The cost of implementing a chatbot is the cost of a defined job, not a universal price per bot. Add one-time content preparation, configuration or engineering, integrations, testing and launch work. Then add the subscription or infrastructure, usage charges and the staff time needed to keep answers current. A free software plan can still require paid staff work.
Reviewed October 7, 2026. If your quote has a “training data” line, ask what source preparation, approval and updates it covers. The source-work example below shows how to estimate that work.
If you are comparing quotes, copy this record for each proposed implementation. Ask every supplier to fill in the same fields before comparing totals:
Audience and supported questions: Source inventory (pages, files, owners, update frequency): Required actions and integrations: Channels, languages and access restrictions: One-time work: task | owner | hours | loaded rate or fixed fee Monthly work: task | owner | hours | loaded rate or fixed fee Software, model and hosting: billing unit | allowance | expected usage | overage rule Acceptance test: real questions | expected sources | pass rule | handoff cases Excluded work and unresolved prices: Year-one estimate = one-time work + 12 × monthly work + annual charges
A loaded rate includes salary, taxes, benefits or agency overhead. Use your own rate. Mark an unpriced item unknown, rather than entering zero. The editable chatbot budget worksheet gives you a separate place to collect the numbers; its amounts are blank and it has no automatic formulas.
What belongs in the implementation estimate?
A useful quote separates the following work. If a line is included in a flat fee, ask what the fee actually covers.
| Workstream | Evidence to request before pricing |
|---|---|
| Content preparation | Number and size of sources; who checks outdated or conflicting answers; who approves corrections |
| Configuration or build | Supported topics, model and retrieval setup, website placement, access and domain rules |
| Integrations | Each system, read or write permission, authentication, failure path and test environment |
| Evaluation and launch | Real questions, expected answers, handoff cases, accessibility check, owner and release condition |
| Recurring operation | Usage unit, subscription or hosting, source updates, answer review, software maintenance and support |
This record also prevents double counting. If a vendor's onboarding fee includes document preparation, do not add the same hours again. If it covers only uploading the files, keep editorial review as a separate line.
A full chatbot pricing comparison covers published plan prices and billing units. This page concerns the work required to make a chosen option useful and keep it useful.
A worked budget, with the assumptions exposed
Consider a small website chatbot that answers approved help pages and sends difficult questions to a person. It has no account lookup or payment action. Suppose the owner uses a hosted plan priced at $119.99 per month, and estimates these internal costs:
| Item | Assumption | Cost |
|---|---|---|
| Review and correct sources | 12 hours × $60/hour | $720 |
| Configure and place chatbot | 6 hours × $90/hour | $540 |
| Test answers and handoff | 8 hours × $60/hour | $480 |
| Subscription | 12 months × $119.99 | $1,439.88 |
| Review conversations and update sources | 2 hours/month × $60/hour × 12 | $1,440 |
| Illustrative first year | $4,619.88 |
The $119.99 is Agentkit's published Standard monthly price, checked October 7, 2026. At that check, Standard included 12,000 messages per month, two chatbots and three team seats. Confirm the current price and plan limits, including source capacity and billing interval, before buying. The hours, loaded rates and monthly review load above are assumptions, not measured customer averages. The estimate excludes taxes, additional usage, outside services and any custom integration. Estimate messages from observed visitor chats, then check what happens at the allowance; do not assume that extra messages can simply be bought.
To compare a custom build, keep the same audience, sources and acceptance test. Replace the configuration line with engineering tasks for retrieval, the website interface, security, deployment and monitoring. Add model and hosting bills, upgrades and an owner for production incidents. A model API price alone is not a comparable custom-build estimate. Get scoped hours and rates for those tasks; a broad market range cannot tell you what your own integration costs.
Price the source work, not a vague “training data” line
For a website chatbot that answers from pages and documents, ask what happens to each source before it is used. A page may need an owner, a current policy check, duplicate removal, or a rewrite. A PDF may need extraction and a check that tables survive. Someone must decide which source wins when two pages disagree. That work has a labor cost even if the platform includes ingestion.
For example, suppose a content owner must inspect 40 help pages at 10 minutes each. Six pages need another 20 minutes of correction, and eight missing answers take 15 minutes each to write. The estimate is 40 × 10 + 6 × 20 + 8 × 15 = 640 minutes, or 10 hours 40 minutes. At an assumed loaded rate of $60/hour, that is $640. Add a second reviewer's time if approval is required. These counts and rates are illustrative; a short FAQ and a regulated policy manual need different review effort.
“Training data pricing” can also mean labeling examples for an intent classifier or fine-tuning a model. Those are different projects. Do not assume the quote includes them, or that they are required for a chatbot that retrieves existing content. Ask whether the fee covers source collection, parsing, editorial correction, Q&A writing, evaluation questions, re-indexing and ongoing updates. The source preparation guide explains how to turn approved material into a useful answer set.
When an answer changes later, identify who updates the source and checks the bot again. Scheduled website retraining, if available on the selected plan, does not decide whether the revised source is correct. Uploaded documents also need to be updated when their contents change.
Integrations change the scope
A website embed and a CRM write are different jobs. For every proposed action, document the input, destination, permission, expected output and failure response. Test with a non-production account or environment where possible. Budget for the person who can approve access and for the person who will investigate failed records after launch.
A quote that says “CRM integration included” should name the exact workflow. Is the bot collecting a lead, creating a CRM record, or reading existing customer information? Does a failed write alert anyone? Is there a duplicate check? An integration can be technically connected while still losing leads or exposing more data than intended. The chatbot integration guide provides a production checklist.
An embed snippet shows how to place code, but it cannot estimate the surrounding work. Copy the current code from the product's Channels screen, place it on a private test page and verify it there before public rollout. Price the setup and acceptance work rather than promising a fixed number of minutes to paste code.
Keep the calendar and the budget separate
Hours determine labor cost. Dependencies, staff capacity and approval waits determine elapsed time. A task can cost four hours of work and still delay launch for a week while a data owner approves the source. Record effort and available capacity beside every task. The build-time planning guide has a copyable work-item record and a worked calendar calculation.
Before launch, choose questions from actual site traffic or support logs and record the source each answer should use. Include a missing-answer case, a conflicting-source case and a human-handoff case. Name the pass rule and the owner who can reject the release. A demo that answers a few handpicked questions is not an acceptance test for the published scope. The QA sampling guide can help you keep checking answers after launch.
What should an implementation quote include?
Ask for the workstream record near the top of this article, plus these terms in writing:
- Who owns the chatbot account, source files, configuration and conversation data.
- Which tasks are one-time, which repeat monthly, and which are outside the fee.
- The subscription or usage unit, allowance, overage rule and billing interval.
- The exact systems an integration reads or writes, and who supplies credentials.
- Acceptance questions, source checks, handoff behavior and a release decision owner.
- Support, monitoring, incident response, export and cancellation terms.
A quote is incomplete if it calls content preparation or maintenance “free” without naming the person who will do it. Revisit the estimate after a small pilot using real message counts, corrections and review time. Replace assumptions with those observations.
If a hosted website chatbot meets your required scope, check Agentkit's current plans. For unusual data access or workflow requirements, use this same scope record to get a custom engineering estimate.
