Section 1
Five categories, not fifty logos
Model providers sell raw capability through an API, priced by usage. They do nothing on their own. You are buying an engine, not a car. Workflow and orchestration tools connect systems and decide what runs when. This is where most small-company automation actually lives. Embedded features ship inside software you already pay for: your CRM, help desk, accounting package, document editor. Vertical point solutions do one job for one industry, with the rules of that industry already built in. Agent frameworks let a model plan multi-step work with access to tools. This is the newest category and the least settled. Start with the fundamentals in [What Is AI Automation? A Plain-English Guide for Founders](/blog/what-is-ai-automation-a-plain-english-guide-for-founders).
Section 2
Start from your system of record
The system where your work already lives should anchor the decision, not the tool with the best demo. For most service businesses that is the CRM, the help desk or the accounting package. Two reasons. Adoption collapses when people have to leave their normal environment to use a new thing, no matter how good it is. And integration is where budgets quietly disappear, so the tool that already sits inside your system of record starts with a large head start it did not earn on features. The corollary is uncomfortable for buyers who like shopping: the best first move is often turning on something you are already paying for. It is cheaper, it is faster, and it tells you whether the workflow was ever the problem. Category detail is in [Choosing the Right AI Automation Tools for Your Business](/blog/choosing-the-right-ai-automation-tools-for-your-business).
Section 3
Read the pricing model, not the price
Three pricing shapes dominate, and they behave very differently as you grow. Per-seat pricing is predictable and punishes you for giving access widely, which is exactly what you want to do once something works. Per-run or per-credit pricing looks trivial at pilot volume and becomes a real line item at production volume. Model it at ten times current usage before signing anything. Usage passed through from a model provider means your bill moves when the underlying prices or your input sizes move. Long documents cost more than short ones, and nobody mentions that in the demo. Also ask what happens to your data and your configured workflows if you leave. Migration cost is the real price of a platform.
Section 4
Buy, wire together, or wait
Three honest options and a rough rule for each. Buy a vertical solution when your process is standard for your industry and the vendor has the domain rules already encoded. You are paying to skip the specification work. Wire it together yourself, using an orchestration tool plus a model API, when your process is the differentiator and no product matches it. Budget for maintenance, because you now own it. Wait when the feature is visibly on your existing vendor's roadmap and your volume is low. Waiting six months for something bundled into software you already pay for beats a migration you will regret.
Section 5
What no platform does for you
Trustworthiness, design, evaluation and use are how NIST frames AI risk, and none of that transfers to a vendor when you sign a contract. You still decide which data the tool may touch, which records it may alter, which calls stay with a person, and who is accountable when something reaches a customer wrong. You still own disclosure. You still own the review step. Ask three questions in every procurement conversation: where is our data processed and retained, is it used to train anything, and can we export a full log of what the system did. Vendors with good answers give them immediately. The ones that route you to a sales engineer are telling you something. On explaining the change to your team, see [Storytelling in the Age of AI and Automation](/blog/storytelling-in-the-age-of-ai-and-automation).
Section 6
Running an evaluation that means something
Trials fail as buying tools because everyone tests the happy path. Build a set of twenty real cases from your own history, including the five that were genuinely hard, and score every candidate on the same set. Score four things: accuracy on your cases, effort to integrate with your system of record, cost at projected volume, and whether the team wants to use it after a week. That last one predicts more about the outcome than the first. Give each candidate the same fortnight, the same cases and the same reviewer. Two structured trials beat a market survey of thirty products you will never install.