Section 1
Readiness is about the work, not the stack
The instinct is to audit tools. The tools are the least of it. Almost every company already has the software required for a first automation, sitting unused inside a subscription they renewed last month. What actually gates a project is whether the work has a shape. If two people do the same task differently and both are right, an automation will encode one of them and quietly annoy the other. If nobody can say what a correct output looks like, there is nothing to evaluate against. So run the assessment at workflow level. Pick the task you would most like to stop doing, and put it through the four tests below. Definitions first if you need them, in [What Is AI Automation? A Plain-English Guide for Founders](/blog/what-is-ai-automation-a-plain-english-guide-for-founders), then a roadmap in [Building an AI Automation Roadmap for Your Startup](/blog/building-an-ai-automation-roadmap-for-your-startup).
Section 2
Test one: can you write the work down
Ask the person who does the task to write the steps as instructions for a competent new hire. Not a flowchart. Prose, with the decisions named. Three things fall out immediately. Steps that exist only in one person's head. Decisions that turn out to be judgment calls rather than rules. And branches nobody documented because they only happen twice a month. If the document cannot be written, the workflow is not ready and no tool will rescue it. This is also the cheapest possible diagnostic: an hour of somebody's time, and it frequently reveals that the process itself is the problem and the automation is unnecessary.
Section 3
Test two: can the system reach the data
Trace every input the task consumes and ask where it lives. A shared drive, an email inbox, a CRM field, someone's memory, a supplier portal that requires a login and a captcha. Anything behind a manual export is a maintenance liability. Anything in an inbox is unstructured and unreliable. Anything in someone's memory is not data yet. You are looking for a workflow where the inputs sit in systems with an interface, and the fields the task depends on are populated most of the time. Fifteen trustworthy fields beat a warehouse of stale ones. If the inputs cannot be reached without a human fetching them, the automation will save less than you think.
Section 4
Test three: who owns a wrong answer
NIST's framing runs through trustworthiness, design, evaluation and use. The compact version for a startup is a named owner per system. Before building, write down what the automation can read, what it can change, what it must never decide alone, and whose name is attached when a customer receives something wrong. If the honest answer to the last one is the founder, that is fine. If it is nobody, you are not ready. Add two mechanics while you are there: a log of what the system did, and a rule that machine involvement is disclosed where a customer would care. On the human side of the transition, see [How to Find Your Startup's Origin Story](/blog/how-to-find-your-startup-s-origin-story).
Section 5
Test four: what error rate can you live with
Every automated workflow has an error rate. The question is whether you have decided what is acceptable before the first mistake, or after. Sort the work by consequence. Reversible and internal, such as drafting or filing, tolerates a high error rate because the cost of a fix is a keystroke. Irreversible or external, such as sending money, making an employment decision or messaging a customer, tolerates almost none and needs a human gate. Name the number. If you cannot say whether nine correct out of ten is good enough for this task, you have no basis for judging the pilot, and you will end up arguing about anecdotes.
Section 6
Scoring yourself honestly
You are ready to automate a workflow if you can write it down, reach its data, name an owner for wrong answers, and state the error rate you accept. Four out of four, on one workflow, is a green light. You are not ready if the process changes depending on who runs it, if the inputs require a person to fetch them, if the output is irreversible and there is no review step, or if the reason for the project is that someone asked about your AI strategy. Three out of four is common and usually fixable in a fortnight. The missing test tells you exactly what the next piece of work is, which is more than a maturity score will ever give you.