HannaAI

Ask your business a question. Get an answer.

Business Insight is a reporting layer you talk to in plain English. Every figure is computed from your own records against one fixed definition — and when something genuinely can't be measured, Hanna says so instead of estimating.

One definition of every figure Computed, never generated Declines rather than guesses

Six questions · your week

I ask. It answers — or tells me it can't.

Plain English, on the till or your phone. Every answer carries the basis it was computed from.

YouWhat did we take last week, and what was the average spend?

£17,940.15

Across 412 covers, an average of £43.55 a head. Ahead of the week before by 4.1%.

Measured · net of tax and service · ties to the daily close

YouHow many walk-ins did we turn away on Saturday?

I can't answer that yet.

Turned-away walk-ins aren't recorded unless someone logs them at the door, so any number I gave you would be invented. I can tell you Saturday ran at 104 of 104 covers with a 22-minute average wait — which is probably the question underneath.

Not measurable · will not estimate

One definition of every figure, everywhere

Replaces: a dashboard you never have time to read

Monday tells me what changed, before I ask.

A like-for-like comparison, run before you're in. Nothing material happened? It says nothing.

Thursday covers · against a day-of-week baseline

Base65
Wk 158
Wk 257
Wk 358

What Hanna checked outside your four walls

School half-term · 2 of 3 Weather · normal Bank holiday · none Local events · none

Thursday is down 11.4%, three weeks running.

Every other night is level or ahead. Two of the three fell inside half-term — a candidate cause, not a conclusion. Worth a look before you change anything.

Metric: covers_served · vs day-of-week baseline · Confidence: medium

Replaces: finding out at month end

I can see where my covers actually come from.

Channel, lead time, party size and who comes back — the shape of demand before it arrives.

78%Booked direct through Hanna
41%Covers from returning guests
3.2 daysMedian booking lead time
1.8%No-show rate, from 4.1%

Party size mix · last 30 days

252%
3–434%
5–610%
7+4%

Deposits cut no-shows by more than half. Two-tops fill late — holding four back until 48 hours out is worth roughly 9 covers a week.

Replaces: guessing which channel works

I can see which nights carry the business.

Four weeks of actuals, and what next week is expected to do.

Average covers by day · last four weeks

Mon48
Tue52
Wed61
Thu58
Fri96
Sat104
Sun71

Next week, expected

Tue46
Wed55
Fri97

Friday and Saturday carry 58% of the week's revenue. The capacity — and the opportunity — sits Monday to Thursday.

Replaces: gut feel about which nights matter

I roster to what's coming, not to last week.

The forecast sets the demand. The clock sets the cost. Both sit next to your target before you publish.

Next week · 486 covers, £21,180HoursCostOf revenue
Rota as drafted372£8,55640.4%
Your target£7,20134.0%
Hanna suggests−14h Tuesday, −8h Wednesday. Friday and Saturday untouched.350£8,05038.0%
Labour as a share of revenue40.4%
0%Target 34%50%

Saves £506 next week. The 4 points still above target are salaried roles — structural, not rota. Worth a conversation, not a cut.

Replaces: copying last week's rota

I know exactly which night needs filling.

Effort spent on a night that will fill itself is wasted. Hanna points it at the gap.

Next week · forecast against capacity

Tue32 free
Wed23 free
Frifull
Satfull

The audience that fits the gap

43Regulars who come midweek
118Lapsed 60–90 days
£435–610Expected in return

Nothing to Friday or Saturday guests — those nights are full, and a discount there would only cost you money. Hanna will never suggest one.

Replaces: emailing the whole list and hoping

Why it can be trusted

Hanna does not do arithmetic.

This is the most important sentence on the page. The language model never calculates anything. A reporting engine computes every figure against your database using one fixed definition per measure, and Hanna is handed the finished number to put into a sentence.

01 · The question

You ask in your own words

"What did we take last week?" Hanna's only job here is to work out which defined measure you mean — revenue, net of tax and service, for a date range.

02 · The engine

A report is run, not written

The same code that produces your daily close computes the figure, from the same source events, with the same filters. There is one definition of revenue in the whole platform and this is it.

03 · The answer

Hanna narrates the result

It receives £17,940.15 and 412 covers and writes the sentence. It cannot invent a total, because it was never asked to produce one.

So the failure mode everyone rightly fears from an AI — a confident, plausible, wrong number — is designed out rather than tested for. If a question maps to no defined measure, there is nothing to narrate, and Hanna tells you that instead.

What it won't do

A shorter list of answers is worth more than a longer one.

Every measure in Hanna is either defined and computable or explicitly declined. Nothing sits in between, and the declined list is published inside the product rather than hidden behind a shrug.

Answers, computed from your records

Revenue, covers, average spendMeasured
Cost of goods and gross marginMeasured
Stock value and supplier spendMeasured
Labour hours and costMeasured
Booking channel, lead time, no-showsMeasured
Dish mix, margin and pairingMeasured
Returning and lapsed guestsMeasured

Declined, and why

Walk-ins turned away
Not recorded unless logged at the door
No data
Why a guest didn't return
Unknowable without asking them
No data
Whether service felt slow
Timings exist; the feeling doesn't
No data
Competitor performance
Not ours to know or guess at
Out of scope
What a dish "should" be priced at
A judgement, not a measure
Yours

The second column is why the first one is worth reading. Any system can produce a number for anything. Only a system that refuses is telling you when it knows.

Outside your four walls

Your trading doesn't happen in a vacuum.

A quiet Thursday means something different in half-term than it does in October. Hanna reads a small number of external signals alongside your own history, and uses them to qualify a finding rather than to explain it away.

School terms

Term and holiday dates for England and Wales, which move demand in almost every neighbourhood restaurant and move it differently by area.

Bank holidays

The official calendar, including the ones that fall differently in Scotland and Northern Ireland.

Weather

Observed conditions for your postcode by day. Rain on a Tuesday is a candidate explanation; a wet fortnight is a pattern.

Your own seasonality

The strongest signal of the four, and the one that improves most with time. Last year in your room beats any general rule about restaurants.

Day-of-week baselines

Every comparison is like-for-like. A Thursday is only ever compared with Thursdays.

Stated as candidates

Hanna says "two of the three fell inside half-term — a candidate cause, not a conclusion." It will not tell you the weather cost you £900.

Where this is going

From explaining to expecting.

Told honestly, because a forecast you can't trust is worse than no forecast. Here is what is live now and what isn't.

Live

Answering

Ask anything that maps to a defined measure and get it in seconds, with its basis. Working in daily service since 2024.

Live

Explaining

Monday findings on a like-for-like basis, with external context and a stated confidence. Silent when nothing material moved.

Partly

Expecting

Next week's covers, drafted orders and labour suggestions are built from your own trading history and the external signals above. They are good enough to plan against and not yet good enough to publish a rota unread.

Not yet

Learning across venues

Patterns that hold across many rooms — how a cold snap moves a counter differently from a dining room — need many rooms. We will build it when we have them, and we will say so when we do rather than before.

In service

Answering questions tonight, in three London venues.

The same definitions across all three, which is what makes them comparable at all. Each name here will take your call.

Hotori Yakitori & Edomae sushi · London EC4ALive since 2024
Charcoal Champ Chinese barbecue · London N1Live since 2026
Brass Bar Cocktail bar · LondonLive since 2026

Before you ask

The questions operators actually ask.

Can the AI make up a number?

No, because it never does the arithmetic. Every figure is computed by a reporting engine against your database using one fixed definition per measure. The language model receives the finished figure and puts it into a sentence. It cannot invent a total, because it was never given the job of working one out. If a question maps to no defined measure, there is nothing for it to narrate and it says so.

How is this different from a dashboard?

A dashboard shows you what somebody decided in advance you would want to see, and leaves the thinking to you. Hanna answers the question you actually asked, in the words you asked it, and tells you what it means. The difference matters most on a Tuesday morning when you have four minutes and a specific worry, which is exactly when nobody opens a dashboard.

Why can't it answer everything?

Because some things are not recorded anywhere. Walk-ins you turned away, a table that felt slow, why a guest didn't come back — none of that exists as data unless somebody captured it. Hanna could produce a plausible number for any of them and it would be fiction. Declining is not a limitation we are hiding; it is the feature that makes the answers you do get worth acting on.

Does it work with my accountant's figures?

It should reconcile, because it is built from the same source events — your sales, your invoices, your clock-ins. Where it differs from a set of management accounts it is usually timing: Hanna accrues to the day the trading happened, and bookkeeping often posts to the day the paperwork arrived. Both are correct; they are answering slightly different questions.

How much history does it need?

It answers questions about what happened from the first day of trading. Comparisons need something to compare against, so like-for-like findings become useful after about six weeks and genuinely reliable after a full seasonal cycle. Forecasting is the slowest to earn trust, and we would rather tell you that than have you act on a projection built from three weeks of data.

Book a demo

Try to catch it out.

Bring the questions you'd actually ask on a bad Tuesday, including the ones you suspect it can't answer. Watching it decline is more informative than watching it perform.

Book a demo