Private page. Shared directly, not listed on the site. Please do not forward.
Hatch · Industrial operations

The AI worked. Nobody trusted it.

An engineering firm built a digital twin that could predict optimal chemical dosing. Operators kept overriding it. The model was not the problem.

30%Less energy and chemical useClient reported, plant level
11m→2mTime to act on an alarmObserved in testing
3Product teams adopted the AI patternsInternal rollout
RoleLead UX and AI designer
Team2 design, 3 engineering, 1 data science, 1 PM
UsersPlant operators on rotating shifts
ShippedWeb and tablet console
The diagnosis

Every project turns on one question: is the obvious explanation the right one?

What everyone assumed

Operators need better predictions. If we improve model accuracy and surface more data, adoption will follow.

What was actually wrong

Operators were not doubting the numbers. They were being asked to sign off on a decision they could not explain to a regulator or to the next shift. Accuracy was never the blocker. Accountability was.

I sat through shift handovers. The pattern showed up there, not in the interviews. Operators were reconstructing the previous shift from memory and paper, then being handed a recommendation with no reasoning attached. Nobody signs their name to a number they cannot defend.

The live plant view: a 3D model of the treatment plant with a status dot on every stage, and cost and process readings down the right side.
Hatch AI9:45AM
Ammonia levels approaching noncompliance.
Current effluent ammonia levels are increasing, and nitrate levels are decreasing, which could lead to noncompliance in 2 hours.

Weather Warning

Heavy Rain in 24 hours
Predicted Rainfall: 100mm
Energy Consumption Peaks
Air Flow Rate Fluctuation
The real console. The plant is live, every stage has a status, and an alert says what is drifting and how long until it matters.
The calls I made

Show the reasoning, not just the recommendation

Every suggestion carries the two or three readings that drove it, in plain language.An operator has to justify the call to a regulator. If the system cannot explain itself, the operator becomes the one taking the risk.

Make the operator the decider, not the approver

Apply and Dismiss sit side by side with equal weight. Dismissing is logged as a normal action, not an exception.Approval flows imply the machine is right and you are the obstacle. That framing is what people push back against.

Design the handover, not just the dashboard

The console carries state across shifts, so the incoming operator sees what happened and what was decided.The handover was where trust actually broke down. Nobody had designed it because it was not considered part of the product.

A recommendation open beside its charts: ammonia, fraction to pass, dissolved oxygen and pH, each showing the current path and the path after the change.

Recommendation: Ammonia levels approaching noncompliance.

Plan outline
  • Disconnect the DO control to stabilize the oxygen supply.
  • Adjust Total Air Flow with specific changes:
    • Set Fraction to Pass 1 to 70%.
    • Set Fraction to Pass 2 to 30%.
Rationale
  • Reviewed real-time data showing rising ammonia levels and decreasing nitrate levels, along with lower-than-optimal dissolved oxygen readings, indicating a need for immediate intervention.
  • The recommendation to adjust air flow rates is based on historical data and previous successful interventions under similar conditions, ensuring reliability and effectiveness in the response strategy.
  • Simulated the impact of different air flow settings on oxygen levels and the nitrification-denitrification balance, leading to precise adjustment recommendations that target the specific imbalance.
  • Adjustments to the "Fraction to Pass" settings are tailored to quickly stabilize critical parameters, aiming to bring ammonia and nitrate levels back within compliance thresholds efficiently.
Every recommendation shows its plan and the reasoning behind it. Passing it to someone else takes a name and a sentence.
What shipped
Aeration demand, liveBefore

11 minutes before anyone acts

was 11 minutesdeviationacted

Live aeration demand. The shaded gap is the distance between the plant deviating and anyone acting on it, which is the part that cost money.

The plant overview, redrawn. The leader lines are the point: an alert that names the stage it came from is one an operator can check.

Plant overview, liveWatching

Holding target every stage inside its band

BlowersDigestersAerationClarifiersHeavy rain in 24 hours100mm predictedEnergy consumption peaksLow nitrogen removal2 recommendations foundEach one shows the readings behind it.
Basin 2 · turbidityNight shift

Holding inside threshold

Threshold 1.0 NTUbackwashshift change · 06:00day shift operatorResolved02:45Cindy JamesBackwash held 20 minutes.Reading returned to 1.0 NTU.Night shiftDay shift

The same record, before and after a shift change. The decision, who made it and what it did are all still there when the next person walks in.

The argument redrawn: one recommendation forks toward two operators, and only the trust survives differently.

Same modelIssued

Same recommendationissued to both

Recommendation issuedOverriddenOperator decided alone.Applied, loggedOn record next shift.Energy consumption peaksLow nitrogen removalno readings attachedReasoning, not accuracyThe gap that mattered.
The plant process view: influent, screening, grit removal, clarification, aeration, anoxic tank, disinfection and sludge, each as a card with its key reading.

Estimated Energy Cost

$103,400 /yr
0.5% over 24h
$105,000/yr target

Estimated Chemical Cost

$52,600 /yr
$60,000/yr target

Key Process Indicators

  • MLSS2723 mg/L
  • DO in Aeration Tank2.0 mg/L
  • Total Airflow2720 ft3/min
  • Sludge Production3.0 ton/d
  • Prim. Clarifier Loading0.41 lb/(ft2.d)
  • Sec. Clarifier Loading6.83 lb/(ft2.d)

Influent Flow Rate

48 Hours
30,00020,0000Aug-1823:00Aug-1905:00Aug-1911:00Aug-1917:00Aug-1923:00Aug-2005:00Aug-2011:00Aug-2017:00
Costs and process readings sit beside the plant flow. Charts show the forecast against the warning and limit lines, so a problem is visible before it lands.
If this sounds familiar

If your AI feature has low adoption, measure whether users disagree with the output or cannot defend it. Those are different problems with different fixes, and teams usually assume the first one.

Have a problem like this one?

Send a link and a sentence. If I can help, I will record a teardown. If I cannot, I will say so.

Start a conversation
Next case studyHatch · Mining value chain