Operational processes wired together so the repetitive parts run themselves, with a person kept in the loop exactly where the cost of being wrong requires one.
In one paragraph
When is a process ready to automate?
When four things are true: it is repetitive, it is already documented or can be written down in an afternoon, the data it runs on is reliable, and the cost of a mistake is known. Automating a process that fails any of those makes it faster and more confusing rather than better, which is the most common way these projects disappoint. The second rule is where the person goes: anything with a financial, legal or customer-facing consequence keeps a review step, and the automation is judged on how much work it removes before that step rather than on removing the step itself. On the record: 90% of B2B order and quote processing automated on a Magento platform.
What this is
Every growing business accumulates work that exists only because two systems do not talk: someone copying orders into a spreadsheet, someone re-keying a supplier feed, someone producing the same report every Monday. It is invisible on the org chart and expensive on the payroll.
The honest first step is usually not automation. It is writing the process down, because a surprising share of these turn out to be undefined, performed differently by each person doing them, and an undefined process does not become better when it is automated. It becomes confusing faster.
Where AI belongs in this is narrower than the current enthusiasm suggests, and more useful than the scepticism allows. A model is good at the judgement-shaped middle of a process: classifying a message, drafting a description, extracting fields from a document a human wrote. It is not a replacement for the deterministic parts, and wiring it in where a rule would have done is how an automation becomes unpredictable.
What this page is not
This page is about operational processes. If the question is whether a model should be in the product itself, retrieval, agents and evaluation, that is the generative AI page.
What usually brings this in
Someone exports a spreadsheet every morning.
A full-time person doing integration work by hand, and a daily window where the data is wrong.
Orders are re-keyed into a second system.
Transcription errors reaching customers, and a process that cannot scale without another hire.
The same report is rebuilt every week.
Hours spent on assembly rather than on the decision the report exists to support.
We automated something and now nobody trusts it.
It was built without logging or a review step, so a single bad outcome removed confidence in all of it.
What gets built
- Process mapping before tooling
- What actually happens, including the exceptions people handle without mentioning them. The map frequently shortens the process on its own, before anything is automated.
- Order and fulfilment operations
- Routing, status updates, supplier notifications, returns and exception handling, wired between the store and the systems behind it so the normal case runs without a person.
- Catalogue and content operations
- Enrichment, attribute completion, categorisation and translation drafted automatically and reviewed by a person, which is the arrangement that makes catalogue work tractable at scale.
- Reporting and alerting
- Scheduled reports assembled from the systems that hold the data, and alerts on the conditions that need a decision, so nobody is manually checking whether something happened.
- AI in the judgement-shaped steps
- Classification, extraction and drafting where a rule cannot express the task, with the deterministic steps left deterministic and the model constrained to the part it is actually good at.
- A review step where it is warranted
- Financial, legal and customer-facing outputs keep a person in the loop. The measure is how much work is removed before that step, not whether the step was eliminated.
Technologies
- Automation
- n8n
- Webhooks
- Scheduled jobs
- Queues
- AI
- LLMs
- RAG
- Prompt engineering
- Evaluation sets
- Integration
- REST APIs
- GraphQL
- ERP
- CRM
- Spreadsheets
- Commerce
- Magento 2
- Shopify
- WooCommerce
- Order management
How delivery runs
Write the process down
Including the exceptions, which is where the real work is hiding. A process that cannot be written down in an afternoon is not ready to be automated, and that finding is itself worth having.
Cost the manual version
Hours per week, error rate, and what a mistake costs. Without this the automation cannot be judged afterwards, and most of them never are.
Decide where the person stays
Draw the review step before building, based on the consequence of being wrong rather than on how confident the technology feels.
Build the smallest version
The highest-volume path first, running alongside the manual process. Automating every branch at once is how these projects stall before anything ships.
Measure, then widen
Against the number from step two. If it has not moved, that is information, and the next branch waits until it has.
Relevant builds
- Custom platform · B2B
Procurement and approvals platform
Scalable B2B procurement platform streamlining purchase requests and approvals.
- B2B workflows
- Approvals
- Integrations
- Magento 2 · Marketplace
Multi-category marketplace
A scalable Magento 2 platform powering a wide multi-category retail catalogue.
- Magento 2
- Multi-category
- Scale
Questions
- Do we need AI to automate a process?
- Usually not, and reaching for it first is the common mistake. Most operational waste is two systems that do not talk, and the fix is an integration with rules, which is cheaper, faster and entirely predictable. A model earns its place in the judgement-shaped step: classifying, extracting or drafting where a rule cannot express the task.
- What should not be automated?
- Anything undefined, anything running on unreliable data, and anything where a mistake is expensive and undetectable. The first two get worse when automated. The third can be automated up to the decision and then handed to a person, which is nearly always the right shape.
- Why n8n rather than a bigger platform?
- Because it self-hosts, the workflows are readable by the people who own the process, and there is no per-task pricing that punishes success. Where a business already runs something else, I will build in that instead rather than add a tool to the stack for its own sake.
- Who maintains the automation afterwards?
- Your team, by design. The workflows are built to be read and changed by the operations people who own the process, documented accordingly, and I am available on a retainer where that is wanted rather than assumed.
Describe what is in front of you and you get a written scope and a figure, usually within a few days of a short call. No charge for either.