Section 1
The six jobs in the stack
Orchestration. The thing that runs when something happens: a form submission, a new row, a scheduled hour. Zapier, Make, and n8n occupy this layer at different levels of technical depth. Model access. Where the reasoning happens, either through a provider's API directly or through a layer that lets you switch between them. Knowledge and retrieval. Getting your own documents, policies, and records in front of the model so it quotes your business rather than its training data. Actions and connectors. Writing back into the CRM, the helpdesk, the accounting system. Usually the hardest part and the least discussed. Interface. Where a person meets the system: chat, inbox, spreadsheet, or an existing tool the team already lives in. Observability and evaluation. Logs, cost tracking, and a way to tell whether quality moved. Startups skip this one and pay for it later.
Section 2
The rule that removes most candidates
Fill the layer the work already touches. If your team runs on a helpdesk, adding intelligence inside that helpdesk beats a better standalone product that requires opening a second tab. That one rule cuts most shortlists faster than a feature comparison. Adoption does not fail because a tool was mediocre. It fails because using it required a detour, and under pressure people take the shortest path they know. The corollary is uncomfortable for anyone who enjoys tooling: your existing platform's adequate built-in AI feature will often outperform a superior specialist product, purely on usage. Prefer the boring option that lives where the work lives, and reserve the specialist purchase for the one workflow where the difference is material. The full selection method is in [Choosing the Right AI Automation Tools for Your Business](/blog/choosing-the-right-ai-automation-tools-for-your-business).
Section 3
What to actually compare inside a category
Feature lists converge. These do not. Pricing unit. Per seat, per task, per run, per step, or per token. At low volume this is trivia. At the volume you are planning for, it is the difference between a line item and a problem. Error behaviour. What happens on a failed step, whether retries are automatic, whether a partial run leaves half-written records behind. Data handling. Where processing happens, retention on prompts and outputs, and whether your inputs train their models. Exit. Can you export your workflows, prompts, and history in a form that is usable elsewhere, or is the configuration the lock-in. Depth ceiling. How far the tool goes before you need code. Every no-code tool has a ceiling, and hitting it two years in is expensive.
Section 4
Sequencing for a startup
Do not buy the stack. Buy the layer the current workflow needs and stop. Most startups should begin with orchestration plus model access, since that pair covers classification, drafting, extraction, and routing, which is the bulk of early value. Add retrieval when the system needs to know things specific to your business. Add observability the moment more than one automation is live, because the alternative is debugging blind. Agent frameworks come last, if at all. They are the layer most likely to be interesting and least likely to be necessary at your stage.
Section 5
The costs the pricing page does not show
NIST frames AI risk management around trustworthiness, design, evaluation, and use, and a tool decision is partly a risk decision. Know what each product can access, what it can change, and what it retains. The unpriced costs are consistent across categories. Integration effort, usually larger than the build. Maintenance when an upstream API changes. Review time, which is real staff cost and rarely modelled. Cost variance, since usage-based pricing follows your busiest month rather than your average one. And the switching cost you accepted when you put your logic into a proprietary configuration format. Ask a vendor where they make money. A tool priced per task wants more tasks. A tool priced per seat wants more seats. Neither is dishonest, but both shape the advice you get.
Section 6
How to judge whether the tool earned its place
Track cost per completed unit of work, not cost per licence. Track the share of runs that complete without human intervention. Track time from failure to detection. Track how many automations are live versus how many still run for a reason someone can state. Then cull quarterly. Any tool with no owner, no active workflow, or no measured effect gets cancelled. Stacks accumulate, and almost nothing removes an item unless a process forces the question. How you describe the resulting system to your team and market matters too: [Storytelling in the Age of AI and Automation](/blog/storytelling-in-the-age-of-ai-and-automation).