Choosing the first workflow to automate
Where to start with agentic automation: pick work that is bounded, observable, and forgiving of mistakes.

Choosing the first workflow to automate
The first agentic workflow you ship matters more than the second. It sets expectations — for the business, and for the team building it. Pick the wrong one and you spend months explaining a failure; pick the right one and you earn the right to do more.
Bounded
Start with work that has clear edges: a defined input, a defined output, and a defined owner. "Summarise the support tickets from yesterday and route them" is bounded. "Make the company more efficient" is not.
Observable
Pick a workflow where you can already see what good looks like. If you cannot tell whether a human did the task well, you will not be able to tell whether an agent did. Existing review steps, logs, or acceptance criteria are a good sign.
Forgiving
Choose work where a mistake is recoverable. Drafting, summarising, routing, and triaging are forgiving — a human reviews before anything irreversible happens. Work that moves money, deletes data, or changes production is not where to learn.
A useful first step
A good first workflow is bounded, observable, and forgiving. It produces something a human already checks. It lets you build the Ground, Orchestrate, and Prove disciplines around a small, real problem — and then expand from there.
Related articles

Sparks of ASI: The Human Strategy to Align ASI in 2027
If AI can run a company, humans still need to define success. Alignment benchmarks, Jev-style policing and hard access controls could make that practical.
What is agentic software? How AI agents work and where they’re useful
A clear introduction to agentic software: what it is, how it differs from chatbots and fixed automation, where it can help, and what it takes to use it well.

Agentic Software plans accelerator application focused on AI training security
We’re planning an application to the 2027 CrowdStrike, AWS and NVIDIA accelerator, focused on data-poisoning risks in the AI training supply chain.
