Case Study 4 of 4SHIPPEDEcolab Intelligence Platforms

Agentic AI That Notices First

Ecolab’s clients had people anchored to dashboards, watching for problems the data could have flagged on its own.

  • I designed agents that monitor data and installed equipment and turn what they find into Recommended Agent Insights: cards a person decides on.
  • Each card carries a type, an urgency and an owner, plus a path to help: the assistant, a live agent, or someone dispatched on site.
  • Piloted with nearly a dozen client partnerships. Industries, clients and images are generalized because the specifics belong to Ecolab and its clients.
The Ecolab3D Intelligence home page with the assistant open beside it during an onboarding step. The page greets the user and shows Top Priorities as cards with status and urgency tags, Key Metrics, and Insights and Opportunities. In the panel, the assistant explains what an insight card shows, repeats one card, and asks whether to continue to the next step.

At a glance

  • ~12

    Client partnerships in the pilot program, each adapted from one core experience, first per industry and then per client.

  • 76%

    Higher completion of Agent Insights when they lived in the chat than on a deterministic dashboard, in a two-week internal test.

  • 8–10 weeks

    To establish the core platform experience every industry adapted, at the end of 2025 and into 2026.

The Decision

What I Chose

I chose to turn what the agent notices into a Recommended Agent Insight: a card that asks a person to decide. A task only has a due date. Every Agent Insight also carries a type, an urgency, its freshness (how long it has waited), who created it and who owns it now, and a path to help, from troubleshooting with the assistant to a live agent to dispatching someone on site.

When a card surfaces in the chat, the user deals with it before moving on: review, act, delegate or defer. Deferring is always allowed, because someone on a site can’t always act right now.

What I Rejected, and Why

  • Our own first version: a static dashboard with an insight tagged on it and a chat beside it. The card didn’t know what the chat knew, and the action had to be found somewhere else.
  • A more deterministic dashboard, filled with everything a normal dashboard has, on the reasoning that the chat was enough. There is only so much room on a dashboard, and a two-week internal test settled it in favor of the chat.

Context & Constraints

A diagram of the architecture behind the Intelligence Platforms. Under a row of commercial subscriptions sit seven Intelligence products, each with its own data collection and tools. An agent for each Intelligence feeds an agentic orchestrator, the Ecolab Virtual Agent, which is driven by a large language model, draws on data, tools and documents, and creates the personalized experience.
Three views of the Ecolab Virtual Agent moving a user from the side panel to a focused full-screen view of an insight. In the panel, the assistant gives an overview of an insight and asks if the user is ready to continue. It then explains that it can switch to a focused full-screen mode and how to switch back, and the user says yes. In the full-screen view beside the product, the same insight shows its root cause, recommended actions, comments and activity.

Constraints

  • Autonomy Set at the Sale
  • Four-Tier Hierarchy
  • Deskless Frontline
  • One Core, Many Clients
  • “Agentic” Confusing

Case Study 1 shipped an assistant the user opens. The Intelligence Platforms asked the opposite question: what if the agent comes to the user first?

The brief to every team: no more dashboards to decipher, less time in the product, more time on the work and the people.

  • What agents watch (data sources and installed equipment) and when they fire is set with the client at the sale, not by the end user.
  • Who it served: Ecolab’s clients, and Ecolab’s own field service teams, who service the equipment behind them.
  • How it scaled: one core experience, adapted per industry, then per client by that industry’s Lead Designer, with design working like an in-house agency.

Each tier meets the agent on a different surface:

  • Owner or admin: a text outside the product above a threshold, desktop otherwise
  • Regional manager (three or four locations): Agent Insights at sign-in
  • Site manager: Agent Insights for one location
  • Frontline worker: not at a desk, hands busy, picks up a phone only when it’s vital

What We Built First, and Why It Broke

I designed a first version where the dashboard, the insight and the chat sat side by side and never met, and it taught me that what an agent notices has to arrive as a decision, not a chart with a tag.

Four Ecolab3D Intelligence screens showing the assistant's first moments: a greeting to the user, a getting-started message that offers to take them to their top priorities, and an insight opening in a larger view beside the assistant with its activity, comments and a chart.

Back to the drawing board...

Back to the drawing board...

Version one looked like progress: metrics, an insight card on top, the assistant in a panel alongside, and a text bringing the owner in from outside. But nothing connected:

  • The canvas was still a static dashboard.
  • The card didn’t carry the user into the chat, and the chat didn’t know which card brought them.
  • The action the insight implied had to be found somewhere else.
  • Nobody shared a definition of an insight, so every team drew the seam differently.

A Calm Start, and a Yes Before Full Screen

I designed the first moment of the day to calm and orient before it asks for anything, and made full screen, or any change to the canvas, something the user says yes to.

Four screens of the Ecolab Virtual Agent panel. The first shows a Good Morning greeting that says the agent is AI, not a person, with a Support Cases menu of three options and a Tools and Help list. The second shows the greeting scrolled, with a still-working message while the assistant thinks. The third shows an answer with a Double-Check Before You Apply warning and Review Protocol and Continue to Dosing buttons. The fourth shows an automated-decision result with a model confidence score and a Request Human Review button.

The day starts calm, with what needs attention first.

The day starts calm, with what needs attention first.

When a user signs in, the Ecolab Virtual Agent speaks first, with a calm welcome that adapts to the product, what happened last, the time of day and how they like to be greeted. Then the copilot panel opens with:

  • One sentence on what happened since they were last in
  • The AI disclosure
  • The top Recommended Insights with affordances
  • An offer to switch to full screen

From there the user can decide, delegate to their team, schedule meetings, send Teams and Outlook messages, and search product and sales documents for prospective clients. Both layouts share one anatomy (Agent Insights, favorite tools, the chat and persona-specific metrics), and which one appears follows the size of the task, the same rule as Case Study 1.

I also demoed changing the data itself through chat: narrowing the view or adjusting filters by describing it. Its consent model carries Case Study 2 into the workspace:

  • The user chooses full screen or not
  • Confirms a change before it is made
  • Can always undo it
  • Can escalate if it is serious

Any sense of AI taking over is a risk, so confirmations look different from chat text, and the user always knows they decided.

The Loop That Shipped

I designed the loop so the agent notices first and a person decides: sense, surface, show its work, decide, act or delegate, log, with “reviewed” counted as an action.

The agent notices first. The person decides.

  1. 1 · Sense

    Agents watch data sources and installed equipment. What they watch is set with the customer, at the sale.

  2. 2 · Surface

    A Recommended Agent Insight: AI label · type · urgency · owner · date.

  3. 3 · Show Its Work

    In the assistant, carrying the card, the page, and why the user is here: the problem · why it matters · how I know · proposed solution.

    → hands the decision to the right person

  4. 4 · Decide

    Persona-scoped controls: act · delegate · defer. "Reviewed" counts as an action.

  5. Approved Scope

    Nothing runs outside what the user approved.

  6. 5 · Act · Delegate

    The agent carries it out, or it is delegated or escalated, and it stays tracked.

  7. 6 · Log

    Every decision and every action, recorded.

Press Next to start with a signal. The agent notices first.

Signals decided: 0

Icons show who acts. Ink marks a gate.

  1. 1 · Sense

    Agents watch data sources and installed equipment. What they watch is set with the customer, at the sale.

  2. 2 · Surface

    A Recommended Agent Insight: AI label · type · urgency · owner · date.

  3. 3 · Show Its Work

    In the assistant, carrying the card, the page, and why the user is here: the problem · why it matters · how I know · proposed solution.

  4. 4 · Decide

    Persona-scoped controls: act · delegate · defer. "Reviewed" counts as an action.

  5. Approved Scope

    Nothing runs outside what the user approved.

  6. 5 · Act · Delegate

    The agent carries it out, or it is delegated or escalated, and it stays tracked.

  7. 6 · Log

    Every decision and every action, recorded.

  • hands the decision to the right person
  • the next signal
  • The agent notices first. The person decides.
  • User
  • Agent
  • Gate
  • Log

Icons show who acts. Ink marks a gate.

Three rules sit underneath the loop:

  • Controls are scoped to the persona. An owner can delegate to a specialist; a manager can act or defer; the agent acts only inside the scope approved at the sale.
  • “Reviewed” counts as an action, so a person looking and deciding not to act is recorded as a decision, not lost as silence.
  • Thresholds come from two places. Ecolab sets presets, some legally or industry mandated; each client tunes others within a range to match its risk posture.

The agent keeps monitoring whether or not the user acts. The card’s label, evidence and consent gate are the disclosure system from Case Study 2, and asking for a person follows the ladder from Case Study 1.

From the Working Files

From the messy middle: the research and the work in progress, as it happened.

  • A working board. At the top, an Ecolab product screen with the assistant open on a tickets and support menu, and a card saying Agentic Mode is active that lists three planned actions with Cancel, Review and Allow buttons. Below it, a comparison of generative AI (prompt, response, ends) and agentic AI (detect, reason, propose, consent, act, then loops back). At the bottom, an agent card with the problem detected, why it matters, how it knows and a proposed solution, with Decline and Consent buttons.
  • A research board comparing how other products place their AI assistants: side panels, full-screen chat and homepage chat in Google Workspace Gemini, Microsoft 365 Copilot and Rovo, plus design system references for agentic experiences.
  • Early designs for an Intelligence platform: a phone lock screen alert about a water issue, an Intelligence home page with a Good morning prompt and insight cards, and the Ecolab Virtual Agent panel walking a user through an insight and its recommended actions.

Where the User Stays in Charge

This is where Case Studies 1 and 2 become one system. The card says it is AI, shows its evidence, asks for a decision sized to the risk, records the decision and any feedback, and keeps a person reachable. The law required disclosure and oversight; the gates were my choices, set where a mistake would cost the most.

Evidence & Outcome

  • Nearly a Dozen Client Partnerships

    A pilot program adapting one core experience per industry and then per client, built as a partnership between Ecolab, its field service teams and clients from the Fortune 1000 to the Fortune 50. No launch numbers exist; these were pilots, and I’d rather say so.

  • 76% Higher Completion

    Two teams each took one concept into a two-week internal test: Agent Insights in the chat, where the user could also reshape the data by describing it, versus a deterministic dashboard. The chat-based concept had a 76% higher completion rate, meaning cards were reviewed or decided rather than left for later. Where a decision lives changes whether anyone makes it.

  • 8–10 Weeks to the Core

    The core platform experience every industry adapted, established at the end of 2025 and into 2026 on the Case Study 1 foundation and the Case Study 3 design system. Owner alerts by text shipped as the entry point from outside the product.

Pilot count and dates are from my own records. The 76% comes from a two-week internal test comparing the two concepts. Client and industry specifics can’t be shown, so the numbers here are the ones I can stand behind without them.

What It Cost, and What I’d Do Differently

What It Cost

  • No dedicated QA or UAT. Whole platforms launched in eight to ten weeks, so user testing turned into client QA, and many first impressions were search with suggested prompts that never reached the canvas.
  • A frozen foundation. It was set in those weeks, leaving little room for the research I kept bringing in.
  • No shared definition. Nobody agreed on what an insight was, so type, urgency, freshness and ownership were argued team by team instead of designed once.
  • Limited feedback. The design team saw pilot feedback only in the first few weeks.

What I’d Do Differently

  • Define the card with leadership before the first platform: what “agentic” means, the five types, what urgency and freshness mean, and who can be assigned, including the Ecolab Virtual Agent itself, which was still a future concept.
  • Ask for three to four more weeks of QA and UAT, so clients focus on the experience instead of bugs.
  • Give testers a guide every week or two and keep synthesizing feedback, so each round asks them to notice something new.
  • Let a card close its own loop when the decision is small, instead of forcing every Agent Insight into a conversation.