The first question on almost every strategy call is some version of "what does this cost?" The second is "why won't anyone just tell me?" Both are fair. This guide is our attempt at a straight answer: not a number, because a number without a scope is a guess, but a complete map of what moves the number, so you can walk into any conversation with any AI automation agency already knowing where the money goes.
One reassuring fact up front. Cost is not the main thing stopping small businesses from adopting AI. Intuit QuickBooks' 2026 AI Impact Report found the top barriers across the US, Canada, UK and Australia were data privacy and security concerns, fear of errors, and limited knowledge of what AI can do. Cost was not among them. And of US small businesses that paid for dedicated AI tools in 2024, 86% were still paying for them a year later, according to the same report's analysis of anonymised payment records. Businesses that buy well tend to keep what they bought.
The five things that actually drive the price
1. Integrations, not intelligence
The model is the cheap part. A voice agent's language model costs cents per conversation. What costs engineering hours is making it read and write your CRM correctly, book against your real calendar with your real rules, hand off to the right human in the right team, and log everything where your reports already look. Each system you connect adds scope. A build touching one CRM is a different project from a build touching a CRM, a calendar, a payment processor and a ticketing tool.
2. The state of your data
An AI chatbot trained on your business needs your business written down. If your policies, pricing and processes live in a tidy knowledge base, the retrieval layer is quick to build. If they live in the head of your longest-serving employee, the first two weeks of the project are interviewing that person. A document-processing system needs example documents, ideally hundreds, and ideally with the correct answers already known. The less of this you have, the more of the budget goes on producing it.
3. Telephony and messaging compliance
Anything that calls or texts customers in the US has to be registered with carriers under A2P 10DLC rules, with documented opt-in and working opt-out. Skip this and messages do not bounce; they are silently filtered. A proper build includes registration, quiet hours, consent handling and deliverability monitoring. It is unglamorous, and it is where most do-it-yourself setups we inherit were quietly failing. Budget for it.
4. Evaluation and guardrails
This is the line item resellers skip and engineers insist on. Before an AI system goes live it needs a test set of known-good examples it must pass, refusal rules for what it must never do, and a pause button that stops outbound actions in seconds. S&P Global Market Intelligence found the share of companies abandoning most of their AI initiatives jumped to 42% in 2025 from 17% a year earlier, and that the average organisation scrapped 46% of proof-of-concepts before production. Evaluation is the difference between the projects that ship and the ones that join that statistic.
5. Who is on the keyboard
Senior engineers cost more per hour and far less per working system. A junior team will build the happy path in a week and spend two months discovering the edge cases the senior team would have designed around on day one. When comparing quotes, ask who actually does the work and whether they are the people on the sales call.
The AI is cents per conversation. The plumbing, the data, the compliance and the testing are the project.
The three engagement shapes
Almost every honest AI automation engagement takes one of three shapes. Knowing which one you need is the fastest way to size a budget.
What it is: One to two weeks. Architecture, integration map, risks, data audit and success metrics, written down.
When you need it: You know something should be automated but not exactly what, or the workflow spans several systems and nobody has mapped it.
What you keep: The document, whether or not you build with the agency that wrote it. That is the test of a good discovery: it should be worth paying for on its own.
What it is: A production system shipped against a written scope, typically in four to twelve weeks, with evaluation, guardrails, monitoring and a handover included.
When you need it: You already know the system you want, or discovery has defined it. A fixed fee is possible precisely because the scope is written.
What moves the price: Everything in the five drivers above, in roughly that order.
What it is: A monthly retainer for monitoring, evaluation, iteration and second-generation builds as the business changes.
When you need it: The system touches revenue every day and you want someone accountable for it who is not you. Models change, carriers change rules, your team changes scripts. Somebody has to keep up.
What it is not: A licence fee. You still own the system. The retainer buys attention, not access.
At BaigOps we add a fourth step before any of these for founding clients: a working prototype on your own data, free. The commercial conversation happens after you have seen the system work, not before. The engagement shapes and what each includes are on the pricing section of our homepage.
The running costs after launch
A production AI system has a monthly bill even with no agency involved. Model usage, usually cents per conversation or per document. Telephony, per minute and per message at wholesale rates when the numbers are in your own Twilio account. Hosting, which for most small-business systems is modest. And the human review time your team spends on the exceptions the system escalates. Ask for an estimate of all four before you sign, and ask which accounts they are billed to. The right answer is yours.
Where budgets get wasted
- Paying to replace tools that work. A CRM migration bolted onto an automation project doubles the timeline and halves the chance of shipping.
- Buying a platform before defining the workflow. The subscription starts on day one. The value starts when the workflow is built, which is often never.
- Automating the wrong end first. The missed call is worth more than the tidy report. Invoca's benchmarks found 37% of phone leads convert during the call itself. Start where the money leaks.
- Skipping evaluation to save two weeks. Gartner has warned that in 2026 a third of companies will harm customer experiences by deploying AI prematurely. The two weeks come back with interest.
- Building in-house because it looks cheaper. MIT's GenAI Divide research found tools bought from specialised vendors succeeded about 67% of the time, while internally built tools succeeded about one third as often.
How to size your own budget
Count the systems the automation must touch. Count the documents or conversations you can hand over as examples. Note whether it calls or texts customers. Decide whether you know the system you want or need it defined. Those four answers place you in a shape and a rough scale, and any competent AI automation agency should be able to quote in writing within a week of hearing them. If they cannot, that tells you something too.
If you would rather have the number than the framework, book a strategy call. We will scope the bottleneck, and founding clients see a working prototype before the first invoice exists. The full picture of what we build is on the AI automation agency page.