Automation, AI or custom software? Start with the work.
If the same steps happen every time, automate the workflow. If someone has to read or judge something, add AI. If the business needs an interface no existing product can provide, build software. The right answer is usually smaller than the first idea.
The short version
| Choose this | When | First useful version | Not when |
|---|---|---|---|
| Workflow automation | The same steps happen in a known order. | Capture an enquiry, assign an owner, update the CRM and send the follow-up. | The work needs judgement rather than rules. |
| AI integration | Someone needs to read, search, summarise, classify or draft. | Read an invoice, find an answer in company documents or triage a support request. | There is no clear source material or human review path. |
| Supervised AI agent | The task has several steps and needs a controlled loop. | Research a defined set of sources, monitor a portal or prepare a request for approval. | The goal is vague or the agent would make an irreversible decision alone. |
| Custom software | People need accounts, permissions, dashboards or a workflow no product fits. | Build an internal tool, customer portal, MVP or mobile experience around the real process. | A simple rule or existing product already solves the need. |
1. Start with the handoff
Look for the point where work changes hands. An enquiry moves from a form to an inbox. A quote becomes a job. A document becomes a decision. A report is assembled before a meeting. That handoff is usually a better starting point than “we need AI”.
Write down who does it, what they copy, which system they update and what happens when something is missing. You have just described the first version of the system.
2. Use rules before intelligence
If a workflow is predictable, use a rule. Rules are easier to test, explain and maintain. A lead can be assigned by location. An invoice can trigger a reminder. A job can notify the person who owns the next step.
That is not less advanced. It is often the safer and cheaper answer.
3. Add AI where the input is messy
AI earns its place when the work involves language, documents or judgement. It can extract fields from a PDF, find an answer in an internal knowledge base, classify an inbox or prepare a response for a person to approve.
The UK Government AI Adoption Research looks specifically at adoption barriers and business impact. Its practical message is useful: adoption needs a clear use case, skills and a way through the risks.
4. Build an agent only when the loop is clear
An agent is not a more impressive name for a workflow. It is useful when the task has several steps, the information can change, and a controlled loop is worth the extra complexity. Keep approval points, logs and limits in the design.
Tools such as n8n's workflow catalogue show the range of patterns businesses explore, from lead handling and document operations to support and AI. The business still needs to decide what the agent is allowed to do.
5. Build custom software when the interface is the problem
Sometimes the work is not missing automation. It is missing a place for people to see the right information and take the next action. That is when a dashboard, portal, internal tool, MVP or mobile experience makes sense.
Start with one role and one workflow. Give that person a useful first version. Expand after the process has earned it.
A GOOD FIRST CONVERSATION
Bring one example of work that gets copied, chased or forgotten. We will help you choose the right level of solution.
Talk through the problem