The post-mortem library

Eight failures. Named, costed, sourced.

Public, documented, and expensive. Every one of them was avoidable with a practice this playbook writes down.

Figures come from primary sources wherever one exists: the Slaughter and May review of TSB, the Queensland Health Commission of Inquiry, the California State Auditor, IEEE Spectrum, and contemporaneous reporting. Where a number circulates without a traceable primary source, it is flagged as such rather than repeated as fact. Dollar figures are approximate conversions at the time of the event, with the original currency in brackets.

What went wrong, in public

Eight failures. Named, costed, sourced.

Grouped by the mistake, so the pattern is visible before you read a single case. Every figure links to a primary source.

Went live before it was proventhree of the eight
TSB Bank
5.2 million customers moved in one weekend, with no tested way back
~$1.3B (£1B) · UK · 2018
Queensland Health
Ten aborted attempts, then go-live carrying 2,422 known defects
~$4M → ~$900M (AU$6.2M → AU$1.2B) · 2010
HealthCare.gov
End-to-end testing compressed from months into weeks by a fixed date
$464M → $824M · 2013
Rebuilt from scratchthe rewrite trap, at federal scale
FBI Virtual Case File
Requirements churned for five years and nothing ever deployed
$170M scrapped · 2005
Assumed the data meant the same thinga semantic mismatch, not a technical one
Lidl SAP
One system keyed on purchase price, the other on retail price
~$590M written off (€500M) · 2018
Governance failed the systemthe humans, not the software
Post Office Horizon
An institution prosecuted its own staff rather than doubt its software
~900 wrongful prosecutions
Birmingham City Council
Customisation broke core accounting and helped bankrupt a city
~$24M → ~$180M+ (£19M → £144M+) · 2022
Waited too longthe cost of not acting
California EDD
A deferred COBOL estate met a once-in-a-century load spike
~$20B fraud · 2020
The pattern behind the pattern

Five mistakes, not eight accidents.

Read together, the cases collapse into a short list of repeated decisions. That is the argument for the playbook: these are not bad luck, they are known failure modes with known counters.

Going live without evidence
TSB, Queensland Health and HealthCare.gov all shipped against a date rather than against proof. The counter is a parity bar and time in parallel treated as entry criteria, not aspirations.
Rebuilding instead of replacing
The FBI burned $170M on a big-bang rewrite that never deployed. The counter is incremental replacement with the legacy system as the source of truth until evidence says otherwise.
Assuming data means what you think
Lidl wrote off roughly €500M over a purchase-price versus retail-price assumption. The counter is treating semantics as a first-class workstream, before build.
Governance that cannot say "stop"
Post Office Horizon and Birmingham are failures of accountability, not code. The counter is one named owner per decision and a register that survives contact with politics.

Waiting is the fifth. California EDD is on this list for what was not done: a deferred COBOL estate that met a once-in-a-century load spike. See the twelve failure categories →