What we saw
The messy path people actually used.
Case Reference Before.
Before.
Change. After.
We worked in real buildings, sales systems, and AI platforms. The pictures below are the work. Names of companies stay out. Made-up savings stay out.
What we named
Who does what, in what order, when it breaks.
What we write down
The start, the change, the check, the keeper.
The shared path
Every loop has four boxes.
Selected stage Start
What is true today, before anyone changes anything.
Evidence produced The old path, who waits, and the number we will compare later.
Where the judgment comes from
Three kinds of real systems.
A building can say it saved power. We ask: saved compared to what?
- Energy saving strategy and verification
- Building operating systems
- Chiller optimization
- Overall system efficiency and retro-commissioning
If four teams count the same sale four ways, nobody knows the number.
- Sales and marketing data platform
- CRM data flows
- Product UX monitoring and tracking
A room full of computers is not a system until someone owns the next experiment.
- Data governance lake and marketplace
- NVIDIA DGX H800 computing
- MLOps workflows
- Agent platform
- AI ecosystem enablement
- AI computing platform
Operating loops
Left is messy. Right is named.
Energy verification loop
If you never write the starting number, “we saved energy” is only a story.
Before · what hurts
- 01 The bill dropped, so the plant calls it a saving
- 02 A colder week and empty floors are mixed into that story
- 03 The chiller was tuned, but the comparison lives in a notebook
- 04 Next year’s auditor cannot rerun the same comparison
After · what we named
- 01 Baseline is the last comparable season, weather-adjusted
- 02 The change is named: which chiller, setpoint, and hours
- 03 The new reading uses the same meter and the same method
- 04 Facilities can open the notes and defend the number
A saving is only real if a stranger can replay the comparison next year.
Baseline methodIntervention scopeMeasured resultBusiness decision
Building operating loop
A sensor that nobody answers is just a noisy toy.
Before · what hurts
- 01 A sensor alarms at 2am on one screen
- 02 The dashboard is green on someone else’s laptop
- 03 The night operator remembers “last time we raised the setpoint”
- 04 The morning shift cannot see what was actually done
After · what we named
- 01 This signal means this action — not “look at the chart”
- 02 This role is on the hook, including after hours
- 03 The allowed action and the stop condition are written
- 04 The next shift can see whether it worked
Telemetry is useless until a named person has a named next move.
Response pathManual steps removedControl boundaryAdoption owner
Commercial number, one rule
If four people count the same marbles four ways, nobody knows how many there are.
Before · what hurts
- 01 Marketing counts a lead as a form fill
- 02 Sales counts a lead as a booked meeting
- 03 Product counts a lead as first activation
- 04 The weekly slide uses a fourth number, rebuilt by hand
After · what we named
- 01 “Lead” means one event, in one system
- 02 Ads, CRM, and product use that meaning
- 03 A broken feed is a quality fail, not a meeting argument
- 04 One person owns the dictionary
The fight stops being “whose number” and becomes “what should we do”.
Source inventoryData contractQuality checkDecision owner
Technology bet, small test
A cool demo is not a yes. A small test with a written answer is.
Before · what hurts
- 01 The H800 demo looks fast on clean sample data
- 02 Real records are slower, messier, and access-limited
- 03 Nobody wrote what would count as “this failed”
- 04 The bet is already spoken about as if it were production
After · what we named
- 01 The question is one business decision, not “is AI cool”
- 02 The test is the smallest run that could fail on real data
- 03 A boring or negative result is written without spin
- 04 Continue, stop, or reshape — with a date
A demo can start a conversation. Only a scored test can spend the next quarter.
QuestionTest setupOrdinary resultGo/no-go reason
AI platform with a front door
A kitchen full of ovens still burns dinner if nobody writes the recipe.
Before · what hurts
- 01 Anyone in the programme can start another agent project
- 02 Training data and tenant files move without a gate
- 03 There is no score for “good enough to keep running”
- 04 When it breaks at 11pm, there is no named owner
After · what we named
- 01 A project starts only after intake: data, compute, owner
- 02 Access is listed and limited before the first run
- 03 An eval exists before the experiment can scale
- 04 The same owner stays from proposal to handover
Compute and agents stay cheap only if starting a project has a door.
Intake gateEvaluation metricRuntime boundaryOwner
Findable, then citeable
A good book with no library card cannot be borrowed.
Before · what hurts
- 01 The article is live, so the team thinks it is findable
- 02 The crawler is blocked or sees an empty shell
- 03 Search and answer engines quote a different company
- 04 The response is to publish more, not to fix access
After · what we named
- 01 First: can a machine open this URL?
- 02 Then: is the entity labeled so it can be cited?
- 03 Then: is the source trusted enough to quote?
- 04 Fixes are ordered. More pages come last
If a machine cannot read the page, publishing another page does not help.
Access checkSchema checkSource authorityAnswer fit
Next
Pick one messy path.
We write the start, name the change, and agree what “checked” means before anyone builds.