Answers with their sources
Questions companies ask about bringing AI in.
Each page answers one question in its first three sentences, then shows the evidence. Where a number appears, its source is on the page.
Answers
- 1.1Can AI actually modernize a legacy codebase, or does it just make a mess?It can, if the tests come first. AI is very good at the three things that made modernization unaffordable: reading code nobody understands, writing tests for behaviour that was never specified, and drafting the migration of one module at a time. It is bad at knowing which behaviour matters, which is why a migration run by AI alone produces something that compiles and is wrong. Sequence it and it works.
- 1.2Does adopting AI mean fewer people, or different work for the same people?The evidence so far points to different work more than fewer people. Companies most exposed to AI grew headcount 52% between 2018 and 2025 against 36% for the least exposed, and the one large field study of an AI assistant found attrition 8.6% lower among the workers who had it. Both cuts and hiring do happen, and they happen more in AI-adopting organizations than elsewhere. The organizations that keep their people are the ones that decide what the new work is before the tools arrive.
- 1.3How do you measure whether AI is actually helping your engineering team?Measure the system, not the individual. Baseline the four DORA metrics (deployment frequency, lead time for changes, change failure rate, time to restore) before the tools arrive, then add two AI-specific ones: the share of AI-authored changes that get reverted, and reviewer time per change. If lead time falls while change failure rate holds, it is helping. If throughput rises and reverts rise with it, you have bought speed with quality.
- 1.4What does an AI acceptable-use policy actually need to say?Eight things: which tools are approved, what data may never go into them, who approves AI output before it reaches a customer or a production system, how AI use is disclosed, who owns the policy, how incidents are reported, what the training expectation is, and how the policy changes. One page for staff, an appendix for engineers. Anything longer will not be read.
- 1.5What does AI consulting cost, and what are you actually paying for?In 2026 a mid-market AI readiness assessment runs about $10,000 to $25,000 over three to six weeks. Implementation engagements start around $40,000 and run into the low six figures depending on how many teams and systems are in scope. Fractional AI leadership is typically $5,000 to $10,000 a month. What you are paying for is judgment about what not to do, which is the part that saves money.
- 1.6Which internal processes should you automate with AI agents first, and which never?Start with work that is high-volume, repetitive, and cheap to undo when the agent gets it wrong: ticket triage and summaries, first drafts of documentation, data entry between systems, routine reporting. Never automate anything that changes a production system, moves money, or touches a customer without a named person approving the action. The order matters more than the tool.
If one of these is your situation
The three-week assessment works out which of them applies to you, in writing, before anyone builds anything.