Everybody talks about AI. Few companies actually use it for something that saves money. Over the past 12 months we've shipped 14 AI integrations across different businesses — here are four use cases that paid back inside 90 days.
1. Support ticket classification and routing
Problem: Customer service spends 35% of its time just sorting tickets before answering.
Solution: An LLM classifier (Claude Haiku or a local model) that reads the ticket, assigns category, priority and team, and — if the answer is simple — drafts a response.
Result on one client: 78% of tickets correctly classified, 40% reduction in first-response time.
2. Extraction from incoming invoices and contracts
Problem: Accounting types in 200+ invoices a month manually, with errors.
Solution: OCR + LLM that pulls supplier, invoice number, line items, VAT and due date. Output goes straight into Pantheon or a similar ERP.
Result: 92% of invoices processed automatically, manual review only on exceptions.
3. Smart search across internal docs
Problem: Engineers lose an hour a day looking for "how did we solve this for client X".
Solution: RAG (Retrieval-Augmented Generation) over a Confluence/Notion knowledge base — an internal assistant that answers with a source link.
Result: ~7h saved per engineer per week.
4. Anomaly detection in operations
Problem: The warehouse occasionally ships the wrong quantity — only caught when the customer complains.
Solution: A pattern detector comparing today's outbound shipments to historical baselines, flagging outliers before they leave.
Result: 6× reduction in mis-shipments in the first 60 days.
What we don't do
- We don't slap "GPT chatbot on the site" unless the client knows exactly which problem it solves.
- We don't recommend fine-tuning on small datasets — RAG + a solid prompt covers 90% of cases.
- We don't put AI on a critical path where hallucination has a cost — there's always a guardrail and human-in-the-loop.
AI works when it's treated as an engineering project with a measurable outcome, not a demo. If you're considering AI in your operations, let's start from a clear problem, not the technology.