All docs pages

AI in Miseline — every assist, explained

Miseline uses AI as an assistant, never an autopilot. Every AI feature follows the same rules:

  • A person triggers it or reviews it. Assists marked ✦ run when someone

clicks them. The few automatic analyses (noted below) only draft or flag — they never change operational records on their own, and anything AI drafts waits for a person to accept it.

  • Every call is metered. Each AI call writes an accounting record with

the acting user, the feature used, the model, input/output token counts, and whether it ran on the platform key or your own key. Admins see the roll-up in Billing & plan → AI usage (by user, by feature, 7/30/90 days) — usage tracking is built in at the user and token level, not bolted on at a gateway.

  • Budgets are enforced. On the platform key, each plan includes a monthly

token pool (Starter 200k · Pro 2M · Enterprise 10M). When it's used up, AI assists pause until the monthly reset (or an upgrade) — manual workflows are never blocked. Bring-your-own-key usage is logged but never counted against the plan budget.

  • Your key, your custody (Enterprise). Organizations can supply their own

API key (Billing & plan → AI settings) and, separately, their own speech-to-text endpoint (Settings → Interview & huddle transcription) so audio and prompts stay on infrastructure they choose.

  • Nothing breaks without AI. If AI isn't configured or the budget is

reached, every screen keeps a fully manual path and says so quietly.

Where AI appears, feature by feature

Culinary & nutrition

  • Recipe import ✦ — paste any recipe text; AI structures ingredients,

steps, and yields for your review before saving.

  • Nutrition estimate ✦ — estimates an ingredient's nutrition when no

database match exists; clearly labeled as an estimate.

  • Menu nutrition narrative ✦ — plain-language summary of how a menu

scores against the population targets you assigned.

  • Menu engineering advice ✦ — commentary on the popularity/margin

quadrant analysis.

Operations

  • Huddle summary — after a recorded huddle is transcribed, AI drafts the

TL;DR, wins, issues, and action items; the manager reviews before publishing, and extracted actions become tasks only on publish.

  • Preventive maintenance insights ✦ — risk-ranks equipment from its

event and work-order history; creating a task from an insight is a click, never automatic.

  • Work-order triage ✦ — drafts severity/assignment context from the

asset's history when a work order is opened.

  • Cost variance commentary ✦ — root-cause narrative on

theoretical-vs-actual usage variance.

Audits, compliance & review programs

  • Template drafting ✦ — describe an audit or survey in a sentence; AI

drafts the template for the builder, where you edit and publish.

  • Mock survey generation — builds a draft audit from a compliance

framework's controls (deterministic, not generative — listed here because people expect it on this page).

  • Inspection summary — a short AI summary attached when an audit is

submitted.

  • Interview question guides ✦ — drafts a role-aware question guide for a

scheduled interview.

  • Interview & walkthrough processing ✦ — turns a transcript into draft

findings, each categorized against the maturity model with a supporting quote. Every finding waits in a validation queue for a person to accept, edit, or dismiss.

  • Survey digest ✦ — aggregates survey responses into the same

draft-findings queue.

  • Maturity rating suggestions ✦ — proposes 1–5 ratings from accepted

findings into a separate draft column; reviewers apply them per category.

  • Goal suggestions ✦ — proposes improvement goals from a finalized

assessment's gaps; you choose which to create.

  • Evidence synthesis & engagement drafting ✦ — assembles engagement

evidence into workpapers, then drafts findings/recommendations for review before anything is finalized.

  • Document extraction ✦ — reads a photographed certificate, SDS, or

service record and prefills the record for confirmation.

Clinical

  • Diet-order parsing ✦ — converts free-text diet orders ("puree, nectar

thick, no citrus") into structured orders for RD verification.

  • Intake risk summary — daily nutrition-risk triage narrative from

intake/refusal/weight data (runs with the daily checks; flags only).

Front of house & catering

  • Catering quote drafting ✦ — drafts a priced quote from a client brief

using your real recipe costs.

Reporting & the org agent

  • Report narratives — operations- and board-audience narratives on

scheduled site/account reports.

  • Executive insights ✦ — fleet-level analysis and recommendations on the

operator dashboard.

  • Org agent chat ✦ — your named assistant (each org can rename it) that

answers from live operational context; available on web and the staff app.

Speech to text

  • Interview, walkthrough, and huddle transcription — audio recorded in

the apps is transcribed through the configured speech-to-text endpoint (platform-provided, or your own per-organization endpoint). Transcripts are editable; you can also paste a transcript manually and skip audio entirely.

The accounting model (for administrators)

Every call above lands one metering record: who (the signed-in user that triggered it, or the system identity for scheduled analyses), what (feature key, e.g. interview_process), model, tokens (input + output), and key mode (platform or your own key). From Billing & plan → AI usage you can answer, at any time: who is using AI, on which features, how much of the monthly pool is consumed, and what ran on your own key. Failed or unavailable calls are recorded too, so gaps are visible rather than silent. The same records back the plan budget enforcement — the numbers you see are the numbers that gate spending.