AI agent cost for a business: from the pilot to ongoing operations

A usable pilot has a written scope, test data, limited access, and verifiable results. It must handle usual requests and show what happens in case of failure. A demonstration with a single question makes these costs hard to estimate.

Ask what items you will receive: workflow, business rules, test set, operational documentation, and suspension procedure. Also specify the changes included after the first review. The time an employee spends explaining the process and validating results is part of the project, even if it does not appear on the contractor's invoice.

An assistant consulting documents can start with read access. A workflow that writes in the CRM must handle approval, errors, and duplicates. A voice agent adds telephony, transcription, speech synthesis, and transfers. These functions explain budget deviations that a "per agent" price easily masks.

These scenarios are fictional. They use a common rate of CA$120 per hour to illustrate the calculation, without constituting an offer or a Daillac pricing range. The durations assume available access and a limited scope. Taxes, subscriptions, and telephony costs are excluded from the initial budget.

A CRM without a test environment, poorly filed documents, or conflicting service rules can alter these durations. Request that the quote separates data reorganization from the agent integration. This way, you can decide which tasks should precede the pilot.

In a voice project, specify the languages, schedules, and supported exceptions. The twelve tests before deploying a voice agent guide provides a template to define the acceptance tests. A confirmed booking deserves a more thorough verification than a simple transcription.

As of October 12, 2026, Anthropic pricing grid shows for Claude Haiku 5.5 a price of $0.10 USD per million input tokens and $0.50 USD per million output tokens for requests whose prompt does not exceed 100,000 tokens. The grid distinguishes an upper tier. These model rates do not cover integration or other agent services.

For a simulation of 1,000 requests, with 4,000 input tokens and 500 output tokens per request, the calculation gives 4 million input tokens and 0.5 million output tokens. At the indicated tier, this represents $0.40 USD plus $0.25 USD, totaling $0.65 USD. This example excludes retries, tool calls, caching, and other fees. This amount should not be directly compared to a budget in Canadian dollars.

The chosen model must pass your quality tests. A task that requires long research, multiple conversation turns, or detailed answers will consume more. The news on Haiku 5.5 and the cost of support tasks explains why the cost of a resolved request is more useful than an isolated token rate.

The orchestration platform can charge a subscription or consumption. n8n pricing are notably based on workflow executions; other tools use credits per action. Check the frequency of triggers, retries, and auxiliary workflows. An automatic query every few minutes can consume resources even when no new requests arrive.

Here is a second simulation, in Canadian dollars. We consider $100 in monthly services and five hours of supervision at CA$80 per hour. The total is CA$500 per month. Out of 800 requests, of which 720 are accepted according to the pilot criteria, the simulated operating cost is 500 ÷ 720, or about CA$0.69 per successful task. The initial creation budget remains separate.

If you amortize a hypothetical initial budget of CA$9,600 over twelve months, you add CA$800 to the monthly calculation. The total then rises to CA$1,300, or about CA$1.81 per successful task for the same 720 successes. This choice of amortization is used for project management; your accounting treatment may follow other rules.

Account for the time spent on refusals, corrections, and transferred requests. An agent may cost little in model calls but require a lot of proofreading. Also measure the work avoided on accepted requests, using the same method before and after the pilot. guide to enterprise AI agents helps define this measure without turning all activity into claimed gain.

The quote should specify the connected tools, accessible fields, and authorized actions. Ask who owns the accounts, where the traces will be stored, and how to export workflow elements. Check what the contract covers when the CRM API or the model changes.

Have the acceptance criteria and exclusions recorded. For a customer request, success can be a correct record approved by the employee. For a voice agent, it can be a reservation actually recorded and read back to the customer. The success rate only makes sense with this defined result.

Reserve a follow-up budget after the pilot and designate the person who can suspend the automation. Our AI integration service can scope the scope before costing, based on available access and the requests you wish to process.