A client came to us with a problem they could feel but could not describe. Around forty people, several projects running at once, and a growing sense that work was taking longer than it should. Nobody could point at where the time was going.
That is a more common situation than it sounds. The information existed — people knew what they had worked on, project managers knew roughly where their projects stood, finance knew what had been spent. It was just scattered across individuals, spreadsheets and inboxes, and nobody had the full picture. Assembling it manually would have taken someone a day a week, so it never got assembled.
So the question was not how to collect more data. It was how to bring what already existed into one place, and then make sure somebody actually looked at it.
## What we built
A web application with three audiences, each seeing a different slice.
Team members logged their work against projects. This had to be fast — anything that takes more than a minute at the end of the day silently stops happening, and then your data is worthless. Short forms, sensible defaults, and nothing asked for twice.
Project managers got a live view of their own projects: where the workflow had reached, what had been invoiced, what had been spent against budget. Before this they were reconstructing that picture from memory and email threads every time someone asked.
The manager got a daily email. This turned out to be the most important decision in the project.
## Why the email mattered more than the dashboard
We built a dashboard first, and it was good. The manager visited it in the first week and then stopped, because visiting a dashboard is a task you have to remember, and it competes with everything else in a working day.
The email does not need remembering. It arrives every morning with the previous day's picture: hours logged against each project, workflow movement, spend against budget, and anything that had drifted. He reads it with coffee before the day starts.
The lesson generalises. A reporting system that requires people to go and look at it will be abandoned within a month. Push the information to where they already are.
## Adding AI, carefully
Raw numbers still need interpreting, and interpretation is exactly what a busy manager does not have time for. So each daily email includes a short AI-generated summary: what stands out, which projects are drifting, where the pattern differs from previous weeks.
Two rules shaped how this was built.
The AI only sees the data. It summarises what the system already knows, and it is not asked to speculate beyond that. Everything in the summary can be traced back to a figure in the same email.
It advises, it does not decide. The summary points at things worth a closer look. It does not tell anyone who is underperforming, and it never produces a number the underlying report does not contain. A system that quietly grades people would have been abandoned by the team within weeks, and rightly so.
That constraint is the whole trick with AI in internal tools. Used to summarise data you already trust, it saves real time. Used to generate conclusions the data does not support, it produces confident nonsense that someone eventually acts on.
## What changed
The immediate change was visibility. Questions that previously required someone to go and find out — how is that project tracking, what have we spent, where did last week go — became things you could answer in a few seconds.
The less obvious change was the conversations. When everyone can see the same numbers, discussions about workload stop being arguments about whose impression is right. The disagreement moves to what to do about it, which is a far more useful place for it to be.
## What I would tell anyone building something similar
Make logging effortless, because incomplete data is worse than none — it produces confident wrong answers.
Push the report out rather than waiting for people to come and get it.
Keep AI on interpreting data, not inventing it.
And decide early that the tool measures work, not people. That decision shapes every design choice after it, and getting it wrong destroys trust in a system faster than any bug.