CONTEXT
Team
2 Product Designers, 1 Product Manager, 3 Developers

My Role
Senior Product Designer
Timeline
2 weeks (Oct 2026)
Business Model
B2B SaaS - Construction & trades
I led the end-to-end design: research synthesis, the AI suggestion and confirmation pattern, and the Claude prototype used in reviews with the founder and engineering team.
Problem
Email enquiries were a dead end in Banta, so office admins retyped every client, site and job detail by hand.
With no trustworthy way to match existing clients, duplicates and junk enquiries slipped through to jobs.
AI reads each enquiry and suggests the client and site as inline chips that admins confirm or change.
Continue opens one review to accept all suggestions, then one click converts it to a job or quote with every detail accurate.
Outcome
Built for enquiry volume that grew from 10–12 to 40–50 a month: admins confirm details instead of retyping them.
Duplicate and junk records are stopped at intake, and one inline suggestion pattern is now ready to reuse across Banta.
Banta’s AI suggests which client and site each enquiry belongs to. The real challenge was designing how admins confirm those suggestions, quickly and with confidence.
Enquiries reach trade businesses from website forms, emails and phone calls. In Banta, each one has to be matched to the right client and site before it can become a job or quote. Admins check carefully, because a form or email doesn’t prove who sent it.
The old enquiry page worked against them: alert banners for every missing match, and email enquiries retyped by hand.
With around 80% of enquiries becoming jobs, every wrong match had a cost: duplicate clients, the wrong site on a job, and invoices that didn’t add up.
My role
As Senior Product Designer at Banta, I owned enquiry intake end to end, from research through to high-fidelity design. I walked through real intake workflows with an office admin, mapped pain points and existing product patterns in Miro, and built an interactive prototype in Claude to test the AI suggestion flow before committing to high-fidelity screens. I designed the inline suggestion pattern and the confirm-before-accept rules, working closely with the founder and engineering team to weigh trust, effort and technical constraints, and held the interface to WCAG 2.2 AA throughout.
THE PROBLEM
Every enquiry forced a trade-off between speed and accuracy
Admins could retype each enquiry carefully and fall behind, or move fast and risk linking the wrong client or site. Banta's earlier enquiry page didn't help either way: email enquiries were a dead end, and missing matches showed up as alert banners with three equal-weight buttons each. Underneath that were a handful of recurring problems:
The goal
Make converting an enquiry quick and trustworthy: no retyping, no duplicate clients, and the right site on every job.
About 80% of enquiries become jobs, so every intake error carries through to scheduling and invoicing.
Retyped by hand
Opportunity
By turning the details in every enquiry into AI suggestions that admins confirm in a click, we could end the trade-off between fast intake and clean data.
RESEARCH & INSIGHTS
I sat down with an office admin to walk through how enquiries really reach her and what she does with each one. Her habit stood out: email enquiries stay unconverted until she has called the customer to confirm they're genuine. With so many arriving, she doesn't trust any single one on its own. That told me the AI's job wasn't to skip her check, but to make it faster.

“There are so many emails coming in. I don’t turn one into a job until I’ve called them.”
— Office admin, trade business

What we learned, and what it changed
01
We learned
Admins don't trust an enquiry until they've checked it themselves.
↓
So we designed
The AI only suggests. Nothing is applied until the admin confirms it.
02
We learned
Admins take notes during the enquiry and refer back to them long after it becomes a job.
↓
So we designed
Notes and the original request carry over to the job or quote, so the context travels with the work.
03
We learned
An email address on a form isn't proof of who sent it.
↓
So we designed
Client matches always need confirming, even when they look right.
04
We learned
Skipped suggestions led to duplicate clients and junk records.
↓
So we designed
Accepting goes through one review that shows the suggested client and site together, so nothing is set without a check.
PROCESS
Reuse what works, design what’s missing
I started from how admins actually handle enquiries, mapped what Banta already had, and reused existing patterns wherever possible. I only designed something new where there was a genuine gap, then refined it through rapid prototypes and reviews with the founder and engineering team.
I walked through intake with an office admin to see how enquiries arrive and which ones she trusts. Synthesising pain points, goals and needs surfaced the core insight: admins don't trust an enquiry until they've checked it.
Decisions and trade-offs
Decision 01
Auto-link known senders
✓
Always confirm
An email on a form doesn’t prove who sent it. One wrong link puts the job, site and invoice on the wrong client.
Why it matters
Invoices go to the right client, not someone with a similar email.
Decision 02
Alert banners
✓
Inline suggestions
Banners interrupted the page with three equal-weight buttons. A suggestion sits where the data lives and asks for one decision.
Why it matters
Admins decide faster, without leaving the enquiry.
Decision 03
Confirm one by one
✓
Accept all
Separate confirms added clicks without adding care. One step still shows both suggestions, and either can be changed first.
Why it matters
One check instead of two, on every enquiry an admin handles.
AI guardrails
I designed the AI’s limits as deliberately as its features. Four rules shaped every screen:
✦
Rule 01
Extract, don’t guess
It reads names, emails and addresses from the form or email. It never invents details.
✦
Rule 02
Suggest, don’t fill in
Matches show as suggestions. Nothing is set until the admin accepts.
✦
Rule 03
The admin always decides
Accept all opens one review of the suggested client and site, and either can be changed before anything is set.
✦
Rule 04
A manual way through
When detection misses, like an email with only an attachment, convert to works by hand.
SOLUTION
AI does the reading. The admin makes the call.
01
Suggested, not filled in
The AI reads the website form or email and shows the matching client and site as dashed suggestion chips, right where the data lives. Nothing is set yet.
02
One review, accept all
Continue opens a single review of both suggestions. The admin can change either one before anything is set, then accepts them together.
03
Source in view while converting
The conversion modal is a split screen: the original form or email on one side, Notes one tab away, and the choice of job or quote on the other.
04
Everything carries over
The new job opens with the client, site, notes and original request already attached, and the list steps aside so the job gets the full width.
When it doesn’t match
No match found
Create a new client, or link an existing one.
The client modal shows previous site addresses to reuse, so a new client doesn’t become a duplicate.
Detection misses
Convert it by hand.
An email with only an attachment can still become an enquiry, job or quote from its Convert to menu.
Not a real enquiry
Reject it without creating work.
Junk is closed at intake instead of turning into a job that has to be cleaned up at invoicing.
The inline suggestion is now a shared pattern in Banta, already planned for supplier documents.
IMPACT
Intake accuracy carries through to scheduling and invoicing. Faster intake also means admins can call customers back sooner, and calls convert better than email.
What we intended
No retyping at intake
No duplicate or junk records
The right client and site on every job
What we observed
Reviews with the founder and engineering team settled modal-first conversion, AI detection with a manual fallback, and admin confirmation for every match.
The inline suggestion became a shared Banta pattern, already planned for supplier documents.
Enquiry volume grew from 10–12 to 40–50 a month, which made reliable intake more urgent.
What we’ll measure next · goal → signal → metric
Goal: trustworthy matching
Suggestions accepted without changes
Signal: admins accept the suggested client and site as they are, and don’t correct them later.
If it’s low: tune matching rules and show why a match was suggested.
Goal: no retyping at intake
Time from enquiry to job or quote
Signal: enquiries are converted in one pass, without admins leaving the flow to copy details.
If it’s slow: find the step where enquiries stall, and why.
Goal: clean client data
Duplicate clients created each month
Signal: enquiries link to existing clients instead of creating new ones that get merged later.
If they persist: make linking an existing client easier than creating one.
Business check: about 80% of enquiries become jobs and phone beats email for conversion, so faster intake should show up as earlier call-backs and a steady or higher enquiry-to-job rate.
REFLECTION
The hard part wasn’t the AI; it was deciding how much to trust it. Every review pushed the same way: keep the admin in control and make confirming fast. Next, I’d test the build against the prototype and track the measures above to see whether intake actually gets faster and cleaner.
Key learnings
01
Design the limits, not just the feature
02
Model the data before the screens
03
Prototype in code when the flow is the question
That’s it. Thank you for reading. 😇






