further increment of automation reaches into interactions that were creating more value, and it rises faster the larger s is. Near c = 0 it can be negative for high-s services: removing friction the customer never valued — the order for collection, the check-in, the receipt — adds value rather than destroying it, and moves the optimum a little to the right of zero. The service-productivity literature treats the human–automation balance as a decision variable to be set for profit rather than a quantity to be maximized, which is the stance this term encodes.[2]

Four archetypes

The share of value runs from near zero to near one, and four points on it are recognizable. The optima are illustrative positions to make the shape visible, not measured results; an operation finds its own by adding V to its sweep. Two things move the optimum along the row. For low-s work it is the cost terms themselves: cascade and rebound are small where automation rarely fails and rarely induces demand, so the four-term optimum sits further right than the reference case's 40–60%. For high-s work it is V, which pulls the optimum left of wherever the cost terms put it.

Archetype How the interaction creates value Illustrative optimum Example
Commodity service It does not. The customer does not differentiate the purchase by the service received to make it; the interaction is a cost of delivery Wherever the four cost terms put it, which for low-cascade, low-rebound work can be 90% or more; V adds nothing Balance enquiry, password reset, delivery status
Upside in the interaction The interaction is where fees are earned, upgrades sold, retention won and lifetime value extended. The product is bought elsewhere; the margin is made in the conversation Around 70%: automate the routine, keep the moments that carry the upside with people Travel servicing with upgrade and ancillary revenue; insurance renewal; banking with cross-sale
Interaction as part of the experience The customer regards the service received as part of what is being paid for. Automating it changes the product, not only its cost Around 30%: automation for the parts the customer is indifferent to, people for the rest Premium and managed accounts; concierge and assisted services; care coordination
Experience is the product The interaction is the purchase. Remove the person and the product is gone Near 0, with small increments of automation that add value at the edges Fine dining; a guided tour; the doorman at a luxury hotel

The last archetype is where the intuition is sharpest, and it has a name. The doorman fallacy is the error of defining a role by its most visible mechanical function, replacing it with a machine that performs that function, and discovering afterwards what else the role did. A hotel doorman opens doors; the same person also hails taxis, provides security, discourages vagrants, recognizes returning customers and signals the status of the hotel. The automatic door does the first and none of the others. The saving is the doorman's salary; the loss is everything the doorman was for.[3] The experience-economy argument makes the general case as a progression: goods are props and services are the stage, and where a business has progressed to staging experiences, the experience is the offering.[4] For such a business V dominates the cost terms almost from the first increment of automation.

Estimating the term

s is measured per interaction type, not per operation. The Value Routing Model scores each type on lifetime-value impact, customer effort, revenue opportunity and churn risk, separately for human and automated handling, and routes on the differential. Normalized to the score range, that differential is s at type level: a high-value interaction that automation handles as well as a person has high value but low s. A first estimate of V is then a sum over the types automated at containment c of each type's s applied as a fraction to the value at stake in that type's volume — the value enters once, through the stake, and s says what fraction of it the handler change forgoes. The estimate is coarse and should be carried as a band, like the four cost terms.

What the share of value changes in practice

  • Find s before sweeping c. The sweep in "How to find your operation's optimum" assumes V is zero. Run without s, it returns an optimum that is too far right by an amount proportional to s.
  • Segment by s, not by channel. Most operations are not one archetype. A travel servicing operation carries commodity work, upside work and experience work in the same queue, and each has a different optimum. Averaging them produces an optimum wrong for all three; the multiple-local-optima caution above applies with force.
  • s is a design choice as well as a measurement. Two businesses with the same cost curve can choose different s: one decides its interactions will be where value is created and invests in them; another commoditizes them deliberately. The optimum then follows the strategy.
  • Automation at the edges of a high-s service can add value. That is the negative region of V near zero: the door can be automated without automating the doorman.
  • Level 4 is where s becomes measurable. Below Level 4 the planning system holds volume and cost and can compute only the four-term curve. V needs a measured value attribute per interaction type and per node, which Level 4 supplies; The Value Destruction Risk in Service Automation is the account of what goes wrong without it.

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 share of value is itself an estimate. The value-lost term depends on how much of the customer's value is created in the interaction, which is measured per interaction type only at Level 4 and is a design choice as much as a measurement; the four archetypes are illustrative positions, not findings.
  • 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 [[Multi-Objective Optimization in Contact Center|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

References

  1. ↑ Lango, T. (2026). Value-Based Models for Customer Operations — From Traditional Queuing to Bottom-Up Value Planning. WFM Labs white paper.
  2. ↑ Rust, R. T., & Huang, M.-H. (2012). "Optimizing Service Productivity". Journal of Marketing 76 (2), 47–66. doi:10.1509/jm.10.0441.
  3. ↑ Sutherland, R. (2019). Alchemy: The Surprising Power of Ideas That Don't Make Sense. WH Allen.
  4. ↑ Pine, B. J., & Gilmore, J. H. (1998). "Welcome to the Experience Economy". Harvard Business Review 76 (4), 97–105.
Retrieved from "https://wiki.wfmlabs.org/w/index.php?title=Interior_Optimum_(containment_rate)&oldid=4296"