Case Studies

What recurs across every engagement

The case studies above are different businesses with different problems. The judgement underneath them is the same seven calls, made over and over. This is the part worth assessing before hiring anyone: not what they did, but how they decided.

1. 01

   Configuration is a hypothesis. Behaviour is the evidence.

   What a system is configured to do and what it has actually done diverge constantly.

   A workflow can be enabled, correct by inspection, and have enrolled nobody for two years. Routing can branch four ways by territory and rotate into two pools. A permission set can be the real cause of a symptom every configuration screen attributes to teams. So a finding derived from configuration alone is a hypothesis — it does not get a severity until the action log confirms it. This is also why a second reviewer reading the same configuration is worthless: they agree with the first one, because the configuration genuinely does say what it appears to say.

   • Three workflows, enabled and enrolling nobody
   • Four branches, two pools
   • Agents that show their evidence
2. 02

   Derive from ground truth, not from the document that describes it.

   The map is always older than the territory, and usually wrong in a way nobody has noticed.

   A territory map rebuilt from who actually owns records corrected two assignments the documented list had wrong. A fleet join rebuilt on a system-generated identifier held where the human-readable stock number silently failed on part of the fleet. Channel revenue read against the field that is actually populated came out at $5.07m against a documented $22,500. In each case the authoritative-looking source was the one that was wrong, and the data the business generates by operating was the one that was right.

   • Territory map from record ownership
   • Finding a join key that holds
   • A channel worth $5.07m
3. 03

   The symptom you were called about is rarely the problem.

   Fixing the thing you were blamed for is not the same as fixing what is wrong.

   A visibility complaint traced to a permission set rather than the team change everyone assumed. A report that an expiry date had stopped copying turned out to matter because a derived flag was telling churn reporting that paying customers had no contract. A channel reported as dead was a report grouping on a field that had stopped being populated. The discipline is to separate the reported symptom from the diagnosis and let each be decided on its own merits — including saying plainly when the thing you were asked to fix was not the thing that was broken.

   • A visibility complaint that was not about teams
   • The symptom nobody reported
   • Why a channel looked dead
4. 04

   Silent failure is the normal failure.

   Systems that fail loudly get fixed. The expensive ones report success.

   An API that accepts a filter, stores it, and never matches anything. A form endpoint returning success while discarding the submission. A parse template that captures field labels instead of values and creates every record perfectly. Webhook delivery that stops, which looks exactly like a quiet week. A report that renders a number that is wrong. None of these produce an error, a queue backlog or an alert. The only reliable defence is to verify at the far end — read the record, read the filter back, subscribe to a heartbeat — rather than trusting the response at the near end.

   • Accepted, stored, permanently inert
   • Detecting silence
   • A report that renders and lies
5. 05

   Decide where a fact lives, once.

   Two sources of the same truth always diverge. The question is when, not whether.

   Three systems each held part of the truth about a machine, and the answer was not to nominate a winner but to decide which owned which part of its life and what single pipe carried readings between them. A ticket association label was rejected because the pipeline already carried that distinction. A derived contract flag is never hand-set, because the date behind it is the fact and the flag is a consequence. Duplication is not always wrong — two records for one asset were correct — but duplicated meaning always is.

   • Arbitrating three systems
   • Never hand-set a derived flag
6. 06

   Reversibility before speed.

   The cost of being wrong should be an inconvenience, not an incident.

   A corrected control is date-bounded before it is enabled, so switching it on does not replay years of history. Existing rotation pools stay alongside new ones so rotation never empties mid-change. Requested deletions are held until the exports that make them safe exist, because stage-timing history does not survive even un-deleting a record. A deal and a company disagreeing on a date is left visible rather than overwritten blind, because overwriting replaces a disagreement you can see with a wrong answer you cannot.

   • Enabling without retro-blasting
   • Holding a requested deletion
7. 07

   Measure the week, not the backlog someone inherited.

   Most of what a CRM can count is a stock. Almost nothing worth managing is.

   Six of eight candidate scorecard components failed against live data — one field populated on a single deal portal-wide, another already at a hundred percent, a third measuring the three workflows that write it automatically rather than the person. Deals stuck in stage, deal value and close date all measure a backlog that existed on Monday morning. The two components that survived rank the team almost inversely, which is the evidence they measure different behaviours rather than the same one twice. And where the data cannot carry a metric honestly, the right answer is to say so rather than to ship it anyway.

   • Six metrics that measured the past