Practical AI automation, and where it should stop
Most successful AI projects do something unglamorous: classify a document, summarise a thread, or extract fields from an email. The failures try to replace a decision that needs accountability.
Good first tasks
Sorting inbound messages, extracting structured fields from documents, drafting replies for human review, and summarising long records. Each has a human check, a measurable accuracy target and a clear fallback.
Keep deterministic code for the rules
Tax calculations, eligibility rules, access control and money movement should be deterministic. A model that is usually right is not a control; it is a risk with a confidence score.
Design the review loop first
If a human cannot review and correct the output in less time than doing the task, automation has not helped. Build the review interface before scaling the model.
Common questions
- Do we need our own model?
- Rarely. A hosted model with good prompting and retrieval handles most business tasks, and avoids the cost and skill required to train and operate one.
- How do we measure success?
- Accuracy against a labelled sample, time saved per task, and the share of outputs accepted without edit. Track all three before and after.
Next step
If this question is live in your business right now, describe the situation and you will get a written view on the options, including whether you should build at all.
Discuss your projectRelated notes
All insightsWhere AI should not be allowed to decide
The categories of decision that need human accountability, and how to design automation that assists without taking responsibility.
ReadAI and automationAutomating invoice and document workflows
A practical architecture for extracting, validating and posting document data, including the exception queue that makes it reliable.
ReadAI and automationAI customer support that does not invent answers
Retrieval, refusal behaviour and escalation rules that keep an AI support assistant accurate and honest.
Read