AI Agents Stall at 30% Contact Containment. Here Is Why.
The first wave of agentic customer support gets to 30% containment in six months, then improvement stops. Not because the AI is not capable. Because 40% of contacts require a policy decision the agent cannot make.
The remaining contacts are not technically harder. The agent can hear the question. It understands the intent. It has the full account history. What it does not have is an answer to the question underneath the question: What is this customer entitled to?
The Four Contact Categories That Hit the Policy Wall
The 40% of contacts that AI agents cannot resolve without a policy layer are not randomly distributed. They cluster into four identifiable categories. Each shares the same structure: The agent can handle the investigation. It cannot authorize the resolution.
Billing disputes and credit requests (20 to 25% of all contacts)
A customer contacts support about a disputed charge or service failure and wants a credit or refund. The agent can investigate and explain: verify the charge, trace the usage, confirm what was billed and why. The policy wall appears when the customer asks for relief. The agent cannot determine what credit the situation justifies, whether a credit has already been applied this billing cycle, or what the maximum adjustment within policy is for this account type.
Plan changes and upgrade eligibility (10 to 12%)
A customer wants to change their plan, add a line, or upgrade their device. The agent knows what plans exist. It cannot determine whether this customer is eligible for the requested change, what the proration terms are, or whether a device upgrade requires a contract extension under current policy.
Retention offers (8 to 10%)
A customer is threatening to cancel. The agent detects the churn signal. It cannot determine what retention offer this customer qualifies for, whether they have already received one in the past 90 days, or what the approval path is for an exception beyond the standard offer.
SLA and outage compensation (3 to 5%)
A service disruption occurred. The customer is asking what they are owed. The agent can confirm the outage. It cannot determine the compensation this customer is entitled to under the mobile operator's service level commitments.
None of these are AI capability problems. They are policy infrastructure problems. The agent does not know the answer because the answer does not exist anywhere the agent can reliably access.
Stage 1 vs. Stage 2: What the Ceiling Costs
The economics are documented for a mobile operator with 10 million subscribers and 22 million inbound contacts per year. The baseline cost of customer support before AI deployment is $440 million per year. The question is how far AI can take it.
Stage 1: Limited AI, no policy layer.
The first wave handles technical, informational, and self-service contacts. It saves $168 million per year, a 38% reduction. Then it stalls.
| Handling type | Share | Annual volume | Cost / contact | Annual cost |
|---|---|---|---|---|
| AI fully contained | 30% | 6.6M | $0.50 | $3.3M |
| Human + AI assist | 30% | 6.6M | $14.00 | $92.4M |
| Full human (policy decisions) | 40% | 8.8M | $20.00 | $176.0M |
| Stage 1 total | 100% | 22.0M | $12.36 | $272M |
Stage 2: Full AI with Polidex as the policy layer.
The agent calls Polidex before acting on a policy-requiring contact. Polidex returns a versioned, auditable decision: What the customer is entitled to, what rules apply, what requires human approval. The four contact categories that stalled at Stage 1 now resolve autonomously, with a record on file for every decision.
| Handling type | Share | Annual volume | Cost / contact | Annual cost |
|---|---|---|---|---|
| AI, no policy engine | 30% | 6.6M | $0.50 | $3.3M |
| AI, Polidex-enabled | 30% | 6.6M | $0.56 | $3.7M |
| Human + AI assist | 20% | 4.4M | $14.00 | $61.6M |
| Full human (exceptions) | 20% | 4.4M | $20.00 | $88.0M |
| Polidex platform fee | $0.6M | |||
| Stage 2 total | 100% | 22.0M | $7.14 | $157M |
AI containment moves from 30% to 60%+. Effective cost per contact drops from $12.36 to $7.14. The $115 million gap between Stage 1 and Stage 2 is the gap a policy layer closes. It is not recoverable by any other means.
The ROI on a Policy Layer
$115 million in incremental annual savings, unlocked by moving from Stage 1 to Stage 2 with a policy layer.
$7.14 effective cost per contact at Stage 2, consistent with BCG and McKinsey benchmarks for AI-first operators at mature deployment.
182x ROI on Polidex. At ~$630K per year for a 10-million-subscriber operator, Polidex represents $0.063 per subscriber per year. Less than 0.015% of average revenue per user. It does not appear in any meaningful cost-per-subscriber analysis. What does appear: whether the AI stack reaches 60%+ containment or stalls at 30%.
What This Means for You
If you have not yet deployed agentic customer support.
Build the policy layer in from day one. Operators who define their policy layer before deployment do not hit the ceiling at all. They go straight to 60%+ containment, with the full ROI available from the start, and skip the 18-month detour at 30% containment that comes from prompt-based architecture. The architecture decision is the single highest-leverage choice in your deployment.
If you are already at Stage 1.
The ceiling is real, it is structural, and it is costing you more than $115 million per year in savings the AI stack cannot reach without a policy layer. Better models do not solve a policy infrastructure gap. More training data does not version a refund policy. Prompt engineering does not produce an audit trail. The investment required (~$630K per year for a 10-million-subscriber operator) is a rounding error against the savings it unlocks. Every quarter at 30% is a quarter a competitor who built the policy layer first extends their lead in cost-per-contact and autonomous resolution rate.
Related
- The architecture decision: the two paths for agentic customer support, and why only one reaches 60%+ containment.
- Compensation and credits: the SLA and outage category, in detail.
- Refund and return policy: the billing dispute category, in detail.
- The policy gap: the structural reason policy in system prompts cannot scale.
Working through how to deploy agentic CS?
If you're at a mobile operator or enterprise evaluating agentic AI for your operation, we'd welcome a conversation about what containment is realistic, what the policy layer needs to look like, and how to make the deployment defensible.
Start a Conversation