smengo

AI assistant

Ask about shifts in plain words — the answer comes from your data

The assistant opens from the header on any page. It doesn't paraphrase the internet: for every question it opens the same tables you do, and only within the limits your role allows.

  • Answers from live organization data, not from a training set
  • Tools are assembled from the asker's permissions — it won't see another department
  • It can propose putting someone on a shift, but only your click schedules it
Start freePricing and credits14 days free · one credit per question

Copilot

who works tomorrow?

Tomorrow, August 12, there are 7 people on shift. Floor, 09:00–18:00: Marina K., Oleg D., Irina S. Kitchen, 08:00–17:00: Artem V., Nikita L. Delivery, 12:00–21:00: Sasha P., Dina Zh. Department minimums for tomorrow are covered.

Looked at: day schedule

AssistantFrom documents
For example: who works tomorrow?
Send

How it works

Five steps between question and answer, none of them blind

The assistant doesn't write database queries. It picks from a closed catalog of tools, each of which answers exactly one question and demands its own permission.

  1. 1Question

    You type in plain text

    Up to 500 characters, no commands or syntax. The panel opens on any page of the app and doesn't pull you away from the one you're working on.

  2. 2Catalog

    The tool list is assembled for you

    The model doesn't see all ten tools — only the ones your role has permission for. What isn't in the list doesn't exist as far as it's concerned.

  3. 3Loop

    The model calls tools, up to five steps

    Ask about one person's hours and they're found in the employee list first, then the hours are counted. Every call is re-checked against permissions, this time on the server.

  4. 4Answer

    Text plus a "looked at" line

    The answer arrives as plain text, and under it — exactly which tools were opened. You can see where the number came from and where to go verify it.

  5. 5Billing

    One credit per question

    Charged once per message, no matter how many tools the assistant opened. Credits are shared with the schedule generator and briefings.

Why not "let the model write the query" The assistant has physically nothing to compose a database call with that isn't in the catalog: every tool has a fixed set of parameters. It's the same constraint that keeps an "export everything" button out of the app.

Boundaries

The assistant knows no more than your role does

Permissions are checked three times: the tool list is built from them, every call is re-checked on the server, and row-level access in the database stands last. You'd have to get past all three at once.

One questionhow many late arrivals were there in July?

Manager

has the employees permission

In July there were 9 late arrivals across 5 people, 214 minutes past schedule in total — with the 5-minute grace period applied. Oleg D. leads: 3 late arrivals, 74 minutes.

Looked at: late arrivals

Shift lead

schedule viewing only

I can't answer that: late-arrival records aren't available to you. I can show who worked in July and how many hours it came to.

The "late arrivals" tool never made the list — no permission

Conversation history is personal Only the author sees their conversation with the assistant — not even the organization owner opens someone else's. The answers in it were assembled under the asker's permissions, and showing them to another person is the same as opening the underlying data to them.

Catalog

Ten tools, each behind its own permission

Eight read, two prepare an action. The list is closed: a new tool is an entry in the catalog, not freedom for the model.

ToolWhat people ask in wordsPermission
get_schedule_dayWho works tomorrow or on July 15: people, departments, times, statusesview_grid
get_coverageWhere the gaps are this week: monthly coverage against department minimumsview_grid
get_employee_hoursHow many hours Anna has in July; who works the mostview_grid
get_availabilityWho can't work weekends; who wants extra shiftsview_grid
get_open_shiftsWhich shifts are open right now; are there swap requestsmanage_shifts
get_latenessWho was late in July and by how many minutescrud_employees
get_demandWhat demand is expected; how many people are needed on the weekendmanage_shifts
get_budget_statusWhether July stayed within the labor budgetmanage_labor_budget
propose_shiftactionPut Anna on July 15; fill the gap with Peteredit_schedule
propose_open_shiftactionPost a floor shift for tomorrowmanage_shifts

The catalog grows by entries, not by permissions A tool is a name, a description, a parameter list and a mandatory permission. Until such an entry exists, the assistant cannot reach the data — however confidently it might claim otherwise.

Actions

It can propose. Only you can do

The assistant never puts people on shifts itself. It prepares a card with a ready decision, and the schedule entry appears the moment you press Execute — under your session and your permissions.

Marina K. → August 14, floor, 09:00–18:00

ExecuteReject

The card waits for your decision. Until the button is pressed, nothing in the schedule has changed.

Artem V. → August 16, kitchen, 08:00–17:00

Executed

At the moment of the press, conditions are re-checked: the cell is free, the labor budget isn't exceeded, the permission still holds.

Nikita L. → August 17, floor, 09:00–18:00

ExecuteReject

There is already an entry for this employee on that date.

While the card waited, the shift was filled by hand. The assistant doesn't overwrite someone else's work — the buttons stay, the call is yours.

  • Execution runs under the session of whoever pressed — not under a service account
  • Budget and cell occupancy are re-validated at press time, not at proposal time
  • A double press doesn't double anything: the card is claimed exactly once
  • Every proposal is a journal row: what was proposed, who decided, how it ended

Two more entrances

Answers from your own policies, and a briefing by eight in the morning

The same conversation answers about more than the schedule: a toggle under the input switches the source, and a button assembles the day's summary.

"From documents" mode

Instead of schedule tables, the assistant answers from your uploaded policies: leave rules, safety instructions, shift standards. The answer carries a "from documents" badge so the source is obvious. Employees get the same mode in the Telegram bot.

Daily briefing

The Daily briefing button assembles today's summary for one credit: who showed up, where people are short, what changed in the schedule. You can subscribe to it — then it arrives in Telegram by eight in the morning, and costs no credits at all.

Comparison

How this differs from "export it and ask a chatbot"

The difference isn't the model — it's where the data comes from and who is accountable for the access boundaries.

Data goes stale the moment it's exported
Read at the moment of the question
The file carries everything exported — other departments included
Tools are assembled from the asker's permissions
Nothing to check the answer against: the file is gone
The answer shows which tools were opened
To change anything you go back and do it by hand
A proposal is executed by a button inside the conversation
Live data
Permissions instead of trust
Action in one click

The same boundaries as the app The assistant reads data by the same access rules as the interface: department, role, row-level permissions. The model receives only what a tool returned under your permissions — not a dump of the database.

Questions

What people usually ask about the assistant

No. Of the ten tools, eight only read and two prepare a proposal with a button. The schedule entry appears at the moment you press, under your session: with no edit permission, the button does nothing either.

The labor budget is available only with the budget permission — without it the tool never makes the list. Employee personal contacts aren't wired to the assistant at all: that data isn't in any of the ten tools.

One AI credit per question, however many tools the assistant opened. The briefing button also costs one credit, while the morning Telegram briefing costs none. Credits are shared across all AI features, and the monthly allowance depends on the plan: 30 a month on Start.

The panel says credits have run out and offers a link to billing. The rest of the app works as usual — AI blocks nothing.

The interface language: Russian, Ukrainian or English. It switches together with the app; there's no separate setting.

Only you. Conversations are closed at the database level, and even the organization owner cannot open someone else's. Credit spending is visible in the shared AI journal — amounts and dates, not the text of questions.

Every answer lists the tools it opened, and the same number is on the corresponding screen in the app. The assistant doesn't do its own math: it calls the same coverage, hours and budget calculations the interface shows.

Ask your own database its first question

14 days free, 30 AI credits a month on the Start plan. The assistant is on from day one — there's nothing to configure.

No card · set up in an evening