Kyle Wiggs / Writing / Technology & AI
Technology & AI
The category is being sold with a level of confidence the technology has not earned. Some of it is genuinely useful right now, and it is not the part being demonstrated.
Meeting notes and summaries. Drafting the follow-up email. Finding the document nobody can locate. Turning a recorded conversation into a record of what was agreed.
None of that demonstrates well. All of it saves hours every week and carries almost no downside if it is wrong, because a human reads it before it goes anywhere.
Assembling everything relevant before a client meeting — recent activity, the last conversation, what changed, what was promised. That is real work, it is tedious, and it is exactly the kind of retrieval and summarisation these systems do well.
Anything presented as producing a recommendation. The output looks authoritative regardless of whether it is correct, and confident wrongness is the specific failure mode of this technology.
In a regulated context, an output that looks like advice and has no traceable basis is a liability rather than a feature.
The question is not whether it is impressive. It is whether you can explain the answer to a regulator eighteen months from now.
Client communications are books and records. A tool that generates or transmits them sits inside your compliance obligation, and needs to be captured, supervised, and covered by policy.
Firms that adopted these tools informally, through individual staff, have generally had to retrofit that, which is more expensive than doing it deliberately.
Whether a vendor trains on your inputs, where data is stored, and who can access it are contract questions with real answers. Ask them before the tool is in use, not after.
It is a very good assistant for work a human will check, and a poor substitute for judgement nobody checks. Firms getting value from it today are almost all using it in the first way.
Read next
What AI-native should mean →