Set up the AI

Mesy AI is the AI you chat with in Mesy. This page helps you choose the model it runs on and decide what it may change without asking. It also covers the monthly allowance, email suggestions and your team’s AI usage. Only admins can do this, on the Models, Usage and AI Permissions tabs of Admin.

Choose the model

A model is the AI engine that writes Mesy AI’s answers. One model runs your whole workspace. Every choice on the Models tab saves as soon as you click it.

To change the model:

1. At the bottom of the sidebar, click Admin.
2. Click the Models tab.
3. In the list of models, click the model you want.
4. A message reads “Active model saved”. The Active model card at the top names your pick, and its row shows Active.

The list groups models by the company that makes them:

• Google: Gemini 3.8 Flash
• Anthropic: Claude Sonnet 5 and Claude Haiku 4.5
• OpenAI: GPT-5.6 Luna
• DeepSeek: DeepSeek V4 Flash
• Tencent: Tencent Hy3

A workspace that has never picked a model runs on GPT-5.6 Luna. Each row has a one-line description. Vision means the model can read pictures. Text-only means it can’t.

Set how long the model thinks

Thinking depth is how long the model reasons before it answers. A lower level replies sooner. A higher level thinks harder first. The level only changes your chats with Mesy AI. When Mesy drafts emails or runs a workflow (a rule that does routine work for you), it uses the model’s default level.

Each row ends with level buttons, such as None, Minimal, Low, Medium, High, X-High and Max. Each model offers its own set.

To change the depth:

1. On the active model’s row, click a level.
2. A message reads “Thinking depth saved”. The level button is filled in, and the Active model card shows the new level, for example “Thinking: Low”.

On any other row, one level has an outline. That’s the model’s default level. To switch to that model, click a level on its row. Mesy switches the model and sets that level together.

If you switch by clicking the model’s name instead, Mesy keeps your current level when the new model offers it. Otherwise it uses the outlined default.

Pick a fallback model

A fallback model is a backup that reads pictures when your main model can’t. DeepSeek V4 Flash and Tencent Hy3 can’t read pictures. If you use one of them, Mesy uses the fallback to read photos and scans, such as a business card.

Only Claude Haiku 4.5, GPT-5.6 Luna and Gemini 3.8 Flash can be the fallback. Until you pick one, the fallback is GPT-5.6 Luna.

To change the fallback:

1. On the row you want, click Use as fallback.
2. A message reads “Fallback model saved”. The row now shows Currently your fallback model, and the Fallback model card names it.

The fallback doesn’t help in the web chat. There, a text-only model can’t see pictures you attach, and Mesy AI only sees the file’s name. To chat about pictures, pick a model marked Vision.

See which model each AI feature uses

The Powered by this model list shows every AI feature in Mesy, such as Chat agent, WhatsApp assistant, Teams assistant and Chat reply suggestions. Each one uses the model you picked, so there’s nothing to set here. Two rows differ:

• If your model is text-only, Vision extraction shows your fallback model instead.
• Voice transcription shows “Not configured (Whisper)”. Mesy sets this row, and you can’t change it.

Check the monthly allowance

The Monthly AI allowance card sits at the top of the Models tab. It shows how much of this month’s allowance your workspace has used. For example, it reads “AI usage: 42% of monthly allowance”. Under the title, it shows the date the allowance resets, which is the 1st of next month.

The bar changes colour at 75% and again at 90%. When you reach the limit, the card shows “limit reached” and AI features pause until the reset. Mesy sets your allowance, so you can’t change it here.

Choose whether Mesy suggests email links

Under the model list, Email linking decides whether Mesy puts emails on customer records for you. A record is one entry in Mesy, such as one customer or one deal. A linked email shows in that record’s history. A new workspace starts on Off.

To change it, click one of the two options:

• With Off, Mesy doesn’t link any email by itself.
• With Suggest, when you email a contact Mesy knows, Mesy links that email straight away. For emails you receive from contacts it knows, Mesy AI checks whether each one is relevant and suggests a link for the ones that are. The person whose mailbox it is accepts or dismisses it.

A message reads “Email linking saved”. A line underneath counts the last 30 days: suggestions accepted, dismissed and still waiting, and the acceptance rate. Until someone answers a suggestion, it reads “acceptance rate not measurable yet”. Shared mailboxes are never linked automatically, whichever option you pick.

To accept or dismiss a suggestion, see Add an email to the CRM.

Decide what Mesy AI may do on its own

On the AI Permissions tab, each area Mesy AI works in has its own row. The tab calls these areas specialists: Admin, Analytics, Calendar, Crm, Email and Research. Each specialist has a tier, which sets how much Mesy AI may do without asking you first.

A confirm card is a box Mesy AI shows before it changes something. Nothing happens until you confirm it. The tiers are:

• Read-only stops Mesy AI changing anything by itself. In the web chat it can still show a confirm card that adds or updates records, and nothing changes unless you confirm it.
• Review shows you a confirm card before every change. Two kinds of change run straight away: your own notification settings, and the workspace’s AI model or thinking depth.
• Auto-write makes changes straight away. It still shows a confirm card before deleting. Bigger changes also get one, such as bulk updates, reassigning a record or overwriting a long text field.
• Full makes changes and deletes without a confirm card.

A specialist nobody has set yet is on Tier 1: Read-only. Some actions show a confirm card even at Full, such as changing a member’s role, sending an invite, adding a calendar event and merging records. At Read-only, they’re blocked. Sending an email always needs your confirmation too.

The confirm card goes to the person who asked Mesy AI. They confirm it themselves, and nobody else needs to approve it.

The same tiers decide what workflow steps may do. At Read-only, a workflow step that changes something is blocked. See Build a workflow.

To change a tier:

1. In the Change column, pick a tier. The row opens and shows what Mesy AI could do at that tier, marked as not saved yet.
2. Click Apply. (To keep the current tier instead, click Reset.)
3. The Current tier column shows the new tier.

Click any specialist’s row to see each thing it can do, from looking things up to deleting. Each one shows what happens at the current tier. Runs live means Mesy AI does it straight away. Confirm card means you confirm first. Blocked means it can’t. Stages for send means Mesy prepares the email and waits for you to confirm the send.

Let people confirm from a chat app

Below the table, the Telegram card has one switch, Inline confirm in chat channels. Despite the card’s name, the switch covers every chat app you use to talk to Mesy: Telegram, WhatsApp and Teams. It starts on.

When it’s on, people tap to confirm in the chat app. When it’s off, the confirm card still shows in the chat app, but without buttons. People confirm it on the Approvals page (in the sidebar) instead. The Telegram card also lists fixed rules, such as which tools only work on the web. Each is marked Always, and you can’t change them.

For how confirming works in each chat app, see Approve AI actions from a chat app.

See how reply suggestions are used

In WhatsApp, Messenger and Instagram chats, Mesy can suggest a reply for your team to send. The Reply suggestions table at the bottom of Models covers the last 30 days. For each kind of question, it shows how many replies Mesy suggested and how many were sent. Auto-sent counts the ones Mesy sent by itself. To turn suggestions on, see Get reply suggestions in customer chats.

Read your AI usage

AI use is counted in tokens, which are small pieces of the words Mesy AI reads and writes. The Usage tab shows the last 30 days to start with.

To see your workspace’s usage:

1. In Admin, click the Usage tab.
2. Next to AI Usage Analytics, click Daily, 7d, 30d or 90d to choose a period.
3. Read the five cards at the top. They show the totals for that period: Total Requests, Input Tokens, Output Tokens, Total Tokens and Avg Tokens / Request.

With Daily, the chart is Today’s Usage (Hourly). With the other periods, it’s Usage Over Time, one bar per day. Click a bar to see that day hour by hour. Click Close to go back.

Below the chart, Usage by Model splits the tokens by model and Usage by Member splits them by person. Click a member’s row to see their Usage Over Time, Model Breakdown and Activity by Hour.

Look back at what Mesy AI did

At the bottom of Usage, three tabs show Mesy AI’s work in detail. Agent turns lists each request Mesy AI worked on: who asked, the model, the outcome, how long it took and the tokens used. It doesn’t show the words of the request or the reply. Tool calls lists each lookup and change Mesy AI made. Raw audit log lists every change anyone made to a record.

For an easier view of your team’s history, see See who changed what.

If something doesn’t work

• If Mesy can’t save a choice on Models, it puts back the previous choice and shows a message saying why.
• If AI features have stopped, check the Monthly AI allowance card. At “limit reached”, they pause until the reset date.
• If Mesy AI says it can’t make a change, check its specialist’s tier on AI Permissions. At Read-only, it can’t make changes by itself.
• If Mesy AI refuses a change even at Full, check the person’s role. A member can’t make manager or admin changes through Mesy AI, and a manager can’t make admin changes.
• If the Usage tab shows no cards, nobody used AI in that period. Pick a longer one.