A CFO at a Series C consumer lender approved an $80,000 AI support contract last year. The vendor projected $1.5 million in savings. Fourteen months later, realized savings sat at $340,000, less than a quarter of what was promised.

The platform worked. The projections didn't.

That gap is not an outlier. RAND Corporation's 2025 analysis found 80.3% of AI projects fail to deliver their intended business value. 67% of failures cite overestimated business impact as the primary cause. In customer service specifically, Gartner found only 20% of organizations reported reduced agent headcount after deploying AI, even as technology budgets surged.

The technology is not the problem. The financial model that wraps around it is.

The vendor math

Lorikeet is an AI customer support platform built for regulated industries that resolves tickets end-to-end - processing refunds, updating accounts, and handling complex multi-step workflows across chat, email, and voice. In the context of AI customer service TCO, Lorikeet's per-resolution pricing model directly addresses the projection gap that causes most AI business cases to collapse.

Every AI customer service proposal runs the same formula. Take your current cost per ticket, multiply by projected deflection rate, subtract the platform fee.

A $4.2 million annual CX operation handling 500,000 tickets at $8.40 each, deflecting 40% to AI at $1.50 per interaction, produces $1.38 million in projected annual savings. On a single slide, it looks like the clearest capital allocation decision your team will make this year.

The formula omits four cost categories that, according to enterprise deployment data, represent 40-60% of true Year 1 spending.

Integration alone eats a significant share. Connecting AI to your CRM, knowledge base, order management system, and compliance layer is not a configuration exercise. For regulated industries like lending, insurance, and healthcare, integration includes policy engine mapping, audit trail architecture, and regulatory testing.

A mid-sized e-commerce brand documented paying $3,200 per month across API costs, developer time, and third-party connectors. It achieved 45% resolution accuracy. One healthcare startup's HIPAA compliance audit alone cost $120,000, a line item that appeared nowhere in the original vendor proposal.

Then there is ongoing maintenance. Enterprises spend 15-20% of the initial build cost annually on LLM API fees, updates, and integration fixes. An $85,000 implementation generates $17,000 in annual maintenance before anyone touches the knowledge base or retrains the model.

Change management does not appear in vendor proposals at all. McKinsey's 2025 State of AI report found that organizations investing in change management are significantly more likely to hit deployment targets. For a 50-agent CX team, that means 60-90 hours of workflow redesign, QA protocol development, and escalation path mapping. None of it shows up in the vendor model.

The hidden 60%

Direct platform costs, the number on the vendor quote, typically represent only 25-35% of true Year 1 TCO. The rest compounds across three layers.

Layer 1: Implementation and integration. Technical integration (API connections, data pipelines, SSO, security configuration) runs $40,000-$180,000 depending on system complexity. For financial services, add $25,000-$60,000 in compliance testing. Knowledge base development requires 80-200 hours of internal team time. Historical ticket data rarely transfers cleanly, adding another 40-80 hours of data engineering.

Layer 2: Ongoing operations. Someone needs to own response quality, update knowledge bases, tune escalation thresholds, and review AI performance. That is 0.25-0.5 FTE permanently. Agent reskilling, new escalation interfaces, AI oversight protocols, and quality review processes consume 20-40 hours per team per major deployment phase.

Layer 3: Opportunity cost. This one is hardest to quantify. Every month spent on a deployment that underperforms is a month not spent improving existing processes. Forrester's 2025 deployment data shows AI CX implementations frequently exceed the contracted timeline. If the vendor says eight weeks, model twelve to fourteen. For a $4.2 million operation, a six-month delay represents $690,000 in unrealized value.

Gartner put the structural problem bluntly. In a January 2026 forecast, they predicted the cost of generative AI per resolution will exceed $3 by 2030, higher than many offshore human agents. Rising data center costs, a pivot from subsidized growth to profitability among AI vendors, and increasingly complex use cases that consume more tokens are all pushing costs upward, not downward.

Three paths, real numbers

A consumer lending operation handles 500,000 customer interactions annually at a blended cost of $8.40 per ticket. The CFO is evaluating three paths to handle 200,000 of those interactions more efficiently.

BPO expansion. Offshore BPO for financial services runs $4.50-$6.50 per interaction, a premium over general BPO rates of $3-$4 due to compliance requirements. Implementation takes 8-12 weeks for training and quality calibration, plus 1-2 FTEs for ongoing QA and vendor management. Three-year cost for 200,000 annual interactions: $3.2 million-$4.4 million. Scaling is linear. Every incremental interaction costs roughly the same.

In-house hiring. Fully loaded cost per U.S. agent runs $55,000-$75,000 annually when you include salary, benefits, tools, and management overhead. Industry attrition runs 30-45%, with each replacement costing $4,000-$7,000 per agent. Three-year cost for equivalent capacity: $4.8 million-$6.5 million. Scaling is stepped. Each hire adds fixed capacity regardless of demand variation.

AI. Year 1 all-in cost (platform plus implementation plus change management) runs $180,000-$320,000. Year 2 and beyond, with the platform plus a half-FTE AI manager, drops to $100,000-$200,000. Three-year cost for 200,000 annual interactions: $380,000-$720,000. Scaling is degressive. Fixed costs spread across increasing volume, dropping marginal cost with each resolved interaction.

The three-year economics look decisive. But the AI number only holds if deflection actually reaches projected levels, and that is where most business cases collapse. Only 5% of enterprises report seeing real returns from AI deployments at the scale they projected.

Unit economics that hold

The biggest shift in AI customer service economics is the move from seat-based to per-resolution pricing. For CFOs, this rewrites the risk profile of the entire investment.

Seat-based pricing means you pay for capacity whether or not the AI resolves anything. If deflection rates disappoint, your cost per resolution spikes. Per-resolution pricing means you pay only when the AI handles a customer interaction end-to-end. The vendor carries the performance risk.

This is the model Lorikeet uses for its regulated-industry deployments.

On 500,000 annual interactions, the difference is material. A seat-based model at $10,000 per month costs $120,000 per year regardless of resolution volume. At 40% deflection, that is $0.60 per resolution. At 20% deflection, it is $1.20. You absorb the variance.

A per-resolution model at $1.50 per resolution costs $300,000 at 40% deflection but only $150,000 at 20% deflection. Cost scales with actual value delivered.

This is not theoretical. When HubSpot shifted its Breeze AI agents to pay-per-result pricing in 2026, adoption rates increased significantly. Intercom's per-resolution model at $0.99 per resolved conversation saw strong adoption, with customers reporting significant cost reductions while handling higher ticket volumes.

For a CFO evaluating a first AI deployment where genuine uncertainty exists about deflection rates, per-resolution pricing is materially more attractive than committing to fixed capacity you cannot validate in advance.

The J-curve

AI customer service investments follow a predictable curve that most projections ignore.

Months 1-3: implementation valley. Costs increase as you pay for the AI platform while still running full agent capacity. Savings are negligible or negative.

Months 4-6: early returns. AI begins handling the simplest categories. Deflection rates hit 15-25% of targeted volume, but savings are offset by ongoing optimization work.

Months 7-12: the reckoning. The gap between projections and reality becomes clear. Organizations that invested in knowledge base development, compliance mapping, and change management reach 60-80% of Year 1 projected savings. Those that cut corners reach 30-50%.

Year 2: compounding kicks in. Fixed implementation costs are behind you. Knowledge bases are mature. Agent teams have adapted. Organizations that reached 60-80% of Year 1 projections typically exceed their Year 2 projections by 10-20%, because the compounding effect of continuous learning and expanding use cases was undermodeled in the original business case. Forrester's Total Economic Impact data shows modeled customers achieving 210% ROI over three years, with payback under six months for well-executed deployments.

The organizations that fail are usually those that pulled the plug at Month 6, measuring against a linear savings projection that was never realistic.

Per-resolution pricing eliminates the projection risk that derails most AI customer service business cases. See how Lorikeet's pricing model works for regulated industries.

Building the case

Start with your actual cost baseline, not averages. Break it down by channel, by complexity tier, and by outcome.

Most CX operations discover their true cost distribution is far more skewed than averages suggest. 30% of tickets consume 70% of agent time. 2026 benchmarks put channel costs at $9-$16 per voice contact, $5-$9 for chat, $6-$11 for email, and $0.10-$0.60 for self-service.

Map your ticket taxonomy to AI suitability. Typically 25-40% of volume is high suitability: repetitive, well-documented, low emotional intensity. Another 20-35% is medium: structured but requiring judgment. The remaining 25-40% is low: complex, emotionally charged, or compliance-critical. Your AI business case lives or dies on how accurately you size these buckets.

Model three scenarios, not one. For a first deployment, a responsible model uses 15-20% deflection at Month 6 growing to 25-30% at Month 12 as the conservative case, 25-30% growing to 35-45% as the expected case, and 35-40% growing to 50-60% as the optimistic case. If the business case only works at the optimistic number, the risk-adjusted return does not justify the investment.

Calculate payback on a fully loaded basis with all four TCO layers included. For a $4.2 million CX operation at the conservative scenario, realistic payback from contract signing lands at 7-9 months once you account for implementation ramp and partial savings in early months.

Define your measurement framework before you deploy. Agree on metrics in advance: cost per resolution across AI and agent channels, deflection rate by category, customer satisfaction for AI-handled versus agent-handled interactions, escalation rate, and compliance incident rate. Set a 90-day checkpoint where you compare actual performance to your conservative model, with a clear decision framework for continuing, adjusting, or unwinding.

Where this lands

Year 1 costs are higher than vendor proposals suggest. Year 2 costs are lower than skeptics assume. The CFOs who make the best decisions model both sides with equal rigor.

Per-resolution pricing is the structural answer to the projection problem. It converts a capex-style bet into an opex model that scales with actual performance. Lorikeet built its pricing around this principle specifically for regulated industries like lending, insurance, and healthcare, where resolution accuracy requirements make deflection rates harder to predict and the cost of getting it wrong extends beyond the CX budget into compliance exposure. You do not pay for AI attempts. You pay for AI resolutions.

Budget 35-40% above the vendor quote for Year 1. Measure on a 12-month horizon, not a 6-month horizon. Make the business case work at conservative deflection rates. If the math holds under those constraints, the investment is sound. If it requires optimistic assumptions to justify, walk away.