Weighed and scooped billing, a topping engine that prices every mix-in, and a display freezer that counts itself down — 12 dessert solutions, narrated by Iris.
Every tub tracked by batch and best-before, hardening times respected, and Iris flagging the flavour that will not survive the weekend.
Dessert margins melt fast — literally. Freezer shrink, over-scooping, weekend batches sized by memory.
Scoop and gram billing at the till, and pricing that knows what a topping actually costs — not a flat charge that averages the margin away.
Inventory live by batch and best-before, hardening times respected, and the flavour that won't survive the weekend flagged early.
Saturday at eight is a different business from Tuesday at three. Production, staffing and the menu move with the queue, not against it.
Scoop yield, freezer shrink, batch cost and weekend demand — live, with Iris naming where the margin melted.
Orders, mix-ins, weighed lines and any tender on one screen — while Iris suggests the right add-on and flags margin drift without adding a second to the queue.
The scale is part of the till. A weighed line prices itself — no estimating by eye, no rounding a gram-level product to the nearest guess.
Every mix-in carries its cost. Toppings price individually rather than as a flat charge that averages the margin away.
Iris upsells at the counter. The add-on, the take-home pack, the gift box — surfaced on the till, timed to what's already in the order.
Scoop yield, freezer shrink, batch cost and weekend demand on one live board — with Iris naming the cause behind the number that moved.
Every module feeds one board. Till, freezer, batch production and workforce all post into the same financial view.
Forecasting, not just reporting. L2 models project where the weekend is heading while the batch plan can still change.
Iris names where it melted. Not that shrink rose — which freezer, which flavour, which shift.
P&L, balance sheet, cash and close on a single balanced ledger — sub-ledgers that reconcile, and a month-end that is a review rather than a rebuild.
Weighed sales post cleanly. A gram-priced line reconciles to the ledger exactly as a unit-priced one does.
Shrink is costed, not estimated. Freezer loss posts at live batch cost rather than as a year-end adjustment.
Ask Iris for any number. Any outlet, any period — answered in the conversation, not a week after close.
Forecast to batch plan to purchase order to the cold-chain van — demand-planned replenishment across every outlet, so the right stock lands before you need it.
The weekend curve drives the order. Forecast demand becomes batch quantities and purchase orders, not a memory of last Saturday.
Cold chain is part of the plan. The van is planned alongside the batch, because in this sector transport is where product is lost.
Feeds the freezer ledger. Receipts and transfers post to Display Freezer Inventory directly, so shrink is caught same-day.
Digital checklists, shift routines and standard operating procedures across every outlet — completion tracked live, exceptions escalated.
Hardening and tempering, signed off. The routines that decide product quality are completed in the system, not assumed.
Temperature logs write themselves. Probe readings roll up by unit, by site, ready for inspection rather than reconstructed.
Exceptions escalate themselves. The routine that didn't complete surfaces to a human; the rest stays quiet.
Rostering, attendance and labour percentage measured against forecast demand. Iris flags the overstaffed lulls and the understaffed peaks before the shift, not after.
Saturday at eight is its own shift. Cover is built from the forecast peak, not from an even spread across opening hours.
Compliance is a guardrail. Break rules and maximum hours are enforced in the roster, not audited after payroll.
Flagged before the shift. An overstaffed lull or an understaffed peak is caught while the roster can still change.
Deterministic thresholds under the hood (the “Zen Rules” layer) — a weight is a weight and a temperature is a temperature. No probability, no judgement call. Two solutions run on this layer.
Unit 3 has held above threshold for 11 minutes. Two batches inside, both within best-before. Raised while the stock is still saveable, not at the next manual check.
A gram-level product billed at gram-level accuracy. The scale talks to the till, the scoop count carries its own price curve, and nothing is estimated by eye at the counter.
A connected scale sends served weight straight to the POS — no manual entry, no rounding a priced-by-gram product to the nearest guess.
Scoop-count pricing runs on its own curve, so a second and third scoop are priced deliberately rather than as simple multiples.
Container and packaging tare is deducted automatically — the guest pays for product, not for the tub.
Serve weight is compared to spec on every line, so over-portioning is visible as a pattern rather than felt at stocktake.
Every weighed transaction is logged, giving the audit trail that gram-level pricing needs.
Every batch freezer, blast chiller and display unit monitored continuously by probe. A breach alerts while the stock is still saveable — and the compliance log writes itself.
Each unit reports its own temperature on interval — there is no clipboard, and no blind gap between manual checks.
A threshold breach alerts immediately, naming the unit, the duration and the batches inside it.
Hardening and tempering windows are tracked as machine state, so product is not served before it is ready or held past where it holds.
A scheduled defrost cycle and a door left open read differently, so routine cycles do not generate false alarms.
A compliant temperature record is generated automatically, ready for inspection rather than reconstructed the night before one.
Forecasting, queue and pricing models under the hood (the “Zen Models” layer) — batch size, peak-hour queue, freezer shrink and markdown depth, predicted before they land. Six solutions run on this layer.
Saturday at eight is a different business from Tuesday at three. The queue is forecast rather than absorbed, and production, staffing and the menu move with it.
Expected queue depth is forecast by hour and by day, so the peak is planned for rather than survived.
A live wait estimate is displayed at the entry, so guests decide whether to join rather than leaving halfway.
Complex builds are the slowest thing at the counter, and the model accounts for order mix rather than treating every order as equal.
Server throughput is tracked, identifying where the counter actually bottlenecks rather than where it is assumed to.
The same forecast drives the roster and the batch plan, so staffing and stock cannot contradict each other.
A display freezer that counts itself down. Every tub tracked by batch and best-before, depleting as the till sells, with shrink surfaced as an anomaly rather than discovered at stocktake.
Each tub is tracked as remaining volume, not as a binary in-stock flag — a tub at 12% reads differently from one at 80%.
Batch and best-before travel with every tub, rotated first-expiry-first-out from the back freezer to the display.
When actual depletion drifts from what the till sold, the gap is raised as an anomaly while the cause is still findable.
Below threshold, back-of-house is told which flavour to bring forward and how long it needs to temper.
A genuine sell-out removes the flavour from the menu board and every ordering channel in the same second.
Frozen product does not travel like a pizza. Each delivery is scored before it is accepted, and the packaging, the dry ice and the serviceable radius follow from that score.
Distance, ambient temperature, product type and rider ETA are scored at acceptance — before the order is on a bike, not after.
Insulated pack tier and dry-ice quantity are prescribed by the score, so the packer is told what to use rather than guessing.
The serviceable radius contracts on a hot afternoon and extends on a cold evening, automatically rather than as a fixed number set once.
Beyond threshold the order is declined at checkout with a reason — one order lost, against a refund, a bad review and the product.
Melt-related refunds are attributed back to the score that allowed the order, so the threshold tunes on real outcomes.
Which flavours earn their freezer slot this season, what a limited-time run should replace, and when to end it — decided on forecast sell-through rather than on affection for a recipe.
Every display slot is scored on forecast sell-through, margin and waste, so the question becomes what a slot earns.
A limited-time run is planned with a start, an expected volume and an end, rather than left to trail off.
A new flavour that would simply take sales from an existing one is flagged before it takes a slot.
When a seasonal favourite leaves, the closest equivalent is suggested to the guests who ordered it rather than a blank space.
Preference clusters by site, so a chain does not run one identical freezer everywhere.
Iris identifies which stock will not clear and recommends the discount that moves it — deep enough to sell, not deeper than necessary, and timed while there are still guests to sell it to.
Items with more remaining than historically sell in the closing window are identified automatically.
A dynamic pricing model trained on past sell-through and markdown outcomes predicts the optimal discount per item, per day.
Batches approaching best-before are prioritised over merely slow ones, because the cost of holding them is not the same.
The manager accepts once and the new price lands on the till, the menu board and every channel together.
Whether the markdown actually cleared the stock is logged per item — the model improves on outcome, not intent.
Weekend batches sized by forecast rather than by memory — with theoretical usage compared to actual, and every kind of loss separated and costed.
Batch quantities come from forecast demand by flavour and day, not from the volume that was made last weekend.
Theoretical usage is compared to what actually left the freezer — the gap is the number that quietly decides food-cost percentage.
Over-portioning, melt loss, best-before discard and spoilage are separated, because the fix for each is completely different.
Iris flags the flavour that will not survive the weekend early enough to promote it, bundle it, or route it into a take-home pack.
Waste by flavour and day-part flows back into the batch plan, so a repeatedly over-churned flavour is corrected rather than repeated.
Not a screenshot — the actual agent. Natural-language answers over the live counter. Four solutions run on this layer.
Every way an order can be built — cone or cup, scoop count, sauces, sprinkles, nuts, fruit, folded mix-ins — captured accurately and priced for what it actually costs.
Each topping and mix-in is priced individually, rather than as a flat charge that averages the expensive ones away.
Vessel choice — cup, cone, waffle, take-home tub — carries its own price and its own packaging cost through to margin.
The combinations guests actually order surface as one-tap presets, which matters most when the queue is at its longest.
Allergen flags travel with every mix-in, and incompatible combinations are blocked rather than left for the server to catch.
Mix-ins deplete the ingredient ledger as they are served, so a topping running low is known before the tub is empty.
Returning guests are recognised at the counter and their usual build — flavour, scoop count, vessel, every mix-in — is reconstructed before they have to describe it again.
Identified by phone, loyalty card or app profile, with history retrieved before the server asks what they want.
The memory holds the whole build, not just the flavour — a dessert order is mostly modifiers, and that is the part guests tire of repeating.
A guest who orders differently on a weekday evening than a weekend afternoon gets the right recall for the occasion.
Group and family orders are remembered as a set, so one buyer ordering four different builds is not four separate conversations.
When a remembered flavour is out or retired, the closest available match is offered rather than an apology.
Points, streaks and a shared family wallet on one ledger across counter, app and web — built for a sector where the buyer and the eaters are often not the same people.
Earned at the counter and redeemed in the app on a single balance, never per-channel silos that confuse guest and staff alike.
A family wallet lets a household share one balance, because one person usually pays for four desserts.
Streaks reward the consecutive-visit habit that drives this sector, rather than only rewarding total spend.
Iris picks the offer with the highest modelled redemption for that guest, instead of a blanket discount that costs margin on people who were coming anyway.
Outstanding points are valued live as a liability and posted to finance, so loyalty never becomes an unbudgeted surprise.
The take-home tub, the gift box, the celebration cake — the highest-value part of the basket, prompted at the right moment and packed so it survives the journey.
Iris prompts the take-home pack or gift box against what is already in the order, not as a scripted question on every ticket.
Packaging tier is matched to journey time, so a tub that has forty minutes to travel is not packed like one going next door.
Celebration and cake orders are captured with lead time, decoration notes and a firm slot, then enter the batch plan rather than a notepad.
Occasion-aware: a birthday order prompts the candle, the box and the message card as a set.
Packaging and dry-ice cost are carried into the line, so gifting margin is real rather than assumed.
No call to book, no deck to sit through — the till, the label printer, the scale, the guest's phone, running on live data and narrated by Iris.
You configure a sector playbook, not a custom project. Iris applies it herself — agentically, from day one.
Onboarding is agentic. Iris connects the till, the scale and every freezer probe directly — no manual data mapping.
The playbook applies itself. Sector thresholds go live immediately, then tune from real trading data, site by site.
Live in weeks, learning from day one. There's no bespoke build to wait on — go-live is a configuration, not a project plan.
The first line of support is agentic — Iris resolves most of it herself. Our engineers pick up from there.
Layer 1 — Iris, 24/7. Configuration questions, anomalies and routine issues resolved directly, instantly.
Layer 2 — our engineers. Anything Iris can't close escalates automatically to a Zentallio engineer.
No blank tickets. Every escalation arrives with Iris's own diagnosis — engineers start from an answer.
“Dessert margins melt fast — literally. Freezer shrink, over-scooping, weekend batches sized by memory: gram-level products with guess-level control.”
Zentallio weighs everything: scoop and gram billing at the till, freezer inventory live, batch production tuned to the weekend curve, and pricing that knows what a topping actually costs. Underneath sits the same platform every sector runs — one data spine from the till to the ledger, three layers of intelligence, and an agent that narrates every screen, flags what needs a human, and acts on the rest.
Over-scooping — a few grams a serve, invisible at 1,840 scoops until it compounds.
Freezer shrink nobody sees until stocktake, and a flavour that won't survive the weekend.
Weekend batches sized by memory, so Saturday runs out and Tuesday gets discarded.
The scale prices every serve and every probe logs itself — weight and temperature are facts, not estimates.
Batch size, queue depth and markdown depth are forecast before the weekend, not reviewed after it.
“Where did the margin melt this week?” — answered from the same ledger the batch is costed from.
Every Zentallio sector runs the same three-layer intelligence, tuned to how that format actually loses margin.