iThe 30% is real. The reason it's real is not.
If you read the trade press, AI is a one-stop discount on IT cost. Replace your helpdesk with a chatbot, save 30%. Replace your developers with copilots, save 30%. Replace your operations team with an LLM agent, save 30%. The number is, in our measured experience, often accurate. The mechanism is almost never the one being advertised.
The 30% comes from removing the parts of the work that should not have been built. The AI is incidental.
iiWhere the savings actually live
1. Removing custom software that does what an LLM can now do
The most reliable savings are at the document-AI layer. Many SMEs spent six figures over the last decade on bespoke OCR pipelines, classification rules, and form-matching scripts. A modern LLM with a good RAG layer replaces 60-80% of that codebase. The savings are real because the maintenance cost was real. The AI itself is the smallest line item.
2. Killing internal "consolidation" projects
Many HK SMEs run with a familiar archaeology: a 2014 PHP system, a 2019 Node service, a 2022 spreadsheet. Three teams have proposed consolidating them, each at six figures. With a small LLM agent reading and writing across all three, consolidation becomes unnecessary. The legacy systems quietly continue; the agent provides the unified view. Cost: about 10% of a consolidation project.
3. Smaller, sharper teams
A senior engineer with good AI tooling produces something between 1.4× and 2.2× the output of the same engineer without it. (Anecdotal but consistent — we measure ourselves.) This is not "fire half the team". It is "do not hire the next two". On a six-person team this is 30% over twelve months.
iiiWhat does not save 30%
Three things people sell as cost saving, that almost never are:
- Marketing copilots. Real ROI exists but is small and slow. Don't lead with these.
- "AI customer support" bots retrofitted onto existing call centres. They displace the cheap interactions and concentrate the expensive ones. Your CS cost per ticket goes up, not down.
- Replacing senior judgement with an LLM. Whatever the demo shows, in production this fails in ways that cost more than the senior's salary.
ivThe order of operations that actually works
- Audit, don't pilot. Spend two weeks reading the source code, the spreadsheets, the support tickets. Find the three places where AI is replacing something rather than adding something.
- Build the smallest copilot. Internal-only, single-use-case, six-week build. Use a major-provider LLM behind your own private endpoint.
- Measure for 90 days. Time saved, tickets resolved, errors caught. Write it down weekly.
- Expand or kill. If you cannot show the 30% in 90 days, kill it and try the next case. The discipline of killing fast is the savings strategy, not the AI itself.
AI doesn't save you 30%. Asking "should this still exist?" — and being willing to act on the answer — saves you 30%. AI just makes the question askable.