AI weekly digest emails: what 30 days of order data should tell you

Most independent owners don't open their analytics dashboard unless something feels off. That's rational — the dashboard is full of charts that may or may not move the needle, and pulling signal out of them takes more time than the average Monday morning allows.

RestoWebMaker turns that on its head: instead of waiting for the owner to come to the dashboard, every Monday at 8am we email a 2-3 sentence narrative of the past 30 days. The math is rolled up on our side. The owner gets it on their phone with their coffee. If the email surfaces something interesting, they tap through to the dashboard. If not, they delete it and get on with their day.

What goes into the digest

The cron job runs every Monday at 08:00 in the restaurant's local timezone (or America/New_York if no TZ is set). For each active restaurant, it computes:

• Total revenue for the past 7 days vs. the prior 7 days (delta and %)
• Same comparison for order count
• Same comparison for average order value
• The top 3 menu items by units sold last 7 days
• Repeat-guest percentage — how many last-7-day orders were placed by guests who'd ordered before
• Any unusual day-of-week or time-of-day patterns vs. the 30-day baseline

That structured data is the prompt input. The output is a 2-3 sentence narrative — not a chart, not a table.

Why narrative beats numbers

Owners can read a sentence in 4 seconds. They can't read a chart in 4 seconds. The whole reason this email exists is to deliver a signal in less time than the owner spends deciding whether to open it.

Compare:

Revenue: $14,238 (-3.2%) · Orders: 287 (+1.4%) · AOV: $49.61 (-4.5%)

vs.

Revenue is down 3% on slightly higher order count — your average ticket dropped $2.30 last week, mostly from a quiet Friday lunch. Wood-fired pizza and the seasonal beet salad are still your top two sellers; the new ramen is climbing.

Same data, different speed of comprehension. The first version leaves the owner to do the synthesis. The second hands it over.

The prompt and the model

The model is Haiku at claude-haiku-4-5. The prompt is roughly:

You are summarizing one restaurant's order data for the owner.
Below is the structured data for the past 7 days and the prior 7
days. Write 2-3 sentences in plain English.

Rules:
- Lead with the most important number or trend.
- Mention specific items and specific days/times.
- Don't hedge ("seems to suggest", "may indicate").
- Don't recommend actions unless the data strongly suggests one.
- 2-3 sentences. No bullet lists. No headers.

[structured data block]

Haiku at this size of context costs roughly $0.0001 per restaurant per week — so a thousand-restaurant run is ten cents. Even with retries and growth headroom we're talking dollars per month at the platform level for a feature that drives meaningful weekly engagement.

Caching

The result is cached on the Restaurant.aiDigest column with a 24-hour TTL. If the owner opens the dashboard mid-week, the digest panel shows the cached summary rather than calling Haiku again. The cron writes a fresh one each Monday and the cache invalidates itself.

Failure modes

Empty week

New restaurants with fewer than 5 orders in the trailing 7 days skip the AI call entirely. The email instead says something like "You had 3 orders last week — too thin to draw conclusions. Tell us when you start picking up volume." This avoids the worst output: a confident-sounding sentence about meaningless sample size.

Anthropic API down

If the API call fails, the email still goes out — just with a deterministic template-rendered version ("Revenue this week: $X. Top items: A, B, C."). Less elegant, but the owner still sees their numbers. The only thing that breaks gracefully is the narrative; the data delivery is rock-solid.

Currency / locale

Restaurants in non-USD currencies get the same digest with their local currency formatting. The prompt receives the currency code and respects it. We've tested with USD, GBP, EUR, and CAD; non-Latin scripts (e.g., for Japanese restaurants) work because the prompt accepts UTF-8 and Haiku passes it through.

What this is not

It is not a daily email — that would be over-promising. Most independents don't have day-over-day signal worth narrating. Weekly is the cadence where the data smooths into something actionable.

It is not a recommendation engine. The prompt explicitly tells Haiku not to recommend actions unless the data strongly suggests one. Owners can pattern-match for themselves; an LLM telling a chef "consider running a Friday lunch promotion" is condescending.

And it is not a replacement for the dashboard. It's a hook — a way to surface the moments where a deeper look is worth taking. The dashboard is still the place where you actually make decisions.