From SOP to automation: the promotion rule
The promotion rule is simple: a process earns automation only after it has been written down as an SOP and executed manually, from the document, without deviation. If a competent human cannot run the process from the written steps alone, software cannot either. Automating an undocumented process does not remove the chaos — it makes the chaos run faster and without supervision.
What is the promotion rule?
Think of it as a ladder with four rungs. Knowledge starts in someone's head. It gets promoted to a written SOP. The SOP gets promoted to "a human runs this from the document, every time". And only the steps that survive that stage — the ones that run identically, execution after execution — get promoted to automation. Steps that keep needing judgement stay human, with the SOP as their guide.
Most automation failures I see in small B2B firms come from skipping rungs. Someone jumps from "Dave knows how we do this" straight to a workflow tool, and the builder has to guess at every decision Dave makes without noticing he is making it. This is the pattern behind most of the disappointments catalogued in what AI automation actually delivers for B2B firms: the technology gets blamed for a documentation failure.
Why do automations built without an SOP fail?
Because the edge cases live in someone's head, and a workflow tool has no head. A human running a process handles the odd input without registering it as an exception — the invoice with no PO number, the lead who fills the form in twice, the client who replies to the wrong thread. An automation built from a verbal description meets that input and either breaks visibly or, worse, does something silently wrong.
Writing the SOP first forces the exceptions into the open. Running it manually from the document is the cheapest possible test environment: every time the person doing the work has to depart from the written steps, you have found a branch the automation would have missed. The document absorbs the correction at the cost of a sentence, not a support incident.
How do you know a step is ready for promotion?
Here is the mechanism I apply. When a step has run its recent executions — twenty is a reasonable working threshold, though it is a rule of thumb, not a law — exactly as written, with the same inputs producing the same action and no one reaching for judgement, then it is a candidate for automation. When the written step still contains words like "usually", "depends" or "use judgement", then it stays with a human, because those words are unhandled branches wearing plain clothes.
Two practical corollaries. First, promote steps, not whole processes — most real processes end up hybrid, with software doing the repetitive middle and a person owning the ends. Second, the SOP does not retire when the automation ships; it becomes the specification you check the workflow against when it misbehaves.
What does the promotion path look like in practice?
Take examples from our own work. Mailbox warm-up is fully promotable: the schedule of gradually increasing sends and engagement checks is identical for every domain, which is exactly why the industry automated it years ago — I have written about what a 91/100 warm-up score actually means and the process behind it. No judgement, no exceptions, perfect automation candidate.
Reply handling is the opposite case. Sorting replies into interested, not-now and unsubscribe looks mechanical, and the triage step often can be promoted — but the response to an interested reply stays human in every system we build, because that step's SOP is full of "depends".
Between those poles sits most of a firm's operations: lead routing, CRM updates, report compilation, invoice chasing. Each one splits into promotable steps and judgement steps once you write it down — and not before.
Where does this leave AI?
Exactly where the ladder predicts. Language models widen what counts as promotable — a step like "summarise the enquiry and flag anything unusual" can now be automated where it could not be five years ago — but they do not exempt you from the documentation rung. A model executing your process needs the same written definition of "unusual" that a new hire would, a point I have made at length in AI doesn't know your business until you write it down.
There is a compounding bonus, too. The SOP library you build on the way up the ladder is also the corpus that powers an AI assistant over your own documents — one set of documents, two payoffs. Write things down once; promote what stabilises; keep judgement where it belongs.
Next step: the Growth System Audit — £450, seven days, credited against any build — maps where your growth system leaks and what to build first.
Total Format builds the systems UK B2B service firms grow on — AI-powered outbound, automation, and reporting — so growth stops depending on the founder's time.
Map your growth system. The £450 audit takes seven days and is credited against any build.
BOOK THE AUDIT