Meet Tsuno at Sirha Lyon 2027 · Stand S.57Learn more
EN FR
Try the demo
Menu

AI and tools

AI that analyses and AI that acts: two different jobs

Analysing (reading, matching, summarising) commits you less than acting (creating, changing, sending). Set them up differently: rights, confirmation, log.

Published on

In the storeroom, a supervisor checks an order list on a tablet against the shelves, a pen in his hand.

AI that analyses reads, matches, summarises and flags; AI that acts creates, changes or sends. A mistake by the first costs time, while a mistake by the second can send out an order, change a rota or deliver a message to the wrong person. So they should not be set up the same way: permissions, confirmation, logging and the ability to undo all need to be tighter as soon as the AI acts.

Thursday, 10:15 am, in the storeroom

Arnaud, the sous chef, is putting together Friday's order. On the tablet, the assistant has done its analysis: expected stock based on the week's sales, the waste logged on Tuesday, the butter price that went up at the supplier. From that it has drafted an order list. It looks right.

Arnaud checks it against the shelves anyway, pen in hand. Two cases of cream are missing that nobody took off the stock after Tuesday's function. And the list asks for twelve kilos of salmon, although the chef pulled the tartare from the weekend menu. Nobody has written that down yet.

He fixes three lines, reads it again, then confirms the send to the supplier.

The analysis was wrong twice. It cost five minutes, because a person read it before anything went out. Had the order gone out on its own, the mistake would have cost twelve kilos of fish.

(Fictional scene and characters.)

Why the two are so easy to confuse

The same window does both

In an AI assistant, you ask a question and request an action in one sentence: "Check what's missing and order it." To the user, it is one gesture. To the restaurant, it is two different jobs with two levels of risk.

A wrong analysis is fixed by reading it

An incomplete summary, a miscalculated gap, a match that does not hold: all of that stays on the screen. You can reread it, discuss it, ignore it. The worst outcome is a bad idea nobody followed.

A wrong action leaves the screen

An order lands at a supplier, or a published rota notifies the whole team. Once it is out, a mistake affects someone else, and it cannot always be taken back. That is where the settings have to change.

The four settings that change when AI acts

1. Permissions

An AI that analyses should already stay within the person's permissions: it only reads what they can see. An AI that acts must also only do what the person can do. A commis should not be able to use the assistant to change a colleague's shifts or approve an order reserved for the chef. And if a permission is removed between drafting and confirming, the action should not go through. The general principle is set out in give your AI the same rights as you, no more.

2. Confirmation

Reviewing an analysis is optional; reviewing an action is not. Hence the two-step rule, covered in AI proposes, a human confirms: the action waits for an explicit yes, given after reading its summary.

3. The log

An analysis does not need to be archived line by line. An action does. You need to find out who confirmed, when, what was done, and the state before and after. Without that, you can neither fix a mistake nor explain it.

4. The ability to undo

Some actions undo in one tap: a task can be closed, an unpublished shift can be moved. Others cannot be taken back: a message that was read, an order that was sent, data deleted for good. A good tool says before you confirm whether the action can be undone. The more final it is, the more carefully it should be reviewed, and some actions should not be handed to the AI at all.

Restaurant examples

On the analysis side, the AI helps without committing anything:

  • reading the week's temperature logs and saying which are missing;
  • matching a supplier invoice to the delivery note and flagging the gap;
  • summarising open incidents before Monday's meeting;
  • spotting that a purchase price has gone up since the last order;
  • flagging that nobody trained on the hot station is on the rota for Saturday lunch.

On the action side, every example affects someone:

  • creating a task for the closing team;
  • moving a shift on the rota;
  • publishing the week's rota;
  • sending an order to a supplier;
  • sending a message to the whole team.

The first list helps the manager see, the second changes someone's work.

The decision grid

For each action, three questions are enough: who is affected if it is wrong, can it be undone easily, and does anyone outside the team receive it. Here is how they play out on everyday cases.

ActionWho is affected if it is wrongCan it be undone?Can the AI prepare it?Who decides
Summarise the week's incidentsNobody, it is a readingNot applicableYes, on its ownThe manager reads
Create a closing taskThe evening teamYes, easilyYesThe person confirms
Move an unpublished shiftThe employee concerned, laterYesYesThe manager confirms
Publish the rotaThe whole team, notified at onceBadly: the notice is outYes, with a reviewThe manager confirms after reviewing
Send a supplier orderThe supplier, the cash flowBadly: you have to call themYes, with a check of quantitiesThe supervisor confirms
Send a message to the teamEveryone who reads itNoA draftThe manager reviews and sends
Payroll, invoicing, permanent deletion and legal actsEmployees, customers, the recordsNoNoThe dedicated screens and authorised people
Discipline, rating or dismissalOne person and their careerNoNo: only dated factsA human, fully

The further down the grid you go, the further the AI steps back: it reads alone at the top, prepares and waits in the middle, then touches nothing at the bottom.

What should stay entirely human

Some actions should not be handed to a conversation, even with a confirmation:

  1. Payroll and invoicing.
  2. Permanent deletion.
  3. Legal documents: contracts, amendments, letters that commit the business.
  4. Disciplinary action and any adverse decision about an employee.
  5. Judging a person: rating them, ranking them, deciding who to hire or let go.

For the last two, the law has a say too. Under EU law, Article 22 of the GDPR gives everyone the right not to be subject to a decision based solely on automated processing when it significantly affects them. The AI can gather the dated facts that prepare a one-to-one. It does not decide. That is the idea behind AI for frontline managers: analyse, advise, reassure.

Common mistakes

  • Setting up every action the same way. Confirming a new task and confirming a published rota do not call for the same attention.
  • Confirming without reading. If "yes" becomes a reflex, the two-step rule protects nothing. A good summary is short, so people actually read it.
  • Thinking a good analysis makes the action safe. In the scene, the analysis was solid. It was just missing what nobody had written down.
  • Giving the assistant a wider account than your own "to make things easier". The AI should have your permissions, not the general manager's.
  • Forgetting the log. An action with no trace is an action nobody can explain the next day.

What to measure

  • The share of AI-prepared actions that get corrected before confirmation. If it is zero, check that people are still reading.
  • The number of actions that had to be undone afterwards, by type. That is the clearest sign of an action placed too high in the grid.
  • Time saved on analysis (Monday's review, the order, checking the logs). That is where the AI pays off with no risk.

Where Tsuno and San come in

In Tsuno, the path is the same whether you go through the app or your usual AI assistant (ChatGPT, Claude, Mistral): drafting, a summary of the effect, confirmation, execution, then a log entry with the state before and after. On the analysis side, the assistant reads within your scope, and its reads are capped. On the action side, it drafts within your permissions, checked again at the moment you say yes. Payroll, invoicing, permanent deletion and legal acts stay out of its reach, even when confirmed. The details are on the Trust and data page, and what the app does is listed on Features.

Tsuno San, the smart voice assistant for teams, follows the same rules: San Voice is available as a demo, and every action goes through the two steps. See Tsuno San.

One last limit: a confirmation does not replace attention. If the team taps "yes" without reading, no tool can protect it.

Further reading

For the role of AI alongside managers, read the pillar article AI for frontline managers. For the confirmation rule, see AI proposes, a human confirms, and for plugging an assistant into your tools, MCP explained for frontline trades. The line between helping an employee and analysing for management is drawn in conversational assistant or job assistant: where is the line, and San's limits in what San never does.

Key takeaways

Analysing and acting are not the same job: a wrong analysis costs time, a wrong action affects someone. Let the AI read and prepare, keep confirmation for anything that changes other people's work, and keep fully human whatever commits a person or cannot be undone.

Frequently asked questions

What is the difference between AI that analyses and AI that acts?

AI that analyses reads, matches, summarises and flags: if it is wrong, you lose time. AI that acts creates, changes or sends something: if it is wrong, an order goes out, a rota changes, a message reaches someone. The two should not be set up the same way.

Should you let an AI send a supplier order on its own?

It is not wise. A sent order commits money and is hard to take back. The AI can draft it and flag the gaps, but a person checks the quantities and confirms the send.

Which actions should stay entirely human?

Payroll, invoicing, permanent deletion and legal acts, but also disciplinary action and any decision about a person: rating them, hiring, letting them go. The AI can gather facts for these, not decide or carry them out.

How do you know if an AI action is safe with a simple confirmation?

Ask three questions: who is affected if it is wrong, can it be undone easily, and does anyone outside the team receive it. The heavier the answers, the more human the action should stay.

Why keep a log of actions taken with AI?

To know who confirmed what, when, and what changed. Without a log you cannot fix a mistake or explain a decision. In Tsuno, the log keeps who, when and what, with the state before and after the action.

Does an AI that acts have more rights than the person using it?

It never should. In Tsuno, the assistant only sees and does what the person can already see and do, and permissions are checked again at the moment of confirmation.