How to Estimate AI Chatbot Costs for a Business Website
Estimate website chatbot costs from the actual connection, conversation workload, token or credit accounting, setup, support and completed customer tasks.
Estimate an AI chatbot's cost by identifying its billing arrangement, measuring representative conversations and adding the software, setup and support needed to run it. Visitor count alone does not tell you the bill: many visitors never start a chat, and one conversation can involve several model requests.
With Niwa, begin by confirming the connection. A supported own-subscription connection has no added Niwa AI usage charges or Niwa-imposed usage limits on that connection; the provider's terms and allowances still apply. If you choose an API-based arrangement, estimate the provider's separately billed usage. Niwa's connection guide shows these as distinct setup choices.
Define the customer work you are estimating
Choose a concrete task, such as helping a visitor find an available product variation or choose a service and complete a booking. Record what counts as a useful outcome and where a person takes over an unresolved question.
A product recommendation may need catalogue information and a follow-up about size. A booking workload includes service selection, availability checks and saving the chosen reservation. Count the model requests and tool interactions needed to carry that task through.
Keep estimates for these tasks separate at first. A long product comparison can have a different workload from a short opening-hours answer. Combining them into one assumed "message cost" hides the mix you need to budget for.
Build a representative usage sample
Record the expected number of conversations during a defined period, the approximate requests per conversation and the model used. Test with representative questions and the languages your customers use. Count measured tokens or service credits rather than guessing from an English word count.
For an API connection, use the provider's response usage fields and account records. OpenAI documents separate input and output fields and notes that some models also report cached-input and reasoning usage. Its dashboard uses UTC, which matters when comparing a business day with local logs. OpenAI usage and cost guidance.
Include retries and work that fails to produce an outcome. A customer asking twice after an error still consumes resources. Also separate setup work, such as initial indexing, from the normal customer workload.
Calculate an API estimate with the right categories
Use the current rates for the exact model and processing mode. Where the provider distinguishes uncached input, cache writes, cache reads and output, calculate each billable category separately and add the results. Check whether cached usage is reported as a subset of total input so you do not count it twice.
For a category priced per million tokens, the calculation is: category tokens divided by 1,000,000, multiplied by that category's rate. Add any tool, storage, audio or other charges your implementation actually uses. The provider's pricing page is the place to verify rates, not an old comparison screenshot.
A worked arithmetic example
On October 9, 2026, OpenAI's Standard short-context prices for gpt-6-luna list USD 0.10 per million uncached input tokens, USD 0.125 for cache writes, USD 0.01 for cached input and USD 0.50 for output. Use this as a worksheet example, rather than a model recommendation or a quoted Niwa service price.
Suppose a hypothetical workload, after separating its billing categories, contains 4 million uncached input tokens, 2 million cache-write tokens, 2 million cached-input tokens and 1 million output tokens. The model estimate is:
- Uncached input: 4 × USD 0.10 = USD 0.40.
- Cache writes: 2 × USD 0.125 = USD 0.25.
- Cached input: 2 × USD 0.01 = USD 0.02.
- Output: 1 × USD 0.50 = USD 0.50.
That totals USD 1.17 for those model categories only. The workload is illustrative, not measured business usage. It excludes other services, taxes, regional processing uplifts and any different context or processing rate. Replace every volume and rate with your own measured categories and current terms before using the figure as a budget.
Estimate a credit plan using its own rules
For a hosted assistant, identify the included allowance, how the selected model consumes it and how extra credits are charged. Chatbase's pricing page, for example, lists message-credit allowances and additional-credit options. Those units should not be substituted for raw API tokens.
Run a representative conversation and inspect the service's usage record. Then estimate the credit requirement for your conversation mix. Check what happens at the allowance: interruption, extra purchase and automatic recharge are materially different arrangements.
For Niwa's supported own-subscription route, budget for the provider subscription and confirm its account conditions. Do not add a fictional per-token Niwa surcharge to that connection.
Add the costs outside inference
Include the plugin or service licence, preparation, configuration, hosting not already included, support and the team's review time. The website cost guide places chatbot operation alongside the broader website budget.
Niwa Business is USD 150 for a custom WordPress website or migration, hosting, maintenance and adapted AI tools. Ask for the billing schedule and separate provider costs. Existing WordPress owners can ask about the standalone plugin without adding a new website to the estimate.
Finally, compare costs with completed useful tasks, not only messages sent. Record unresolved questions and human follow-up as well. A cheap answer that needs repeated correction may be less useful to the business than a reliable completed journey.
Discuss your chatbot workload with Niwa. Share the customer task, a representative question set, expected conversation volume and your preferred AI connection. Ask for a setup proposal that separates provider usage from implementation and care.
