Set up the AI: model, thinking, usage limit and permissions

Pick the model behind every AI feature and set how hard it thinks. Watch the monthly allowance and decide what the AI can do without asking. Three admin-only tabs of the Admin page hold all of it: Models, Usage and AI Permissions.

Choose the model

One model powers the whole workspace. Every choice on the Models tab applies the moment you click it, and there is no Save button anywhere on the tab. A choice the server refuses goes back to what it was, with a message saying why.

To change the model:

1. Open Admin, then Models.
2. Pick a model in the list below the three cards.
3. The three cards and the Powered by this model list now name your pick.

The list groups models by who made them, five makers in this order: Google (Gemini 3.8 Flash), Anthropic (Claude Sonnet 5 and Claude Haiku 4.5), OpenAI (GPT-5.6 Luna), DeepSeek (DeepSeek V4 Flash) and Tencent (Tencent Hy3). A workspace that has never chosen runs on GPT-5.6 Luna.

Each row carries a one-line description and one or two tags. Vision means the model can read images; Text-only means it can’t. ADK-native marks a model Mesy reaches directly rather than through a bridge. The model you are running wears an Active pill.

Set the fallback model

DeepSeek V4 Flash and Tencent Hy3 can’t read images. The fallback model is where an image or document scan runs when the active model can’t do it. Three models can take that job: Claude Haiku 4.5, GPT-5.6 Luna and Gemini 3.8 Flash. To change it:

1. Click Use as fallback on the row you want.
2. The row reads Currently your fallback model.

A text-only row reads Text-only, can’t be the fallback in place of the action. The fallback covers what the active model can’t do, so it has to be a model that can see.

Set the thinking depth

Every row in the model list ends with one chip per thinking level that model offers, from None up to Max. Lower is faster to first reply; higher reasons more deeply before answering. Levels differ per model.

On the model you are running, the chip for the depth in use is filled in. Click another chip to change it. The Active model card then reads Thinking: and the level you picked.

On any other row, the chip for that model’s own default is outlined, to show what would run if you switched. Click any chip on that row and Mesy moves to that model at that depth in one step.

A model that can’t use reasoning on the current stack shows a line saying so in place of its chips.

The depth applies to interactive chat. Background work, such as drafting, capture enrichment and workflow steps, runs at the model’s own default depth.

See what each AI feature runs on

The Powered by this model list names every AI feature and the model behind it, in three groups.

• Interactive: Chat agent, Telegram assistant.
• Background jobs: Email drafting, Capture enrichment, Email link suggestions, Workflow enrichment, Workflow email drafting, Daily insights, Import auto-match.
• Media: Vision extraction, Voice transcription.

Every row reads Inherits followed by the model that runs it. You can’t set a model per feature. When the active model is text-only, the vision row names the fallback model instead and says why. Voice transcription always reads Not configured (Whisper): it runs on a fixed model that doesn’t follow your pick, and there’s nothing to set.

Watch the monthly allowance

The Monthly AI allowance card shows how much of the month’s allowance your workspace has used, as a percentage, and the date it resets. Your operator sets the allowance; you can’t change it here. At the limit the card reads limit reached and AI features pause until the reset. Mesy never shows a currency amount to a workspace.

Turn email linking on or off

The Email linking control at the bottom of Models decides whether incoming and sent mail is linked to records by itself. It has two settings.

• Off: incoming and sent mail is never linked automatically.
• Suggest: mail from a matched contact gets an AI relevance check, and relevant messages become suggestions you accept or dismiss. Sent mail to a matched contact still links directly.

Under the two tiles, a line counts the last 30 days: how many suggestions were accepted, dismissed and are still waiting, plus the acceptance rate. Before anyone has answered a suggestion the line reads acceptance rate not measurable yet, rather than showing a rate of nothing. Shared mailboxes are never linked automatically in either setting.

Read the usage figures

Open Admin, then Usage. AI Usage Analytics breaks token use down by day, model and member. Pick Daily, 7d, 30d or 90d at the top right.

Daily shows Today’s Usage (Hourly); the other three show Usage Over Time. Below those sit Usage by Model and Usage by Member. Admins can click a member’s row to see their usage over time and by hour.

Inspect what the agent did

Three views share one block at the bottom of Usage, one at a time.

Agent turns lists every turn with its outcome, tool count and tokens, plus three speed figures. Time to complete is the whole turn, from your message to the finished reply, tool calls included. Time to first token is the wait before the first word appears. Tokens/s is how fast the rest of the reply came out. The last two only apply to a streamed reply, so a turn that ran on Telegram or as a background job shows a dash for both.

Above the table, Speed by model repeats the same three figures for each model your workspace has used, as medians. A model with no streamed turn shows a dash rather than a zero. Click a turn to open it: you get its speed figures, every tool call inside it, and how long each model call spent waiting and generating.

Tool calls lists every agent read and write with its duration and outcome. Raw audit log is the field-level change ledger for any actor, and it’s read-only.

Decide what the AI may do on its own

Open AI Permissions. Each specialist (admin, analytics, calendar, comms, crm, email and research) has one of four tiers.

At Read-only the AI can only read data. At Review every write shows you a confirm card first. At Auto-write writes run directly and deletes still show a confirm card. At Full writes and deletes run without confirmation, though team changes always confirm.

There’s no separate approval queue: a gated action shows a confirm card to you. Expand a specialist to see every tool it can use, grouped by risk. Each tool shows its behaviour at the current tier: Runs live, Confirm card, Blocked or Stages for send. Sending email always stages a card at every tier.

To change a tier:

1. Pick a tier in the Change column; the row expands to preview it.
2. Click Apply to save, or Reset to keep the current tier.
3. The Current tier badge updates.

The Telegram card lists rules that always apply there, plus one switch, Inline confirm in Telegram. On, you can confirm a staged card with a tap in Telegram; off, the same card appears without buttons and points you to Approvals. Editing a card’s fields is always web-only.