A useful chatbot budget separates the initial build from the cost of running and maintaining it. The model's token bill is only one line. Content preparation, website integration, testing, support and keeping answers current can require more attention than the chat box itself.
For a small business, start with the job the assistant should do. Answering questions from approved service information is a smaller project than identifying customers, checking orders and changing bookings.
The budgets below are illustrative planning calculations, not market averages or a quote for your project. They show which assumptions to replace when comparing real offers.
Decide what the chatbot is allowed to do
| Scope | Example | Main work to budget |
|---|---|---|
| Answer from approved information | Explain services and link to relevant pages. | Content cleanup, retrieval, answer testing and website setup. |
| Collect an enquiry | Ask relevant questions and pass the request to a person. | Form validation, reliable delivery, consent wording and failure handling. |
| Read customer-specific data | Look up an authenticated customer's order. | Identity, access control, system integration and audit records. |
| Perform an action | Change a booking or update a record. | Permission checks, confirmation, recovery and stronger operational testing. |
For a first version, a clearly bounded information assistant is often the most useful project to price. It can answer common questions and offer a direct route to a person when the source material is insufficient.
Do not pay for a transactional assistant when a clear FAQ and contact form solve the problem. Look at actual repeated enquiries first. If visitors mainly ask for a phone number or opening hours, making that information visible may be the better improvement.
What the initial build includes
The knowledge work starts before model integration. Someone needs to identify approved pages, resolve contradictory prices, remove expired policies and decide which questions should be handed to a person.
Then there is the interface: a usable mobile panel, readable messages, keyboard access, loading and error states, and an obvious way to close the conversation. The visitor needs to know whether the assistant is answering a question, collecting an enquiry or confirming an action.
The implementation also needs a protected backend, usage controls, logging choices and a way to update its knowledge. A provider's retrieval service can help search approved files; for example, OpenAI file search retrieves relevant material from an uploaded knowledge base. Connecting it does not establish that every answer is correct.
Testing should cover ordinary questions, missing information, outdated material and attempts to push the assistant outside its role. A message that says “ignore the business rules” must not become permission to change what the application allows.
Three example setup budgets
The following uses an assumed labour rate of €75 per hour. The hour ranges are example allocations for planning, not measured delivery times or Design by HL price commitments. Amounts exclude tax and third-party fees.
| Example project | Assumed effort | Illustrative labour budget |
|---|---|---|
| Configure a hosted FAQ assistant with clean source content | 12–20 hours | €900–€1,500 |
| Build a custom website assistant with approved knowledge and enquiry handoff | 40–70 hours | €3,000–€5,250 |
| Add authenticated account access and a limited action workflow | 80–140 hours | €6,000–€10,500 |
Replace the rate and hours with a supplier's itemised estimate. Poor source material, several languages, unusual authentication or an unreliable existing API can move a project outside these examples.
Ask what “included” means. Does the quote cover content preparation, integration, evaluation, mobile testing, staff training and post-launch corrections? Does it include the hosting or only the implementation? These questions make quotes comparable.
Calculate model usage from a conversation
API providers commonly charge for input and output tokens. Input includes more than the visitor's latest message: it can include instructions, conversation history, retrieved information and tool results. Output includes generated text and, depending on the model, other billable processing described in its pricing. Check the provider's current pricing.
Here is an arithmetic example in US dollars. It assumes $2 per million input tokens and $10 per million output tokens; these are sample rates, not a named product offer.
| Assumption | Example value |
|---|---|
| Conversations per month | 1,000 |
| Model requests per conversation | 4 |
| Average input per request, including context | 1,500 tokens |
| Average output per request | 300 tokens |
Requests: 1,000 × 4 = 4,000
Input: 4,000 × 1,500 = 6,000,000 tokens
Output: 4,000 × 300 = 1,200,000 tokens
Input cost: 6 × $2 = $12
Output cost: 1.2 × $10 = $12
Model total: $24 per month
That $24 excludes retrieval tools, storage, hosting, monitoring, retries, support and any additional billable processing. A ten-percent increase in requests with the same averages would add $2.40 to this particular model calculation. It is not a forecast for a real deployment.
Measure the averages in a pilot. A conversation that repeatedly sends large documents or calls several tools will not behave like this small example. Provider bills may also be in a different currency from your development quote.
Budget the recurring work around the model
A monthly budget should include the platform or hosting, any retrieval and storage charges, monitoring, support and time for reviewing answers. If a hosted service charges by conversations or credits, establish exactly what it counts and what happens when the allowance is used.
For a simple labour example, two hours a month at the assumed €75 rate is €150 before tax. Four hours is €300. That time could cover new service information, a review of failed answers and a small regression check. These examples are not a promised maintenance package.
Set a named owner for content changes. If the business changes its cancellation policy, who updates the source and verifies the answer? A chatbot with accurate launch-day information can still become misleading when nobody maintains it.
Also ask how the assistant is disabled or replaced if it stops working. A visible contact route should remain available during an outage or spending-limit event.
Keep customer data within the intended scope
A public FAQ assistant rarely needs access to a full customer database. Give it the minimum information required for its job, and treat authenticated account lookups as a separate implementation decision.
Decide what conversation data is stored, who can read it, how long it is retained and how deletion works across the application and its suppliers. A provider's statement about training use is not the same as a statement that no data is stored. OpenAI's data controls, for example, distinguish training use, monitoring logs and application state.
For actions, enforce permissions in application code. The model's answer must not decide whether a visitor can see another customer's order or approve a refund. Human handoff and input limits are among the controls discussed in OpenAI's safety guidance.
Test whether the assistant earns its place
Build a review set from real, anonymised customer questions. Include straightforward answers, ambiguous requests, outdated information and questions the business cannot answer. Write the expected answer or handoff before testing.
Record whether the response is supported by approved information, whether its links work and whether the visitor can reach a person. For an enquiry workflow, verify that the request actually arrives in the intended system with the right details.
Review a sample after launch. Measure useful outcomes such as successful handoffs or enquiries that require less follow-up. Message count alone does not establish that customers were helped, and a chat that ends is not automatically a resolved issue.
What to include in a pricing brief
Send a prospective supplier the website, the approved source material, the languages, a sample of repeated questions and an estimate of expected conversation volume. Describe any required systems and distinguish reading information from changing it.
Ask for separate setup, recurring platform, usage and maintenance costs, plus the assumptions behind them. Request a small acceptance test covering correct answers, uncertainty, handoff and an outage.
That produces a more useful budget than asking for “a ChatGPT box on the website”. You can start with a narrow, well-maintained assistant and expand when real questions demonstrate the need.
To discuss a website integration, send me your site and a few questions the assistant should handle. Include the systems it would need to use so the scope can be estimated before choosing a tool.
Putting AI to work on your website?
Choosing a model is one part of a working feature. If your business needs a support assistant, content workflow or connection to an existing system, I can help define and build the interface, integrations and checks it needs.