A flower shop off Bree Street in the Cape Town CBD, half past two on a Thursday in late October, three days before the busiest wedding weekend of the season. The cold room is full. On the counter sits a Multiflora invoice for 480 stems landing at four the next morning, a Netflorist Same-Day queue on the desktop showing eleven pending, and a WhatsApp Business number with forty-three unread. One of those is from a bride asking whether the shop does installations at Molenvliet for a February wedding. One is a sympathy arrangement for a funeral in Bellville tomorrow morning, sent by a family friend who has the wrong crematorium and a bank reference she cannot find. Three are from a Sandton PR agency wanting to move a standing corporate delivery from Wednesdays to Tuesdays. The head florist herself is in the cold room, up to her elbows in a hydrangea-and-protea order for a matric-ball centrepiece the boyfriend is collecting at four.
That is what a working Cape Town flower shop looks like in 2026. Two florists, a driver on the road in a small bakkie, a Yoco Neo Touch on the counter, a Xero login somewhere on the laptop the assistant borrows to run stock. Where AI belongs is not on the bench and not in the cold room. It belongs on the WhatsApp thread and in the quote inbox, the two places that keep pulling the head florist off the flowers.
Where the hours actually go in a small SA flower shop
Break the week down honestly. There is stem work: conditioning, chopping, wiring, arrangement, install, breakdown. Nobody wants a bot near that. There is buying: the pre-dawn drive to Multiflora Cape Town on Auction Road, or the Flower Union delivery arriving in bakkie two. Nobody wants a bot near that either. And then there is admin.
Admin looks like this. Enquiries land on WhatsApp all day. Some are two-line “how much for a bouquet delivered to Rondebosch?” queries that need a five-line reply, a picture, and a Yoco payment link. Some are sympathy conversations dressed up as WhatsApps, arranging tributes for a service the next morning with the wrong crematorium named and a bank reference the sender cannot find. Some are corporate account requests that need onboarding, invoice terms and a purchase order chase through accounts payable. Some are brides.
Brides deserve their own paragraph. A bride enquiry is a long conversation, six months of it, that starts on a Pinterest board and ends with the head florist standing on a step ladder at Nooitgedacht at six on a Saturday morning with two industrial buckets of ranunculus and no phone signal.
None of that gets faster by hiring another florist. It gets faster by keeping the two florists behind the counter and on the bench, and letting a bot handle the first ninety seconds of every message.
The enquiry queue: what AI drafts and what the florist sends
The single most useful pattern in a small SA flower shop is a WhatsApp-first drafting system, connected to the shop's number, that writes a first reply the florist reads and edits before it goes out. Not autopilot. Draft-and-send.
A message comes in. The system has already read it, categorised it (sympathy, wedding, corporate, walk-up gift, complaint, delivery query), pulled the customer's history if there is any, and written a response in the shop's voice. The florist glances, adjusts a word, hits send. Ninety seconds instead of nine minutes.
Two things matter about the draft. First, it names a real price band, not “let me get back to you”. A R550 mixed bouquet with a card in Rondebosch is not a mystery; the model has read the last hundred quotes and knows the ranges. Second, it never quotes a bespoke wedding job. The draft for a wedding enquiry reads something like “thank you, we book weddings six to nine months out, here are three sample installations we did in the Winelands, and I will call you between ten and eleven tomorrow to walk through your day.” The florist keeps the call. The bot keeps the acknowledgement.
Sympathy is the one category where the draft must never send itself, even after edit. It routes to the owner's phone with a suggested three-line reply and a flag that says “call within the hour”.
Flowers for a funeral are not a Yoco link.
The delivery-day loop: the quietest AI win in the week
Deliveries chew hours. Not the driving. The coordination. The recipient not answering the buzzer. The gate code that was never sent. The address in Pinelands versus Pinehurst. The photograph the sender wants of the arrangement on the recipient's desk to prove it landed on the birthday, on the day, in the office they said they would be in.
An AI-drafted delivery loop, running against the day's route sheet, handles this in three moves. Two hours before the drop it sends the recipient a WhatsApp asking for a gate code, a preferred window, and a person to call if the driver cannot get in. It pings the sender when the driver leaves the shop and again when the arrangement lands. And it attaches the photograph the driver takes on arrival, with the recipient's face out of frame if there is a person in the shot, because the recipient never consented to a photograph.
That last detail matters. POPIA does not exempt florists. A “here is her birthday face” video sent to a corporate sender is not compliant, however well-meant. Bake the boundary into the automation and it stops being a case-by-case judgement call at four in the afternoon.
Wedding quote scaffolding: where the human stays and the paperwork speeds up
Wedding work is the highest-margin corner of a small florist. It is also the slowest to quote. A florist can spend three hours building a proposal for a bride who ends up going with a different vendor, and it hurts every time.
AI does not replace the wedding conversation. It shortens the paperwork wrap-up afterwards. After the bride's call, the florist dictates a two-minute voice note into the shop's phone: bride's name, venue, date, palette, arch, aisle markers, four bridesmaid posies, twelve reception centrepieces, one bridal bouquet, breakdown at 23:00. The system builds a first-draft proposal from the note: line items, indicative pricing per stem group, delivery and install costs, breakdown block, POPIA-compliant photo release clause, a Yoco deposit link and a plain-language T&Cs footer.
The florist rewrites it before it goes out. Not because the draft is bad. Because a bride is a person with a specific taste and a specific budget and a specific mother, and the closing paragraphs are what wins the booking. The draft gets the numbers right. The florist gets the tone right. The proposal lands inside four hours instead of four days.
In my experience, faster proposals with cleaner numbers win a meaningful share of the enquiries a shop currently loses in the follow-up gap. Not all of them. Some brides were shopping around and were never going to book you. The ones who wanted a fast, professional reply from a florist who sounded present, you catch.
Corporate accounts: a quiet retention layer
Most SA flower shops carry a handful of standing corporate orders. The weekly reception arrangement at a Cape Town law firm. The monthly boardroom pieces at a Bryanston asset manager. Sympathy arrangements from a Sandton insurance broker to bereaved client families. The revenue is steady, the admin is annoying, invoicing runs through Xero, POs arrive as email attachments, and the person who signs off the flowers is on maternity leave from June.
A quiet retention layer keeps these accounts warm without a sales call. Two weeks before the recurring delivery, a scheduled WhatsApp or email asks whether the palette should stay the same or shift with the season. Photograph of last month's arrangement attached, for context. A “confirm as-is” button that takes ten seconds. An invoice generated in Xero the moment the driver leaves the shop, with the correct PO field, so accounts payable at the client never has a reason to hold payment past thirty days.
None of that requires AI. But AI is what handles the exceptions: the “actually can we skip May, the boss is on leave?” reply, the “different budget for the AGM this quarter” negotiation, the “our office moves in July, here is the new address” change. Draft, review, send. Same pattern as the enquiry queue. Same discipline about what routes to the owner.
Where AI must not go in a flower shop
A short, honest list of what the model must stay out of.
The sympathy conversation itself. Someone is arranging flowers for a father who died on Sunday. A bot reply, however well-worded, is a small betrayal. Route the message to a human every time, and let the florist call.
The bridal consultation. A first-draft proposal is paperwork. A first call with a bride is not. The relationship you build in that thirty-minute conversation is what earns the referral to the sister-in-law two summers later.
Complaints. The bunch that arrived wilted. The arrangement the recipient described as sad. The wedding centrepiece that toppled at 22:14 during the speeches. Those are owner-level conversations. The bot can flag the message and pull the order history. It cannot write the apology.
The flower knowledge itself. A model will happily invent that peonies are seasonal in Cape Town in July. They are not. If a customer asks what is in season, the honest answer comes from the florist who was at Multiflora at four this morning, not from a language model that has never held a stem.
Where to start
If you run a two-to-four-florist SA shop and want to see whether any of this earns its keep in your week, the smallest useful project is the WhatsApp draft-reply loop on the shop number. Not the delivery photos, not the wedding quotes, not the corporate retention feed. Just the ninety-second first reply on the enquiries that already land in your phone every day.
Give it a fortnight. Measure two things: how much time the florist gets back at the bench, and whether the enquiry-to-quote ratio moves. If both go the right way, and in most working shops they do, you have a foundation on which to build the delivery loop and the wedding scaffolding. If they do not, you have learned something honest about your shop for the cost of two weeks of experimentation, which is a fair trade.
The florists in Cape Town and around the country who get real value from this almost always start narrow and stay honest about what the tool does. It writes the first ninety seconds of a reply. It runs the delivery loop. It scaffolds the paperwork. Everything else — the stem, the recipe, the ladder at Nooitgedacht at six on a Saturday morning — is still yours.