Everyone has bought something. Fewer have chosen well.
82% of small business employers have already invested in AI tools, according to a 2026 survey by the SBE Council. In customer service the shift is even sharper: 80% of organizations now use AI agents, up from 47% in 2023. Adoption is no longer the interesting question. The interesting question is which problems deserve an off-the-shelf tool and which deserve automation built around your own data and systems.
Get that split wrong in one direction and you pay engineers to rebuild what a subscription does for cents. Get it wrong in the other and you wire your customer relationships into a rented black box.
What "buy" costs now
The buy side has become remarkably cheap at the entry point. Meta prices its Business Agent at roughly 2.00 USD per million tokens, which works out to about 4-5 cents for a typical US customer conversation, as CNBC reported. On the commerce side, Shopify switched Agentic Storefronts on by default for eligible merchants, making 5.6 million stores discoverable in ChatGPT, Copilot, Google AI Mode and Gemini without those merchants building anything.
At those prices, "build" cannot compete on cost for generic work. It should not try to. Building earns its keep somewhere else entirely.
Five questions that decide it
Run each candidate process through these before signing or specifying anything:
- Volume and exception rate. High volume with few exceptions favors buying. High exception rates mean the tool's happy path will not cover you, and every exception lands back on a human.
- Where does the data live? If answering well requires your inventory, your CRM, your pricing rules, a generic agent without deep access will be confidently wrong. Owning the integration means owning the answer quality.
- Does the process differentiate you? Password resets do not win customers. If the process is why customers choose you, treat its automation as a product, not a purchase.
- Integration depth. Counting how many systems the workflow must touch is a fast proxy: one or two, a connector-rich tool likely covers it; five systems with conditional logic between them, and you are describing custom work whether you admit it or not.
- Exit cost. Ask what leaving looks like before you enter. Conversation history, trained behaviors, and customer contact points held inside someone else's platform are the modern lock-in.
Sequence matters more than the verdict
Whatever you decide, order of operations decides most of the outcome. The rule we apply in our own automation work at Clodron: automate the boring, high-volume path first, and leave the exceptions with humans. Teams that start with the hardest edge case burn their budget proving that AI cannot do the one thing it is worst at, and never collect the easy savings sitting in the repetitive 80 percent.
What to do next
List your five most repetitive processes and score each against the five questions above. Most small businesses land on a hybrid: buy for commodity conversations at commodity prices, build thin custom automation where their own data is the advantage, and keep a clean export path everywhere. Then automate the most boring item on the list first, measure it for a month, and let that number, not a vendor demo, decide what you do second.