How much does AI automation cost a small business?
OpenAI's new Sol and Luna models make the cost question timely. But a cheaper model is only one part of a dependable business system. Here is what belongs in your budget.
The short answer: there is no useful single price for AI automation. A draft-writing assistant and a system that reads enquiries, checks availability and updates your CRM are different projects. Budget separately for implementation, monthly operation and the people who handle exceptions.
OpenAI's September 22 announcement introduces GPT-6 Sol and Luna with lower API rates than their predecessors. That makes it worth revisiting some running-cost assumptions, not assuming that every custom application has become cheap to build.
This is Andrew Matia's independent buying guide for LineWeb. The examples below are planning illustrations, not project quotes or promised savings.
Buy a process that works, not a model name that sounds impressive.
Three costs. One business decision.
Ask a supplier to separate these items so you can compare like with like.
Build and connect
Workflow design, integrations, data preparation, permissions and tests against your real business cases.
Run and maintain
Model usage, hosting, connected services, monitoring and agreed support. Identify fixed and variable charges.
Review and improve
Your team's review time, exception handling and changes when policies, products or integrations evolve.
What changed in OpenAI's pricing?
Lower model rates can reduce usage costs. They do not price the whole project.
The launch table lists GPT-6 Sol at $2 per million input tokens and $10 per million output tokens. GPT-6 Luna is listed at $0.10 and $0.50 respectively. These are API rates in US dollars, checked against the September 22 announcement, not a monthly subscription or a LineWeb service price.
Illustration only: 1,000 requests using 2,000 input tokens and 500 output tokens each would produce $9 of Sol usage or $0.45 of Luna usage at those rates. This assumes only the stated token volumes, without additional reasoning, tools, retries or other billable usage. It is deliberately not a forecast for an agent that takes many steps.
OpenAI: GPT-6 Sol and Luna announcement and launch pricingWhy two automation quotes can look very different
Imagine a service company receiving enquiries through its website and email. Preparing a suggested reply is a narrow task. Reading attachments, detecting duplicates, assigning a territory, checking the current calendar and creating a reliable CRM record requires more integration work. The same model can sit behind both.
Ask what the implementation fee actually includes. Does the system understand your approved answers? What happens when a connection fails? Who can correct a record? Can you export your data and move to another provider? A lower quote that leaves those questions unanswered may simply cover less work.
Use a written acceptance checklist. Agree the input, expected result, human approval points and examples that must be handled correctly before launch. For a first project, a smaller complete workflow is usually easier to evaluate than a broad assistant with unclear responsibilities.
- Separate one-off setup from monthly support and metered usage.
- Specify included integrations and who maintains them.
- Agree a usage ceiling, alerts and an owner for exceptions.
- Define data export, access rights and the process for future changes.
Estimate value from work you already understand
Start with a week's sample of the task. Count how often it happens and time the complete human process, including corrections. Then trial the assisted version on comparable cases. Count accepted results, not merely generated responses.
For example, saving three minutes on 400 enquiries releases 20 hours of capacity per month. At an illustrative internal cost of 20 EUR per hour, that represents 400 EUR of capacity value before system costs. It is not automatically a cash saving: someone must be able to use that time productively, and review work must be included.
A useful calculation is: monthly capacity value plus evidenced additional contribution, minus operating and review costs. Only estimate implementation payback when that net value is positive. Do not count a sale as an AI result simply because an automated reply happened somewhere in the journey.
- Record the baseline before introducing the automation.
- Include failed runs and human review in the cost per result.
- Compare models on your own accepted-output test set.
- Start with the smallest workflow that can demonstrate value.
Custom is valuable when the connections matter
An existing business subscription may be enough for drafting, summarising and individual research. Custom development becomes more relevant when the task must follow your operating rules across several systems, or when a generic tool creates extra copying and manual checks.
Our recommendation is to inspect your current CRM and software first. Use capabilities you already pay for where they fit. Build the missing connection only when the saved work or improved service justifies its ongoing ownership.
Connect AI to the work behind the business
LineWeb designs automation around existing forms, CRMs and operational tools, with a defined scope and a clear handover to the people using it.
Explore automation and integrationsClear answers, without the jargon.
01Is AI automation a one-off cost?+
Usually not. Separate implementation from model usage, infrastructure, third-party subscriptions and maintenance. Ask which costs rise with volume.
02Does a cheap AI model make custom development cheap?+
Not necessarily. Integration, workflow rules, testing and support still need work. Model usage is only one budget line.
03Can I keep my existing CRM?+
Often that is the sensible starting point. Feasibility depends on its interfaces, permissions and data quality, which should be checked before quoting an integration.
04How do I request an accurate AI automation quote?+
Provide a sample of the task, monthly volume, current tools, the desired result and the mistakes you cannot accept. Remove personal or confidential information from initial examples.
Andrew Matia - LineWeb
Founder of LineWeb. I write about the practical side of websites, systems, automation and search: the decisions that make a business easier to run and clearer to find.
Read the founder storyTell us which task is costing you time.
Share the workflow, the tools you use and an approximate monthly volume. We can assess whether existing software is enough or a focused custom integration makes sense.