“Can I trust an AI with my customers?”

01The short version: she’d rather say “let me check” than guess.

Sally only speaks from what you’ve entered, refuses to invent what she doesn’t know, hands anything uncertain to a real person, and is re-tested before every change ships. The rest of the page shows the receipts.

She can’t quote a price you never set.

Every reply is built only from the services, prices and hours you entered. The open internet stays fenced off. Watch it happen below.

02Grounded in your business, not the internet

The single most important design decision, and the one that prevents the scary “AI confidently invents nonsense” problem.

When you set up Suddenly Sorted, you tell Sally your services and prices, your team, your opening hours, and your policies (cancellations, deposits, and so on). That structured information becomes Sally’s knowledge base, and it’s the only thing she’s allowed to speak from.

Every time a customer messages, your business’s real details are loaded fresh and placed in front of Sally as ground truth. She is explicitly instructed, in her core operating rules: “Never make up services, prices, staff names, or hours not listed.” If a customer asks about something you don’t offer, she doesn’t improvise. She says she’s not sure and offers to have the team confirm.

She also never sees (and so can never leak) your private notes. Things you jot against a client (“always runs late”, “allergic to a dye”) are kept out of Sally’s reach entirely. Even if a customer asks “what do you have on file for me?”, she can’t read it back.

03Six ways a bot could burn you. Six rules that block them.

Each one has a specific, enforced rule against it — not vibes. Here’s the list.

• She never says “booked” until it actually is. Sally is forbidden from telling a customer they’re confirmed until the booking has truly been saved and the system confirms it. No phantom appointments.
• She won’t double-book. Bookings are checked against your calendar, your opening hours, and past dates before anything is saved, and the system is built to reject two people grabbing the same slot at once.
• She invents nothing. No made-up prices, services, staff names, hours, or email addresses. If it’s not in your setup, it doesn’t come out of her mouth.
• She moves bookings, she doesn’t duplicate them. When a customer reschedules, Sally changes the existing appointment rather than creating a confusing second one.
• She knows your policies. If a service needs a deposit, she flags it and lets the customer pay it right there in the chat to lock the booking in, through a secure Stripe checkout, so she never handles raw card details herself. The deposit goes straight to your Stripe account and we take no commission.
• She passes it to your team when stuck. Unlisted request, unusual question, or a conversation going in circles? She collects a name and number and logs it for your team as a callback, instead of guessing.

04Nothing reaches your customers untested.

Every change runs 2,200+ automated checks plus real booking conversations replayed against the same model that powers your live chat. If a single check fails, the change doesn’t ship.

Two layers of testing sit between any change we make and the version that reaches your customers.

Layer 1: Automated tests (the safety net)

A suite of over 2,200 automated tests runs on every single change. These are exacting, pass/fail engineering checks of the machinery around Sally, the parts that must never break. We don’t ship a change unless all of them pass. They cover things like:

• Double-booking protection: two customers racing for the same slot can’t both win it.
• Business separation: one business’s data is completely separated from every other business’s.
• No duplicate or late confirmations: saving a booking twice, or sending a stale email, is caught.
• Contact details are valid: emails and phone numbers are sanity-checked before they’re stored.

Layer 2: The evaluation harness (testing Sally herself)

Automated tests check the plumbing. To check Sally’s judgement, we built an evaluation harness: a library of realistic, scripted customer conversations (a regular booking, a reschedule, an awkward question, a complaint, someone trying to book a service you don’t offer) that we replay through the exact same Anthropic Claude model that powers your live chat. Same brain, controlled exam.

Crucially, we don’t just read the transcripts and nod. Each scenario carries machine-checked expectations called oracles, so grading is objective and repeatable.

A few of the everyday scenarios

To make that concrete, here are three of the booking conversations we replay, and the exact thing each one checks. These are the bread-and-butter cases your customers send every day.

05How her answers are scored

Every test scenario asks pointed questions about what Sally did. Most are graded by exact rules; the subtler ones by a second AI acting as an examiner.

Exact, rule-based checks

These are black-and-white: they inspect what actually happened in the system, not just what Sally said:

The AI examiner (for the judgement calls)

Some qualities can’t be measured with a simple yes/no rule, so a second, separate AI model acts as an independent judge, scoring things like:

• Did she invent a price? The judge is given the only prices allowed for that business and fails Sally if she states any figure that isn’t on the list.
• Did she ask instead of guess? When she couldn’t know something, did she actually find out rather than fabricate?
• Did she stay on-brand? Warm, clear, and human, the way a great receptionist sounds.

06Guarding against “making things up”

The thing everyone actually worries about, addressed head-on.

“Hallucination” is when an AI states something false with total confidence. We attack it from three directions at once:

07How she gets better over time

Sally isn’t frozen. She improves, and the way she improves is designed so she can’t quietly get worse.

Inside your dashboard you can review any conversation Sally has had and flag one that wasn’t quite right. That flag isn’t just feedback into the void: it feeds a deliberate loop:

08See what Sally did while you were closed

All this work is no good if you can’t see it. Insights turns Sally’s quiet, around-the-clock effort into plain numbers you can read in a minute.

Open Insights in your dashboard and the first thing you see is the part that’s easy to miss: the bookings Sally captured after hours, while you were closed, with an estimate of what they’re worth. Below that is an honest read on whether you’re actually growing, and a short list of where the next bit of money is.

• While you were closed. An estimated dollar value of the bookings Sally took outside your opening hours, plus the chats she handled, the ones she booked, and the reminders and review requests she sent on your behalf.
• How you’re growing. Bookings, estimated booking value, new versus returning clients and your no-show rate, each compared against the period before so you can see the direction of travel.
• Where the opportunity is. Your busy and quiet windows across the week, the regulars who’ve gone quiet and are worth winning back, and your top services.

The money figures are estimated from your service prices, so the closer your prices and opening hours are to reality, the sharper the numbers. It’s on every plan, with nothing to set up. See how Insights works →

Sceptical was the right place to start. Now try her yourself.

The fastest way to trust her is to watch her answer for your business — set up free in minutes, no card, and grill her yourself. Questions first? Email hello@suddenly.co.nz and a founder reads it.