CentriWebAI automation cost comes in three parts: the build, which you pay once; running costs, such as AI model usage, software subscriptions and text or call charges, which you pay every month; and upkeep, which keeps it working as your business and the tools change. Any single figure you are quoted is a mix of those three, so ask for them separately.
We don’t publish price ranges here, because a range that covers both a simple reminder and a custom phone agent tells you nothing. What follows explains what moves each part up or down, shows how model usage is actually charged, and gives you a worksheet to run with your own numbers.
The three parts of AI automation cost
| Part | When you pay | What it covers | What makes it bigger |
|---|---|---|---|
| Build | Once, at the start | Working out the process, setting up, connecting systems, testing | More systems, messy data, custom logic, voice |
| Running costs | Every month | Model usage, platform subscriptions, phone and text charges, hosting | More conversations, longer conversations, larger models |
| Upkeep | Every month or as needed | Monitoring, fixes, updates to information, changes when tools change | More moving parts, frequent business changes |
Most surprises come from the second and third rows. A build quote is easy to compare. The monthly costs are where proposals hide their differences.
Build: what makes it bigger or smaller
The build is mostly people’s time. These are the things that add to it:
- How many systems it connects. An AI that only answers questions on your website is one thing. One that checks your calendar, writes to your CRM and sends a text is three more connections, each to be set up and tested.
- The state of your information. If your prices, services and policies are written down and current, the AI can be set up to answer from them quickly. If they live in someone’s head, writing them down is part of the job.
- How many exceptions there are. A simple booking flow is quick. A flow that changes for new versus existing customers, urgent versus routine jobs and different locations takes longer.
- Voice or text. Phone agents need more testing than text chat, because callers interrupt, mumble and change their minds.
- Testing. Good builds spend real time running the system against real questions before customers see it. A quote with no testing time in it is cheaper for a reason.
The cheapest way to bring the build down is preparation on your side. Write down how the job is done today, including the exceptions, gather your current prices and policies in one place, and decide who will sign off the answers the AI gives. Time an outside builder spends chasing that information is time you pay for.
Running costs: how AI model usage is charged
AI models are charged by the token, a small chunk of text. Providers charge separately for input tokens (what is sent to the model, including your instructions and business information) and output tokens (what the model writes back), usually priced per million.
As at November 2026, Anthropic’s pricing page lists Claude Haiku 5.5 at US$0.10 per million input tokens and US$0.50 per million output tokens for prompts up to 100,000 tokens, and Claude Sonnet 5.5 at US$2 and US$10. Google’s Gemini pricing page lists gemini-3.5-flash-lite at US$0.30 input and US$2.50 output on its paid tier. Anthropic’s page also says its Batch API, for work that doesn’t need an instant reply, takes 50 per cent off both. Check their pages for current prices.
A worked example
Suppose each enquiry conversation on your website sends 20,000 input tokens to the model across all its turns and gets 2,000 tokens back. Those are assumptions for the example; your own figures will differ.
- On Claude Haiku 5.5: 20,000 input tokens is US$0.002 and 2,000 output tokens is US$0.001, so about US$0.003 a conversation. A thousand conversations a month is about US$3.
- On Claude Sonnet 5.5: US$0.04 plus US$0.02, so about US$0.06 a conversation, or about US$60 for a thousand.
Two lessons. Model choice can change usage costs many times over, so ask which model a system uses and why. And for most small businesses, model usage is rarely the biggest line. The subscriptions and the people usually are.
Platform and channel charges
Many systems run inside a platform that charges its own way. Two examples, as at November 2026, from the vendors’ own pages:
- HighLevel’s AI pricing article lists an AI Employee Unlimited plan at US$97 a month per sub-account, with Conversation AI unlimited subject to fair use, and says text and call charges are billed separately.
- Intercom’s pricing page lists its Fin AI agent at US$0.99 per outcome, on top of seat plans.
Per-outcome, per-response, per-seat and per-token pricing all behave differently as you grow. Run your expected volume through each before you compare. Check their pages for current prices.
Upkeep: why it never goes to zero
An AI system is not finished when it goes live. Things around it keep changing:
- Your business changes. New prices, new services, new staff, a new holiday roster. The AI needs to know.
- The tools change. Apps rename fields and change how they connect. Workflows that ran for months stop overnight.
- Models are retired. Anthropic’s pricing page lists older models as retired, and Google’s page notes that two Gemini 2.5 models are deprecated and will route to newer models from 1 December 2026. A system built on a retired model has to be moved and retested.
- Answers drift. Someone should read a sample of real conversations regularly to catch wrong or awkward answers.
Upkeep can be a staff member’s job, an agency retainer or a mix. What matters is that someone owns it. Our guide to why automations break covers what to watch.
A worksheet for your own AI automation cost
Fill this in for the one job you most want to fix. Rough numbers are fine.
- Volume. How many times a month does the job happen? (Enquiries, calls, invoices, bookings.)
- Time. How many minutes does each one take a person today?
- Value. What is a won customer worth to you, or an hour of the person’s time?
- Leakage. How many are missed, late or forgotten now? Missed calls and unanswered enquiries count.
- Running costs. Ask for model usage, platform fees and text or call charges at your volume, each as a separate line.
- Upkeep. Ask what it costs a month and what it covers.
- Build. Divide the build cost by the months you expect to use it, and add it to the monthly costs.
Then compare. If you miss 20 enquiries a month and one in five would have become a customer worth $2,000, that is four customers, or $8,000 a month, at stake. Set that against the monthly total from steps 5 to 7. If the system saves time instead, multiply hours saved by what that time is worth. Our guide to measuring AI return takes this further.
What changes the price
- Off-the-shelf versus custom. A standard product set up for you costs less to build but you live with its limits. Custom work costs more and fits your process.
- How much it touches. Each extra system connected adds build and upkeep.
- Volume. Usage-based charges grow with you. Flat fees don’t, until you hit a fair use limit.
- Model choice. Larger models cost more per token. Many jobs run well on smaller ones.
- Who maintains it. Doing it yourself costs time. Paying someone costs money but frees you.
Questions to ask before you accept a quote
- Which parts are one-off and which are monthly?
- What volume did you assume for the running costs, and what happens if we double it?
- Which AI model does it use, and can it be changed later?
- Whose name are the accounts in?
- What does upkeep cover, and what counts as extra?
For more on judging who you are dealing with, see our guide on choosing an AI automation agency. If you want the numbers worked through for your business, that is part of how our AI work starts.