AI Automation Guides

Which guide answers your question

Each guide below owns one decision. They are written to be read on their own, but they build on each other in roughly this order.

The idea underneath all of these guides

Nearly every recurring business process is a mix of three kinds of step, and almost all the value in automating one comes from telling them apart correctly.

1. 01Steps that follow fixed rulesCode
   Reading a field, matching a record, applying a threshold, updating a system. These should become plain deterministic code: identical behaviour every run, near-zero cost, and easy to test. Most of a mature process ends up here, and it should.
2. 02Steps that need genuine judgmentAI judgment
   Reading an ambiguous message, deciding which of five categories a messy case belongs to, drafting a reply that fits the situation. These are worth a model. Every one that survives into the running process is a place the procedure could not be pinned down — useful, but a cost to keep small.
3. 03Steps where a person must stay accountableHuman approval
   Paying money, sending something irreversible to a customer, acting on a decision about an individual. These keep a human approval permanently, not because the software is untrustworthy but because somebody has to be answerable.

That third category is not a temporary safety measure. Where a decision produces legal or similarly significant effects for a person, the GDPR gives that person a right to human intervention, and the EU AI Act sets out human oversight requirements for high-risk systems. Designing the gate in from the start is considerably cheaper than retrofitting one.

What to automate first

Google's site reliability engineers use a useful test for which work deserves automating. They call it toil: work that is manual, repetitive, automatable, produces no enduring value, and grows in direct proportion to how well the business is doing. That last property is what makes it worth attacking — it never gets better on its own.

Applied to business operations, the best first candidate is usually a process that is frequent enough to matter, boring enough that nobody enjoys it, and stable enough that you can describe what happens when it goes wrong. That last condition is the one teams skip.

• Frequent. Weekly at minimum. A process that runs twice a year will not repay the effort of pinning it down.
• Describable. You can write out the trigger, inputs, steps, exceptions and finish condition. If you cannot, the procedure guide is the place to start, not the automation.
• Bounded in blast radius. Prefer a first process where a mistake is embarrassing rather than expensive, while you build confidence in the approvals.
• Owned by one person. Someone has to be able to say what correct looks like, and to notice when it stops being correct.

Or start from your function

The guides above are about the general problem. If you would rather start from the work in front of you, each function has its own recurring jobs, its own risks, and its own reason for keeping a person in the loop.

Why the written procedure matters more than the tool

The formal quality-management world reached this conclusion decades ago: ISO 9001 is built on the requirement that processes be documented, maintained and reviewed, because a process nobody has written down cannot be audited, improved or handed over.

Automation makes this sharper rather than softer. A written procedure is what you check the automation against when a result looks wrong, what you hand to a new colleague, and what you edit when the business rule changes. Our own product treats the written procedure as the source of truth and the code as its implementation, for exactly this reason.

If you want to start there, the procedure templates are free, editable and require no account.

In this section

• GuideAI Workflow AutomationFor the operator who owns a recurring process and has become its bottleneck: what this category really does, which of your processes belong in it, and which ones you should leave alone.
• GuideAI Agent GovernanceFor the operations, risk or security owner who has to sign off on an agent touching real systems: the controls to put in place, and the evidence to keep, one process at a time.
• GuideHuman-in-the-Loop AIIf you own a process you are about to hand to software, this guide shows you where the approvals go, which pattern each one should use, and which gates must never come off.
• GuideSOP AutomationIf you own a recurring process and already have it written down for people, this guide shows what to add so software can run it end to end, and which steps should stay with a person.
• GuideAI Agent ReliabilityYou have an automation that works most of the time. This guide names the ways it fails, and shows how to shrink the part of it that can vary until those failures stop.

Limitations and when not to use this

• These guides are about business operations — the recurring, procedural work a team does across a handful of tools. They are not about automating software delivery, data pipelines or infrastructure, where the tooling and the trade-offs are different.
• Nothing here is legal, financial or employment advice. Where a guide touches a regulated subject we cite the primary source, but your obligations depend on your jurisdiction, industry and circumstances.
• Automation is the wrong answer for a process that is still changing shape every month, or one that runs rarely enough that a checklist would do. Both cases are cheaper to solve with a written procedure alone.

Sources

1. Eliminating Toil — Google, The Site Reliability Workbook Accessed 24 July 2026
2. ISO 9001:2015 — Quality management systems, requirements — International Organization for Standardization Accessed 24 July 2026

Related reading

• SOP TemplatesEditable procedure templates for the processes these guides describe.
• ToolsCalculators and decision tools for working out whether a process is worth automating.
• Finance & AccountingThe same material applied to one function, with complete job playbooks.