Enterprise AI Total Cost: Cost per Business Outcome

A cheap pilot can become an expensive operating model

A pilot's monthly model bill is not its enterprise total cost. Production adds retrieval infrastructure, workflow execution, tool API charges, observability, security, evaluation, support and the cost of correcting unsuccessful work. Token pricing matters, but so do retries, concurrency, context growth and the number of actions required to complete a business transaction.

Use a cost-per-outcome model

Calculate cost per completed business case, not only cost per request. A simple baseline is: total operating cost = inference + retrieval/data + workflow/integration + infrastructure + observability/security + support and human review. Divide by successfully completed cases, not initiated runs.

Conceptual reference pattern; enterprise policies and integrations determine implementation.

Example: why retry rates matter

Suppose a workflow starts 100,000 cases per month. A nominal €0.04 model spend per attempt suggests €4,000. If each case requires four model calls, that component becomes roughly €16,000 before retries, retrieval, integration or operational costs. These are illustrative inputs, not market prices or an Archy benchmark.

Choose a deployment model with eyes open

DimensionVendor-hosted serviceCustomer-hosted platformInfrastructureTypically bundled or usage-pricedOperated and paid by customerModel consumptionMay be bundled or meteredCustomer provider accounts and chargesOperationsShared responsibility variesGreater customer operational responsibilityControlVendor deployment constraintsMore direct deployment and data controlsEconomicsLower entry friction; variable spendLicense plus infrastructure and AI usage

Neither option is inherently cheaper. Compare equivalent reliability, security and support obligations rather than only invoice lines.

Model independence is a design capability

Portability requires more than changing an endpoint. Prompts, tool calling conventions, context limits, retrieval quality, evaluation results and compliance approval all affect substitution. Maintain a tested model adapter layer and evaluate quality, latency and unit economics by workload.

Cost controls that do not undermine quality

Classify tasks before model routing, keep context bounded, cache stable evidence when policy permits, use deterministic code for deterministic work, cap expensive loops and monitor failure-driven spend. Document quality thresholds so optimization does not quietly degrade business outcomes.

What Archy's enterprise model changes

For customer-hosted deployments, Archy Platform is licensed separately while customers connect their own LLM providers and pay model usage and infrastructure directly. This improves transparency and choice; it does not eliminate those costs or provide native usage throttling by itself. Evaluate governance and external quota management accordingly.

The denominator changes the decision

Continue the illustrative example with a 90% successful completion rate. If all 100,000 initiated cases consume the four-call baseline, inference alone is about €0.178 per successful case: €16,000 divided by 90,000. At 99% success, the same spend is about €0.162. Neither figure includes the work required to resolve the unsuccessful cases. A cheaper route that fails more often must therefore be compared with the whole recovery path, not just its first attempt.

Human review can dominate the difference. If 10,000 cases each require two minutes of review, that is roughly 333 hours. Apply the organisation’s actual loaded labour rate to that time, and separate capacity released from cash savings. The calculation does not establish that headcount can be removed; it shows the operational capacity consumed by the workflow. These are scenario assumptions, not measured customer results.

Build a ledger that finance can challenge

Attribute model calls, retrieval, tools and reviewer time to a workflow version and business case. Shared platform costs need an explicit allocation method. Record unsuccessful and abandoned cases as well as completed work, otherwise the cost model rewards hiding failures. Change one routing or context policy at a time and compare results on equivalent cases with the same acceptance threshold.

The resulting decision is practical: which route meets the required outcome quality at an acceptable cost and response time? Some workloads justify expensive reasoning; others need a simpler deterministic path. Financial governance is stronger when that choice can be reproduced from evidence rather than justified after the invoice arrives.

Explore Archy Platform.

A reusable cost-per-outcome worksheet

Monthly cost componentInput to collectInferenceCalls, tokens, retries and actual provider chargesRetrieval and toolsIndexing, storage, search and integration chargesOperations and controlsInfrastructure, monitoring, security and support allocationHuman review and recoveryReview hours multiplied by the loaded labor rateSuccessful outcomesAccepted completed cases; exclude failed and abandoned cases from this denominator

Add the monthly cost components and divide by accepted completed cases. Keep failed-case spend in the numerator. Compare the same workload, quality threshold and reporting period; this worksheet is a calculation method, not a measured Archy result.

For the related architecture decisions, see LLM Model Routing: Cost per Successful Workflow; Enterprise AI Workflows: State, Approvals and Recovery.