CEO @ ScalingSmiths | Ex safety engineer building AI tools for safety engineering firms
Glasgow
Joined March 2021
- Tweets970
- Following53
- Followers169
- Likes10.1K
Pinned Tweet
The most useful AI system I've built this year has no gen AI in it (and 2 key constraints)
The most expensive words in a tool build: "let's add AI"
Expensive 2x
Once in trust - output you can't guarantee between runs, where you needed it fixed.
Once in cash - hardware and hosting, if the data is commercially sensitive or client-confidential.
The habit that works:
Before anyone says those words, ask which parts of the job follow rules you could write down.
Code the parts you can. Then look at what's left.
Why I build the AI parts of a tool last
Internal AI tools keep disappointing for the same reason:
They start from the model, not the job
- AI shoved in before the problem is looked at
- demo the model before the job is scoped
- feed it everything
- hope the answers hold up
Here's the order I build in instead:
1. Start with the parts that follow clear rules - plain code, no AI.
I'd sort your data into four tiers before any AI touched it (3&4 for data giving competitive edge)
The four tiers:
3. Built on Microsoft - custom on Azure, or plug-and-play (and limited) Copilot.
Limited increase in risk - assuming your firm stores/processes data with Microsoft
4. Never leaves the building - local models only. Dearer, and limited to open weights models
AI is the wrong tool for 95% of work
What people expect an AI tool to be:
- AI used into every step
- a black box that spits out the golden answer
What the useful version usually is:
5 checks before AI output goes into a report:
- every figure or standard cited, traced to the source
- inputs are right & assumptions stated
- passes the smell test
- you can follow & defend the reasoning
- a named person has signed it off
YOU'RE still accountable
Build the search first and you can test whether AI actually adds anything, instead of guessing.
Would a serious model on top help? Absolutely. The search works either way.