The Technology Leader's AI Handbook
Check the work before you scale it
Technology leaders are told to scale AI, but no system should go wider than the ability to check its work. This handbook gives the executive who answers for a large company's AI systems a method for each stage: measure what the tools actually do, prove the work before it scales, find the cost of running, control agents with permissions rather than prompts, and give the board evidence it can defend. Every chapter starts from a decision a technology leader must make and ends with what to check before building, shipping or signing.
Inside this book
- Measure what AI tools do for engineering speed before committing delivery dates or headcount
- Check the work with tests fixed beforehand and cases the system has not seen
- Find the cost of one output, because a short trial understates the cost of running
- Standardise the process and the data before automating them
- Fund a build one stage at a time with a stop condition agreed in advance
- Set who builds, who checks and who owns the value across central and embedded teams
- Run models inside an agreed error band, with halt rules and re-testing for as long as they run
- Control agents with permissions and approval at steps that cannot be undone, not with prompts
- Treat regulation as design work: classification, logging and human oversight
- Decide build, buy or partner by who owns the software and the data around the model
- Keep an evidence file for every public claim about what the AI does
By Valentin Nagacevschi