Skip to content
allanninal.dev/build
Collaborate

Part 4 of 7 · Supplier lead time watcher series ~5 min read

How slippage gets caught

Slippage is the failure this system exists for. A single late delivery is visible to anybody; a supplier moving from six days to eleven over five months is visible to nobody, because every individual order was only a day or two worse than the last.

Key takeaways

  • Compare the recent window against the established baseline, not against the quote.
  • One late order is never an alert. Several above the old p90 is.
  • The acknowledgement segment drifts first and is the earliest usable signal.
  • Check for a seasonal explanation before raising it as a change.
  • The alert names the numbers and the orders, so the conversation starts with evidence.

Recent against baseline

How supplier lead time slippage is detected against a baselineA vertical chain of five steps entered by a box labelled A delivery is received and its lead time computed. Step one asks whether there is enough baseline, twelve or more prior orders; if not it exits to Just record it, with no comparison yet. Step two asks whether it is above the old ninetieth percentile from the baseline period; if not it exits to Normal, nothing to say. Step three counts how many of the last six were also above it. Step four asks whether that is three or more, a pattern rather than an event; if not it exits to Note it, watch but do not alert. Step five asks whether it is seasonal by comparing the same weeks last year, drawing on a side box for last year's same period; if so it exits to Expected, annotated rather than alerted. A note says four gates fire before anybody is told, and an alert that fires on one late order is ignored.AWS ACCOUNTA delivery is receivedlead time computedEnough baseline?12+ prior ordersJust record itno comparison yetnoAbove the old p90?from the baseline periodNormalnothing to saynoHow many of the last 6?also above itThree or more?a pattern, not an eventNote itwatch, do not alertnoSeasonal?same weeks last yearLast yearsame periodExpectedannotated, not alertedyesFour gates before anybody is told. An alert that fires on one late order is ignored.
Fig 1. How a drift alert is decided. Each gate removes a category of false alarm that would otherwise teach people to skip the email.
  • Database
  • App integration
  • Machine learning
  • Management
  • Analytics

Why the baseline is frozen

The comparison has to be against a fixed earlier period rather than a rolling window, because a rolling window absorbs the drift. A supplier sliding from six days to eleven over five months never looks unusual against their own last three months, which is precisely the failure mode this is supposed to catch.

So the baseline is the distribution from a stated earlier period — the previous twelve months, or the twelve months before the drift was last acknowledged — and it is re-based deliberately by a person rather than automatically.

The signal that arrives first

A supplier's acknowledgement and fulfilment delays over four quartersA stacked bar chart with four bars measured in days. Two series: sent to acknowledged in orange, and acknowledged to received in purple. January to March shows one day to acknowledge and five to fulfil. April to June shows two and five. July to September shows four and six. October to December shows four and nine. A note says the orange band doubled two quarters before the purple one moved at all.05101520~6Jan-Mar~7Apr-Jun~10Jul-Sep~13Oct-DecSent to acknowledged, daysAcknowledged to received, daysThe orange band doubled two quarters before the purple one moved at all.
Fig 2. One supplier over a year, split into the two segments they control. The acknowledgement delay is the leading indicator and it is free to measure.

A supplier taking four days to acknowledge an order when they used to take one is usually short-staffed, and short-staffed becomes late deliveries a quarter or two later. It is the cheapest early warning available and it comes from an email timestamp.

It is also a much easier conversation to open. “We have noticed acknowledgements taking longer — is everything all right at your end?” is a supportive question that often produces useful information, where “you have been late four times” is an accusation that produces a defence.

The seasonal check

A lot of apparent drift is the same slowdown that happened last year. Checking the equivalent weeks in the previous year before alerting removes the most common false positive, and where there is no previous year the alert says so rather than pretending to have checked.

What the alert says

The whole notification

  • Supplier: Hartley Components. Product: 4mm bearing housing.
  • Baseline: 6 days typical, 8 at the 90th, from 31 orders across 2025.
  • Last six orders: 8, 11, 9, 12, 9, 13 days. Four above the old 90th.
  • Acknowledgement: also up, from 1 day to 4.
  • Not seasonal: the same weeks last year averaged 6 days.
  • Effect if unchanged: the reorder point for this part should move from 8 days to 13.

Six lines, all of them facts, and the last one is the sentence that makes it actionable. An alert that reports a change without saying what it means for what somebody has to do gets filed rather than acted on.

Note what the alert does not do: it does not change the reorder point, it does not suggest a different supplier, and it does not assign a rating. Those are all decisions, and the next post is about why they stay with a person.

Next: turning a number into a conversation.

All posts