The problem raised while it is still cheap.
Operational failures are rarely sudden. A stockout was visible in projected quantity a fortnight earlier, a slipped milestone was visible when its dependency moved, a bad margin was visible on the third order. The signal was in the records. Nobody was reading them on a Tuesday.
Reads projected, not actual·Named cause, not an alert·Prompts, does not act
How this goes wrong today.
Not for want of reports. Most businesses have plenty. The gap is between a report existing and somebody reading it in time to do something.
- 01
Reports nobody opens on the right day
The shortage report exists, is accurate, and is opened on a Friday when the order needed placing on Monday.
- 02
Reacting at the point of failure
The line stops, the customer calls, the batch expires — and the response is expediting, which costs several times what planning would have.
- 03
Alerts that are noise
A threshold alert firing so often that it is filtered into a folder, which is functionally the same as not having it.
- 04
Margin found at the end
A job's real cost assembled after delivery, when the only remaining option is to quote better next time.
- 05
Everyone knows, nobody owns
The problem is common knowledge in one team and news to the team that could have prevented it.
What runs instead.
The difference from an alert is that these carry a cause and a proposed action, and they arrive early enough for the cheap fix to still be available.
Shortage against lead time
Projected quantity — actual, plus ordered and planned, less reserved and indented — watched against reorder level and real lead time, not against what is on the shelf.
Slippage with its consequence
When a task's dependency moves, the effect on the milestone and on the billing date is stated, rather than the date quietly changing.
Expiry and slow movement
Batches approaching expiry and stock that has not moved in a period, ranked by value, while there is still time to sell or move it.
Margin drift during delivery
Accumulated project or job cost tracked against estimate as it happens, raised while something can still be done rather than at final invoice.
Failure patterns named
Rejections grouped by item, supplier, batch, workstation, and shift, with the pattern stated instead of left in a table for somebody to notice.
Maintenance and downtime
Scheduled tasks coming due, and assets whose repair cost is running against remaining book value — while replacement is still a decision.
What the agent may and may not do.
This use case is deliberately the most conservative of the four. The agent's job is to be read on a Tuesday, not to act.
- May read
- Stock, orders, projects, jobs, assets, and quality records across the business
- May do
- Raise material requests within a ceiling, and surface findings with a named cause
- May not do
- Commit spend, reschedule production, change a price, or close a finding
- Ceilings
- A value limit on any request it raises, per period and per warehouse
- Ownership
- Every finding routes to a named person, not to a dashboard
- Suppression
- A dismissed finding stays dismissed until the underlying figure changes materially
- Trail
- The records behind each finding are named, so it can be checked rather than believed
The modules underneath this.
The signal comes from across the business, which is exactly why this is difficult when operations and finance sit in different systems.
How to tell whether it worked.
Measure the outcomes, not the number of alerts sent. A system that produces more findings is not thereby better.
- Stockouts against forecast shortages
- How many actual stockouts had been raised in advance, and how many arrived without warning.
- Expediting cost
- Premium freight and short-notice buying, which is the direct cash cost of finding out late.
- Write-off from expiry
- Value written off for expiry or obsolescence, against the value flagged in time to be moved.
- Milestone dates held
- Projects delivering to the date they were last replanned to, rather than to a date quietly abandoned.
- Margin variance at completion
- Final cost against estimate, and how early the variance was first raised.
- Findings acted on
- The share that produced an action rather than being dismissed — the honest test of whether these are worth reading.
The programmes this normally arrives inside.
Next step