Verdigris • sep - dec 2025 • in alpha

from chatbot to proactive ai assistant

i was hired to usability test an ai chatbot for data center technicians. i recommended not shipping it — and led the pivot to arc, a proactive assistant now in alpha.

sole designer (research → strategy → design) · 9 usability sessions + 2 contextual inquiries · team: head of product, engineering, 3 research collaborators

02 — context

verdigris catches equipment failure in data centers before it happens.

they built a chatbot to help technicians understand the data, and asked us to test one thing: is this the right product?

"a lot of data center technicians don't know what to do with the data itself. they don't know how to analyze it. they just want the building to run efficiently."

jimit shah, head of product @ verdigris

03 — the test

i set five ship criteria and tested against them.

the chatbot assumed technicians would notice an issue, open a tool, ask the right question, and judge the answer — all on them.

04 — what i found

0/5

criteria met — recommendation: do not ship

glanceable

scan severity in under 10 seconds

✕ failed

8/9 had to read first

actionable

leads to a clear next step

✕ failed

8/9 wanted what to do

trustworthy

shows source and confidence

✕ failed

9/9 asked where it came from

field-ready

works on mobile and tablet in the field

✕ failed

desktop-bound

recoverable

verify, escalate, or hand off when unsure

✕ failed

no path

the root cause: the chatbot put the work on the technician.

the ai should do the opposite — detect the issue, surface the context, explain the risk, recommend the next step, and show its reasoning. that reframed the question from "how do we make the chatbot easier to use?" to "how do we bring ai into the workflow at the moment it's needed?"

05 — making the case

i grounded the pivot in evidence, not preference.

i showed leadership that a passive chatbot was a business risk — alert fatigue and slower resolution — and used the ship criteria to keep the call objective, not a matter of taste. in the final review, i walked the cto through the mismatch between a blank prompt and a technician's real-time needs.

what changed in the room

from

"can we build a bot?"

to

"how do we design trust?"

06 — the pivot: arc

same ai. new surface. arc starts with the problem, not a prompt.

old — chatbot (all on the user)

notice the issue → know which data matters → open a separate tool → ask the right question → judge the answer → decide what to do

new — arc (system-initiated)

system detects anomaly → push notification → tap → alert view (cause · severity · source · confidence · next step) → act, escalate, or monitor

push notification entry

fixes: blank prompt

alert-first detail view

fixes: not glanceable / no next step

confidence + source data

fixes: no trust

recommended next steps

fixes: no action

persistent handoff context

fixes: outside the workflow

safe by design: severity thresholds control alert fatigue; the dashboard stays as a manual fallback; language is "recommended," never "do this." the technician always decides.

07 — impact

verdigris accepted the recommendation and moved arc into alpha.

my most valuable contribution was the call not to ship — turning usability findings into product criteria changed the product's direction, not just its interface.

08 — reflection

not every ai product should be a chatbot. chat works when users know what to ask. here they didn't — so the answer wasn't a smarter chatbot, but an assistant that notices, contextualizes, and guides.