Crane — Conversational AI UX
Case study · 2026
Balakumar E · Product Designer
● 001
Client’s Pain
The project followed a structured, research-first process.
/ 01
Discovery & research
Brief, stakeholder interviews, current-state workflow, competitor study
/ 02
Synthesis & definition
Interview matrix, affinity map, personas, jobs-to-be-done, MVP scope
/ 03
UX audit
60-screen walkthrough, friction log, 20 issues, priorities
/ 04
Strategy & structure
Design brief, AI principles, engineering constraints, user flow
/ 05
Wireframes & design
Wireframes, low-fi options, 19 hi-fi screens, AI patterns
/ 06
Handoff & measurement
Audit check, review with Dom, dev handoff, launch metrics
● 002
Introduction — Crane
Support teams move years of tickets between help desks. Today that means weeks of manual mapping, spreadsheets and log-reading. Crane was built to remove that work.
● 003
Discovery Phase
/ Problem
Specialists juggle several tools, spreadsheets and manual error handling — so large migrations are slow, fragile and non-linear.
/ Value proposition
An AI-assisted workflow end to end: connect, profile, map, validate, run and reconcile — with pre-execution checks.
/ Success · Dom
Migrations that took weeks or months, done in 1–2 weeks — faster, safer and less manual, without losing accuracy.
Primary users
Businesses that need support data moved
Secondary users
Migration specialists (forward-deployed engineers)
Business goal
Faster migration, low error rate, zero crashes
User goal
Easy mapping, clear errors, reliable filters, templates
D
Dom
Product Owner · vision, strategy, approval
B
Bala (me)
Product Designer · research, design, audit, testing
M
Mani
Dev Lead · feasibility, architecture, CUI
S
Specialists
Workflow validation & data handling
Fourteen steps — and a loop back to mapping every time the sample failed.
01
Migration request
02
Select source
03
Select target
04
Source file
05
Process source
06
AI analysis
07
AI auto-mapping
08
Objects & config
09
Pre-flight settings
10
Run pre-flight + sample
11
Fix mapping / errors
12
Start live migration
13
Data migrated
14
Post-run verification
● 004
Research Phase
Dom
Product Owner
Focus — Success, outcomes, scope
Mani
Dev Lead
Focus — Tech debt, constraints, handoff
Specialist 1
Senior FDE · multi-client
Focus — Day-to-day friction, trust
Specialist 2
Mid-level FDE
Focus — Stability after recent fixes
Specialist 3
Mid-level FDE
Focus — Feature parity, regressions
Specialist 4
Senior FDE · data wrangler
Focus — Failed records, cleanup
“Make complex platform migrations significantly faster, safer and less manual — while keeping data accurate.”
— Dom · Product Owner
“Edits only autosave every 2 minutes; leaving the step or refreshing overwrites work.”
— Mani · Dev Lead
“I’d like a timer that estimates how long a migration will take… and a notification when it pauses.”
— Specialist 1
“When I apply a date-based filter and click Test & Count, the tool sometimes returns an error.”
— Specialist 1
“Retrying failed records with no clear explanation was challenging — we had to work out the issue manually.”
— Specialist 4
“All the problems we used to have are mostly fixed in the current version.”
— Specialist 2
Quotes lightly trimmed from interview notes.
● 005
Synthesis Phase
Error visibility
Blind spots in long jobs, unexpected pauses, raw backend error strings
Mani, Specialists 1 & 4
High
Confidence in data scope
Filter validation errors, record-count mismatches, silent auto-mapping overrides
Specialists 1 & 4
Critical
Manual overhead
Spreadsheets, CSV column fixes, timestamp conversion, sample cleanup
Specialists 1 & 4
Medium
System stability
Missing AI routes, fragile 2-min autosave, no conflict protection, heavy polling
Mani
Critical
Engine maturity
Recent fixes solved legacy issues for some; edge cases remain for others
Specialists 2 & 3 vs 1 & 4
Mixed
/ 01
Data safety & control
1A Save states
Mani: 2-min autosave overwrites work
Specialist 1: refresh anxiety
1B Auto-map overrides
Specialist 1: refresh re-maps my fields
Insight
The system must never change things silently.
/ 02
Visibility during runs
2A Progress & ETA
Specialist 1: wants a timer + ETA
Mani: long jobs time out unseen
2B Pause alerts
Specialist 1: POD pauses go unnoticed
Specialist 4: dataset visibility gaps
Insight
Migrations can’t run as black boxes.
/ 03
Errors & guidance
3A Plain-language errors
Specialist 1: no failure reasons in CSV
Specialist 4: raw log digging
3B Data wrangling
Specialist 4: columns, timestamps, cleanup by hand
3C AI copilot
Specialists 1 & 4: assistant to explain and fix
Insight
Turn backend errors into actionable fixes.
Forward-deployed migration engineer
Goals
— Deliver migrations in 1–2 weeks
— Keep data 100% accurate
— Fewer custom scripts
— Easy customer sign-off
Pain points
— Silent data loss on refresh
— Opaque errors, raw logs
— Manual timestamps and columns
— No real-time progress
Knows
Source and target systems, objects and fields
Doesn’t know
Crane’s agents, scores or internal terms
Wants
To move data fast and know before live that it will work
Fears
The AI changing data she can’t see or undo
“Don’t teach me migration. Tell me what you did and what you need from me.”
Design implication → make Crane’s own steps, decisions and results readable.
01 · Critical
Data safety
Save-status chip · pre-exit warning · lock against auto-map overwrites · filter “Test & Count” states
03 · Medium
Data hygiene
Skeleton loaders · one-click sample cleanup · CSV column checks · timestamp and dropdown helpers
02 · High
Visibility & AI
Pause banners with cause + Resume · ETA timer · long-running job card · failure reasons in CSV · plain-language errors · AI fallback states · preview → confirm → undo
Deferred
Self-healing runs without human confirmation · canvas workflow builder · multi-tenant workspace switching
● 006
Audit Phase
Before · original screens from my test migration
Trust & safety
AI picks labelled “set by you”, no undo, risky Restart
Navigation & structure
Stuck in the wizard, two steppers, no Run sample button
Language & clarity
recipe, transjson, run_preflight, raw JSON in chat
Feedback & counts
Six readiness formats, counts disagree, blocking loaders
High
Blocks or misleads — e.g. no way back from setup, no Run sample button
Medium
Slows or confuses — e.g. pre-flight looks automatic, progress not live
Low
Polish — e.g. “Preview Records” label, no completion moment
Problem 01
No way back
Opening setup from the chat was a dead end.
Problem 02
Hidden next step
A sample run had to be typed as a command.
Problem 03
Log-style AI
Recipes, API paths and agent names in replies.
Problem 04
Decisions without owners
AI picks labelled “set by you”; no undo.
Problem 05
No end moment
Runs stopped with no summary or next step.
Solution
Back on every step
“Back to setup” in the phase bar; Back on every step.
Solution
Always a next step
The main action is a button above the chat input.
Solution
Human language
“Crane AI” replies in plain words; details on demand.
Solution
Show your work
“Suggested by Crane” vs “Set by you”, confidence, Undo.
Solution
A clear ending
Completion card, report, retry and a toast.
01
Always give a way back
Low effort
02
A “Run free test” button
Low effort
03
Plain-language assistant messages
Medium effort
04
Post each blocker list once
Low effort
05
One count and one score everywhere
Medium effort
06
Label AI decisions as AI
Medium effort
07
Undo after every AI or bulk apply
Medium effort
08
Blocker rows open a focused fix
Medium effort
Priority 1 · must fix
#11 way back · #16 Run sample · #13 fixable issues · #12 Resolver · #1 #5 Billing & nav
Priority 2 · should fix
#17 action by input · #15 pre-flight · #7 #8 #9 setup clarity · #18 #19 live progress
Priority 3 · polish
#3 #4 #6 #10 #14 #20 — risky actions, labels, AI wording, completion
● 007
Strategy Phase
/ Phase 01
Setup
01
Dashboard
02
Connect
03
Workspace
04
Choose objects
05
Which records
06
Records fetched
07
Scope
/ Phase 02
Prepare & map
08
Conversation
09
Preparing migration
new
10
Map fields · Resolver
/ Phase 03
Check
11
Pre-flight · issues
new
12
Pre-flight · passed
new
13
Sample run · live
new
14
Sample run · results
new
/ Phase 04
Go live
15
Add records
16
Payment successful
new
17
Live run
18
Run completed
new
19
Dashboard · completed
new
01
Show your work
Every AI step says what it did, where and why — stages, counts, “Suggested by Crane”.
03
Always a next step
The main action is a button above the input. Typing is optional, never required.
02
Every action undoable
Preview → confirm → highlight → undo. Back on every step, Edit and Undo on every mapping.
04
Confidence over speed
Label confidence; low-confidence picks wait for the user.
05
Human language
No agent names, JSON or log lines in the chat.
06
Handoff, not cliff
A blocker opens the exact thing to fix.
● 008
Final Design
Field and value mapping moved into a work surface: plain names first, one primary action per card, “Suggested by Crane” vs “Set by you”.
Blockers become a checklist, posted once. “Fix →” opens the exact field in the Resolver; re-run in one click.
02
Fix opens the exact field
Covers audit #13 — issues weren’t obviously fixable.
Test 25 records before trusting 17,610. One click to start; results show passed and failed records with links into Freshservice.
1 click — was a typed command
Covers audit #16, the most-cited blocker.
● 009
Outcome
UX issues solved · partly solved
readiness formats unified
to start a sample run
end-to-end screens handed off
Start a sample run
Typed command only
1-click button above the input
Readiness score
6 different formats
1 score, everywhere
Blocker list
Posted up to 6 times
Posted once, as a checklist
Way back from setup
None — dead end
Back on every step + phase bar
Who made a decision
AI choices shown as “set by you”
“Suggested by Crane” vs “Set by you”
Undo after applying
Not available
Edit and Undo on every mapping
Progress of a run
Stepped, not live
Live card with counts and %
End of a run
No summary or next step
Completion card, report, retry, toast
/ 01
Designing for AI is designing for trust
Show the reasoning, label the AI, never let it act silently.
/ 02
Research earns the right to redesign
Interviews and a real walkthrough made every change defensible.
/ 03
Constraints are design input
No streaming and call caps became honest states, not hidden failures.
/ 04
Split feedback is a signal
Two happy and two struggling specialists — design for the edges.
