● Result
5 secmedian reply, down from 9 hours
Cake Inspiration
Meet Cake Inspiration's WhatsApp Agent: instant answers that get smarter on the job.
Cake Inspiration was answering ~90 WhatsApp enquiries a day by hand. Voltade's AI agent now handles the repeat ones in seconds, hands real orders to a human, and gets sharper every time staff correct it.
The challenge
- ~90 WhatsApp enquiries a day, all answered by hand: message by message, one staffer at a time.
- Most were repeat questions the website already answered: endless 'lazy-reader' lookups across 21 product categories and the FAQ.
- Replies took up to 9 hours, capped at staffed hours (10am–6pm), so leads went cold and after-hours messages queued overnight.
- High-value custom orders fell through: bespoke leads (photo prints, fondant figurines, logo cakes) were told 'we'll be in touch', but the handoff often reached no one.
- Localisation gap with an offshore sales team: a Philippines-based team struggled with local cake vocabulary (orh nee, MSW durian, money-pulling cakes, peranakan/mahjong themes, 100-day milestones) and the Singlish register, so replies read as generic and off-brand.
- No single source of truth: answers depended on what each staffer happened to remember, with no consistent version of the catalogue or policies.
The solution
One agent, three moves
Voltade built Cake Inspiration a custom WhatsApp agent that sorts every message into one of three moves: answer from a knowledge base crawled from the shop's own website (always in the shop's voice), point a catalogue shopper to the exact product page, or hand a custom order to a human designer and step back so the designer owns the chat. It never improvises an answer, and it never tries to close a bespoke order itself.
It learns on the job
And it keeps improving after launch. The agent learns on the job, with the owner as its supervisor.
From a correction to a lesson
When a reply misses the mark, a staff member leaves a note on the conversation and tags the agent. It reads the correction and proposes a lesson in its own words, with a confidence score and the reason behind it. From then on it's sharper at exactly the questions this shop gets asked. No engineer in the loop, no waiting on a developer.
Governed, with the owner in control
And the learning is governed. A correction runs through the same pipeline as any other change to the system. The agent scores its own confidence, and a router decides whether the fix is safe to apply on its own or should go to the owner for a one-click approval. Higher-stakes changes always wait for sign-off. Approved lessons become permanent memory, loaded into every future conversation, so the fix sticks.
Built for Singapore
PDPA-aligned, hosted in Singapore, PSG-eligible, and deployed in weeks, not months.
I can fix the agent myself now, without waiting on an engineer. I just tell it what I didn't like about a reply, and it learns. Sometimes it will automatically suggest improvements and it will ask me to approve it. I think that's pretty cool.
The learning loop
- 01
Feedback capture
A staff member writes an internal note on a conversation and @-mentions the agent. That wakes it in a strict coaching mode, with its customer-facing tools stripped away, so it can only think and learn.
- 02
The agent proposes a lesson
It drafts a patch rather than overwriting the file: a proposed change with the reason behind it and a confidence score.
- 03
The approval gate
A router weighs the agent's confidence against how sensitive the topic is. Low-stakes facts apply automatically; high-stakes rules go to a pending inbox for a human.
- 04
Human sign-off
The founder sees the proposed fix. Approve, and the lesson is applied cleanly. Reject, and the rejection becomes its own coaching signal, so the agent works out why it was wrong.
- 05
Permanent memory
Next time the agent wakes, the approved lesson is part of its memory. It won't make the same mistake twice, and it won't relearn a lesson it already holds.
Key results
Median reply time
9 hours → 5 sec
Staff corrections applied in week one
7, no engineer in the loop
F&B or retail shop drowning in repeat WhatsApp questions?
Show us a week of your WhatsApp inbox. We'll show you what an agent would answer in seconds, and how your team trains it.