AI customer experience
Your customers write on WhatsApp, open your app, scan a code at the table or talk to their phone. We design and build the AI that answers them on each channel, and hands over to your staff when it should.
What you get
- One concierge on the channels your customers use, giving the same answer on each
- Recommendations a customer can question, because each one carries its reason
- Voice in and out, on the phone itself where privacy calls for it
- Handoffs to staff with the conversation already summarised
- Cost per channel and the most frequent questions on one dashboard
Who it is for
Your customers ask the same questions on several channels and expect an answer in the one already open on their phone. Today your staff answer by hand, or a website chatbot answers some of them and misses the rest. This service is for hotels, restaurants, retailers, wellness providers and service companies that want every channel answered well, with costs they can see.
What we build
- Concierge agents. One reasoning and knowledge layer behind web chat and messaging, so a customer gets the same answer on any channel. Conversations are summarised, so context carries forward.
- Personalisation customers can read. Recommendations that state their reason, from rules your team can audit.
- Voice. Speech in, clean text out and natural read-back, on the device where the data should stay there.
- Ordering and booking with AI inside. QR ordering, booking and payment flows where an assistant helps the guest and the rules that protect them are enforced in code.
- The staff side. Handoff queues, conversation summaries, and a dashboard that clusters customer questions and estimates cost per channel.
Designed around your customers and your staff
We start by mapping every step a customer takes, on every channel, with a lane for agents and a lane for staff. Each handoff gets a trigger, a package of context, a response time and a record. A handoff without a response time becomes a queue nobody owns, so we also design the escalation for when the named person is away.
How an engagement runs
| Stage | What happens | Duration |
|---|---|---|
| 01 · Ground the case | Customer and staff research, the channel map, one metric such as answered questions or bookings | 3-5 weeks |
| 02 · Prove the value | A prototype on one channel, tested against cases from your own conversations | 3-6 weeks |
| 03 · Engineer the system | Channel connectors, knowledge layer, handoff queue and dashboard for the pilot footprint | 8-16 weeks |
| 04 · Pilot in one place | One venue, store or market, with handoff and override logs read every week | 6-12 weeks |
| 05 · Scale what works | More channels, sites and languages | 3-6 months |
| 06 · Run and evolve | Operation, cost per channel, question clusters and a quarterly value review | Ongoing |
Each stage is contracted separately, and the price is quoted on request once its scope is fixed.
Built for Switzerland
Customer conversations contain personal data, so revFADP applies from the first day, and personal data is replaced with aliases before a model sees it wherever the design allows. Allergen information has been mandatory for non-prepacked food in Switzerland since 1 May 2018, so an assistant that answers allergen questions must answer them correctly. In the EU, the European Accessibility Act has covered booking sites, apps and self-service check-in since 28 June 2025. Dates, articles and sources are on the regulatory reference.
Proof from the lab
| Build | Label | What it shows |
|---|---|---|
| A multi-channel AI concierge | In production | Web chat, Telegram, Messenger, Instagram and WhatsApp on one layer, cost-aware routing with fallback, and a staff dashboard with question clusters and cost per channel |
| Zeza | Beta | Speech, writing help and read-back in a natural voice, all on the iPhone |
| BlinkDine | Beta | QR ordering and payment for restaurants, with allergen safety, fees shown before payment, tip prompts only after service and no guest funds held, all enforced in code |
| Wellsura | Beta | Activity recommendations with reason codes from a fixed list; insurance status kept out of the ranking, enforced by a test |
| FoodHalo | Beta | A nutrition coach that must look up every figure it states |
| An adversarial evaluation harness for allergen safety | Internal build | 117 adversarial cases and a 100% pass rate to merge any change |
Built on
Questions
Which channels can a concierge cover?
The concierge we run in production answers on web chat, Telegram, Messenger, Instagram and WhatsApp from one reasoning and knowledge layer. Apps on iOS, Android and the web, and QR codes at a table or a counter, connect to the same layer.
How do you stop it from making things up?
Answers cite the passage they rest on, and the assistant hands the question to a person when nothing supports an answer. Where a wrong answer can harm someone, such as allergens, a frozen test set must pass in full before any change ships.
Can personalisation be explained to a customer?
It can be designed that way. Wellsura’s recommendations come from a weighted ranker that returns reason codes from a fixed list, so a member can always see why an activity was suggested.
When does a person take over?
At triggers written as rules a tester can check, such as a payment question or an answer the agent cannot support from its sources. Your staff receive what the agent saw, what it could not determine and what it recommends.
Does voice need the cloud?
Not always. Zeza runs speech recognition, writing help and read-back on the iPhone itself. The trade-off is hardware: it needs iOS 26 on an iPhone that supports Apple Intelligence.
Next step
Test it on one channel first
The Proof of Value Sprint builds a prototype on one channel in 2-4 weeks and evaluates it against your metric. If the metric is not fixed yet, start with the AI Opportunity Assessment. The price is on request.
See the Proof of Value Sprint