Automate the boring path first
Frequent, repetitive, measurable work is usually a better first target than glamorous “AI transformation”. Email classification, document extraction, draft preparation and information routing can create useful gains without handing over consequential decisions.
Glasgow accountancy firm: emails requesting invoices, statements and receipts arrive dozens of times a day. A rule routes messages by subject and sender; an AI step reads the body to confirm what is being asked and which client it concerns; a draft reply is prepared for a person to approve. No invoice is ever sent without a human.
Rules route · AI interprets · human approvesUse rules where rules are enough
If a postcode is either in a service area or not, use a rule. If an invoice total must equal the sum of its line items, validate it deterministically. Do not ask a language model to guess predictable facts — it will occasionally get them wrong.
Use AI for messy interpretation
AI earns its place when people currently read free text, identify intent, extract variable fields, summarise context or draft language. Keep the resulting action controlled by normal software so AI only produces understanding, not unauthorised decisions.
Monitor the workflow, not just the model
Track throughput, corrections, escalations, failure reasons, cost and time saved. The automation is successful only if the whole process improves — not because an AI step was added.
Automating a broken process with AI just produces more broken output, faster. Fix the process and measure it first; then add automation and compare.
Plain automation
- Rules, scripts and integrations that always run the same way
- Perfect for predictable checks, such as a postcode in or out of a service area
- Reliable, fast and cheap to run
Use it wherever the answer is deterministic.
AI where interpretation adds value
- Reads free text, identifies intent and extracts variable fields, as in the accountancy example
- Keeps the resulting action controlled by normal software and human approval
- Occasionally uncertain, so it needs verification or escalation paths
Use it for the messy reading step, not for the decision.
A dependable workflow nearly always combines both: a rule routes, AI interprets, a person approves. The failure modes below show what happens when the two are confused.
A checklist for a safe first automation
- Choose a frequent, repetitive, measurable task.
- Map it: trigger, input, output, owner.
- Put predictable logic in rules, interpretation in AI.
- Keep consequential actions behind human approval.
- Measure before and after.
Common failure modes
For Scottish SMEs, automation that removes re-keying, chasing and manual document handling is usually the fastest measurable win. Start in one department and prove the number before scaling.
Frequently asked questions
Do I need code to automate AI workflows?
Not necessarily. No-code and low-code platforms can connect inboxes, documents, spreadsheets and CRMs, with rules and AI steps configured visually. Code helps for larger or higher-risk systems.
Can AI reliably read my invoices?
Modern models are good at structured extraction, but always validate totals and flag anything uncertain for a person. Never post an unverified AI-extracted amount.
How is automation different from an AI agent?
Automation follows a fixed workflow; an agent can choose actions toward a goal. Automation is simpler and safer. Start with automation before agent autonomy.
The tooling behind this guide is being built in public. It is open source on github.com/BryanFiFife, so you can inspect rather than take our word for it:
Open source does not mean certified. Read the code, run the tests, and treat it like any tool: check it against your own requirements before you depend on it.
A reliable implementation is not “an LLM with a prompt”. It is a controlled business system with a model only where a model adds value.