Rules, AI or an agent? Choosing the right kind of automation
Most workflows need less AI than the pitch suggests, and more control than the demo shows. A practical way to choose per step.
- Author
- Orivion editorial
- Published
- Last reviewed
- Reading time
- 7 min
Written from Orivion's delivery practice.

Why this matters
AI projects fail in two predictable ways: they use a model where a rule would have been cheaper and more reliable, or they give a model responsibility without the controls that responsibility needs.
Choosing the right mechanism for each step is the single decision that most affects cost, reliability and risk.
A decision checklist before you automate
Use rules when the inputs and decisions are predictable. Consider AI when the task involves language, documents or classification that fixed rules cannot handle well. Consider an agent only when the task needs several steps or tools, with bounded permissions and a way for a person to intervene.
Assess one workflow using its typical volume, input quality, exceptions, cost of an error and the systems it must access. Define representative test cases and an acceptance threshold before implementation; compare time saved against review effort and operating cost. An impressive demonstration is not evidence that the workflow is ready for production.
Rules: when the steps are known
If you can write the logic down as a flowchart, use ordinary automation. Moving data between systems, routing an approval, sending a reminder or updating a CRM field are rule-based tasks. They are cheap to run, deterministic and easy to audit.
Workflow tools such as n8n, Make, Zapier or Power Automate, or a small custom service, handle most of this work.
AI: when the input is language or documents
Models are useful where rules break down: reading an email to understand intent, extracting fields from a scanned document, classifying a request, searching knowledge by meaning, or drafting a response.
The output is probabilistic, so it needs a measured accuracy level and a plan for when it is wrong. Retrieval-augmented generation (RAG) keeps answers grounded in your own approved documents and lets them cite sources.
Agents: when a task needs several steps and tools
An agent plans and executes a sequence of actions, calling tools and APIs along the way. That makes it useful for bounded, multi-step tasks, and it also means it can act, not just answer.
Agents therefore need tighter permissions, logging of every action, and human approval for anything consequential. Start with assistance and review; widen autonomy as evidence builds.
People: where accountability sits
The NIST AI Risk Management Framework and its Generative AI Profile treat risk as something governed, mapped, measured and managed across the lifecycle. In practice that means a person is accountable for each AI-supported process, and people approve decisions with legal, financial or customer consequences.
Examples
- Invoice intake: rules route and validate; AI extracts fields; a person reviews low-confidence cases.
- Customer support: AI answers from approved content with citations; unresolved conversations hand over to a person with context.
- Lead handling: rules route by region and service; AI summarises the enquiry for the salesperson.
Before you automate a workflow
- Can the steps be written as rules? If so, start there.
- Is the input language, documents or unstructured data? Consider AI for that step only.
- What happens when the output is wrong, and who notices?
- Is there a representative test set to measure accuracy before release?
- Which actions need human approval?
- Where does the data go, and under what provider terms?
- Who owns the process after launch?
What to do next
Pick one high-volume workflow, map it step by step, and mark which steps are rules, which need AI and where a person must decide.
Information on this site is general. It is not legal, tax or financial advice. Requirements change, so confirm current rules with the relevant authority or a licensed adviser.
Related articles
Keep reading.
- Data5 min readBefore the dashboard: can the business trust its numbers?Dashboards fail when definitions and data quality are left for later. Start underneath.Read
- Digital Strategy4 min readChoose the CRM before the first customerDesigning customer data early is faster than untangling it later.Read
- Commerce6 min readBuilding a travel or tour company website with online bookingTours are sold by date, group size and pickup point, which makes them harder to book online than a product. How to plan the site so customers can actually pay for one.Read
Want to talk through how this applies to your business?