SERVICE ECONOMICS · OCTOBER 2026
A cheaper interaction can still create a more expensive customer problem.
Optimize the total cost of correctly resolved issues, not the number of chats avoided. Include platform usage, human work, knowledge upkeep, quality review and recovery, and separate cash savings from released capacity.
Start with a cost and issue baseline
Choose a period and a consistent issue definition. A customer may contact you through several channels about the same problem; counting every contact as a separately resolved issue can distort both volume and cost. Record how you link repeats and which reopen window you use.
Include agent handling, supervision, tooling, outsourced service, training and relevant operational overhead. Keep cost allocation assumptions visible. A fully loaded cost measure is useful for operating comparison, while an incremental cost measure helps decide whether a new workflow is worth adding. Do not mix them in the same ROI calculation.
Break the baseline down by intent, channel and complexity. A password-navigation question differs from a disputed payment or technical incident. Select potential improvements based on observed workload and risk rather than claiming that an invented top-twelve list always represents a fixed share of tickets.
Use a formula with a defensible denominator
Full cost per resolved issue equals the period’s attributable service costs divided by issues meeting your resolution definition. The numerator includes AI platform and usage charges, human handling, exception review, knowledge maintenance, quality assurance and recovery. The denominator excludes unresolved or duplicate issue records according to the policy you document.
For a before-and-after comparison, keep the business perimeter and definitions stable. If the new system expands coverage, changes hours or serves a different customer mix, report those changes. Better service may justify higher cost; lower cost alone does not establish success.
Track vendor-billed outcomes separately and reconcile them to your own quality-reviewed cases. Read the supplier’s definitions before forecasting fees. Fin billing and outcome definitions
Read the contract before forecasting usage
Identify the billable unit, eligibility rules, minimum commitments, volume tiers, included seats and additional charges. Ask how failed attempts, reopened conversations and handoffs affect the invoice. Use a representative export to reconcile the provider’s usage count with your cases.
Intercom’s pricing page distinguishes seat and outcome components. API-based implementations add model input, output and potentially tool charges, which vary with context and workload. Estimate usage from observed traces rather than a universal cost per support ticket. Intercom pricing; API cost components
Add the costs that rarely appear in a product demo: updating policies, reviewing exceptions, maintaining integrations, investigating incidents and training agents. Assign an operating owner and budget. An apparent saving created by omitting these duties will disappear when the system enters real service.
Improve the work before increasing autonomy
First remove avoidable demand: unclear product instructions, broken status pages, confusing invoices or missing policy explanations. A better customer journey may eliminate more effort than automating the same confusion.
Next improve agent assistance: retrieve context, prepare drafts and summarize handoffs. Then automate bounded information tasks. Permit consequential actions only when authority and recovery are proven. Compare the incremental benefit and risk at each step rather than assuming the most autonomous design is cheapest.
Use intent-level quality sampling and preserve customer access to a person. If an automation causes repeated contact, include that effort in its cost. The historical NBER support-assistance study is evidence that augmentation can help in a studied setting; it is not a universal cost-saving guarantee for your operation. Historical support-assistance research
Test the business case under realistic conditions
Run a baseline and pilot with comparable cohorts or a staged rollout where practical. Record volume, correct resolutions, reopens, handling time, escalation workload and customer satisfaction. Do not attribute every improvement to AI if staffing, policy or product quality changed at the same time.
Model downside sensitivity: lower automation eligibility, more review, higher usage and slower incident recovery. Include implementation expense separately from ongoing operating cost. Set a review date and conditions for expansion, renegotiation or stopping the pilot.
If the goal is capacity release, name the work people will do with the saved time. If the goal is cash reduction, identify the actual budget line that changes and when. Avoid declaring employee hours as realized savings when payroll and workload remain unchanged.
Compare the choices
| Lever | Potential benefit | Measure the risk |
|---|---|---|
| Remove avoidable demand | Fewer issues enter support | Check whether customers can still obtain help. |
| Assist human agents | Less context gathering and drafting | Correction time and acceptance quality. |
| Automate information | Lower routine handling effort | Wrong answers, abandonment and reopens. |
| Execute bounded actions | Less coordination and completion time | Unauthorized actions, recovery and incident cost. |
Worked example: make the decision concrete
Illustrative monthly scenario: baseline attributable cost is $20,000 for 4,000 resolved issues, or $5 per issue. In the pilot, human and overhead costs are assumed to be $17,000, platform and usage $2,000, and added maintenance/review $1,000. Total cost remains $20,000. If only 3,800 issues meet the same resolution definition, cost is about $5.26 per resolved issue.
The pilot might still improve availability or satisfaction, but it has not demonstrated cost reduction on these assumptions. To show $2,000 monthly cash savings, the team needs a credible $2,000 reduction in actual spending after all added costs. This example is transparent arithmetic, not data from a claimed private deployment sample.
Failure modes to plan for
Beware of denominator inflation, ignored repeat contact and shifting work to customers. A closed conversation may hide an unresolved issue; investigate a sample rather than automatically treating silence as success.
Another failure is double counting: platform-billed outcomes, internal resolutions and human handoffs may all relate to one issue. Maintain a case identifier across systems. Include incident recovery and unauthorized-action losses in governance decisions even when they are infrequent; a low average cost does not excuse unacceptable service risk.
Your useful next step
Build an issue-linked cost worksheet for one intent cohort. Reconcile provider charges, human review and reopens, then compare full cost and customer outcomes using the same definitions before expanding automation.
Questions and answers
What is the best cost metric?
Use full cost per correctly resolved issue with a documented issue and reopen definition. Keep incremental costs and vendor-billed outcomes as separate supporting measures.
Do saved employee hours count as savings?
They count as released capacity when measured. Cash savings require an actual spending reduction; explain how capacity will be redeployed if costs remain unchanged.
Should we pay per outcome or per seat?
Compare the complete contract against your expected workload and outcome definitions. Neither billing model guarantees lower total cost or better service.
How quickly should the investment pay back?
There is no universal payback period. Use measured operating benefit, implementation expense and realistic downside assumptions, then review against your organization’s investment criteria.
Sources and evidence scope
Primary sources checked on 11 October 2026. Product pages establish vendor-described features; they do not establish independent comparative performance. Worked examples are explicitly illustrative. Implementation recommendations are editorial guidance.
- Intercom Fin outcome definitions — Defines billable Fin outcomes and resolution categories.
- Intercom — Fin usage pricing is a vendor billing measure; it does not prove industry-wide independently verified resolution rates.
- Openai Prices — API cost requires separate input/output, caching, tool charges and actual usage.
- Nber — Early support-assistant study reports 14% average issues-resolved/hour improvement; human augmentation differs from autonomous replacement.
Research Path