The assistant said my colleague was not in my team
17 per cent of live employees had no manager recorded, so an AI assistant would say a real colleague was not in your team. Why names stay out of the index.
Topic
AI
What AI genuinely accelerates in delivery, and where it does not.
14 articles
About this topic
These pieces are about using models in production rather than about models. The distinction matters because almost all of the difficulty sits outside the model: what the system does when retrieval returns nothing relevant, how you tell whether last week’s change made the output better or merely different, and what a confident wrong answer costs when a person acts on it.
The other half is AI inside our own delivery — where it genuinely earns its place, which is comprehension of code nobody documented and the mechanical bulk around a change, and where it does not, which is anywhere the answer has to be right rather than plausible. Both are written from having done it, including the times it produced something that read well and was wrong.
The assistant said nothing was waiting. Four things were
An AI assistant told a parent nothing new had been asked for. Four reward requests had sat unanswered for weeks. The fault was in how it was told to look.
The pricier image model labelled the wrong picture perfectly
We asked two AI image models for a pizza with three of eight slices left. The dearer one drew a perfect 3/8 label over a whole pizza. A parent would approve it.
The assistant invented four requests the child never made
A parent asked an assistant what their child had requested. It listed four things. The child had asked for one reward, four times. The prompt already forbade it.
The model should never produce the number
An owner asked what it would cost to let their sales team question their own data. We measured it, our cheap-model theory lost, and the answer was still yes.
AI in your quoting system without sending prices out
An AI helper for a quoting system needs to read every rate and margin you have. So where it runs matters more than how clever it is. A design, not a shipped feature.
Our AI chat showed a parent a line of garbage characters
A reply from an AI assistant was cut off mid-sentence and its hidden control text reached the screen, with a button that would have put it into a child's lesson.
An app that worked for us and failed for every new user
Days before store submission, a fresh API key got a 404 on a model the provider still listed. Availability is per account, and every earlier test used our own key.
362 pricing rules to check, and nobody gets past forty
A fabricator's 362 spreadsheet formulas each needed a sign-off from somebody who knows the trade before a quote could be trusted. The review stalled at item forty.
Asked for fifteen questions, given twelve, told nothing
A children's learning app returned twelve questions when a parent asked for fifteen and said nothing. Fail, truncate quietly, or return twelve and say so.
The AI feature failed half the time and blamed the network
A question generator failed on about half of requests and blamed the network. The network was fine: the system had stopped after one of its five AI suppliers.
Can software sort vehicle photos without a big AI project?
A car rental business had 67,000 handover photos already tagged by staff. Reusing those tags sorted new photos with no training and no new hardware.
Letting staff ask the system in plain English, safely
A plain-English assistant over a business system must not show a junior the finance figures. We built the limits into what it can see, not what it is told.
Will AI make my software project cheaper? Partly
Where AI genuinely speeds up building software, where it quietly costs you more, and the one review rule a supplier should never loosen.