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

Future of work

Conversational assistant or job assistant: where is the line?

A conversational assistant answers anyone with no context; a job assistant knows your location, acts within your permissions and asks for confirmation.

Published on

In the storeroom, a manager checks a delivery on a tablet while a colleague puts away crates of vegetables.

The line between a conversational assistant and a job assistant has nothing to do with how good the conversation is. It depends on what the assistant knows and what it is allowed to do. A general-purpose assistant answers anything, for anyone, without knowing your business or the person talking to it. A job assistant knows the roles, the validated procedures and each person's permissions, prepares actions without carrying them out on its own, and knows when to stay quiet. In hands-on work, its value is not in chatting: it is in moving real work forward and leaving a useful record behind.

Thursday, 3:20 p.m., in the storeroom

The afternoon delivery has just arrived. Camille, the manager, is checking it off on the tablet. Next to her, Moussa is putting the boxes away, a lightweight earpiece at his ear. One box of cream has taken a knock; the packaging is crushed.

First version of the scene. Moussa takes out his phone and asks a general-purpose chatbot: "What should I do with a damaged box of cream at delivery?" The answer is long and sensible. Nothing wrong with it. But nothing about the house rule, nothing about who to tell here, and nothing is recorded. In a week, nobody will remember that box.

Second version. Without putting the boxes down, Moussa asks: "The cream box is damaged, what do I do?" The assistant reminds him of the location's own receiving procedure, the one that was validated, and offers to prepare a report on this delivery. Moussa confirms. Camille sees it on her tablet. And when she wonders that evening whether it is the first time, she can ask and get dated facts, not a hunch.

Both scenes are fictional. Yet they describe the line better than any definition: in the first, there was a conversation. In the second, the work moved forward.

What a general-purpose assistant does well, and what it cannot do

Credit where it is due: general-purpose assistants have become very good at drafting, rewording, summarizing and explaining. The problem is not how smart they are. It is where they sit.

It answers anything, for anyone

A general-purpose assistant is built to answer anyone about anything. That is its strength and its limit. It does not know your walk-in, your roles or your standards. It answers as it would for any restaurant, which means not quite for yours.

It has no context and no permissions

It does not know who is talking. An apprentice who started on Monday and a group operations director get the same kind of answer. In a restaurant, that is a problem: not everyone can see other people's schedules, review notes or a location's results. A tool that ignores permissions cannot safely be plugged into a team's real working life.

It fills the gaps

Faced with a question that has no answer, it produces a plausible one. That feels helpful until the day someone acts on information that never existed. In a kitchen, on a temperature or an allergen, that is unacceptable. We go into this in why an assistant must be able to say "I don't know".

It leaves no trace in the work

The conversation ends, and nothing has changed in the restaurant: no task ticked off, no report filed. That is one reason why a stack of general-purpose tools does not replace a tool built for the job, as we explain in why not build Tsuno with ChatGPT, Notion and Make?.

What makes a job assistant

A job assistant can be recognized by six properties.

  1. It knows the location. Roles, validated and dated procedures, checklists, routine checks, equipment. It answers from what the team has written and validated, not from generalities.
  2. It knows who is talking. Each person has their own session. It acts within their permissions: it only sees and does what they can already see and do.
  3. It acts in two steps. It prepares the action and summarizes it, the person confirms. Nothing runs without that confirmation, and the decision is recorded.
  4. It knows when to stay quiet. On critical topics (allergens, temperatures, safety), it does not summarize from memory: it shows the validated sheet or says it does not know, and points to a person.
  5. It keeps answers short. One sentence, one step, a short list. The person has a task to finish.
  6. It leaves a useful record. A task done, a reminder set, a report filed. What was asked and confirmed becomes a dated fact that others can find later.

Making the day easier without making conversation

A job assistant makes the working day easier: no more searching, interrupting a colleague or guessing. It also brings follow-through, because what it helps people do leaves a record. But for hands-on, operational work, one position has to be clear: its value is not measured by how long people spend talking to it.

In a lot of digital products, the longer you stay, the better. In a kitchen, it is the opposite: every minute spent chatting with a tool is a minute taken from prep, cleaning or service. The right job assistant is the one you use for ten seconds, that moves a task forward, and then goes quiet.

The other layer: analytical AI for managers

The person on the line wants a short answer so they can keep working. The manager needs perspective: seeing what keeps coming back, deciding where to spend their time.

That is a different layer of AI, built for management. It helps to:

  1. Connect facts: an incident that comes back for the third time in a month, a closing checklist often left incomplete on Sundays, an action decided and never closed.
  2. Anticipate, by answering "what deserves my attention today?" before a problem shows up on the floor or in the numbers.
  3. Plan: prepare a week, an opening or a site visit from what is already known.
  4. Communicate clearly: draft a message to the team or a briefing point, which the manager reviews, edits and sends themselves.
  5. Follow up on actions: find what was decided, by whom, and whether it is done.

The doctrine behind this layer fits in two sentences: Tsuno connects the facts. People decide. The AI presents dated facts, says where they come from and what is missing. It does not decide in the manager's place and it does not judge people. We set this out in software recalls the facts, people decide: a doctrine for AI in HR, and how to ask these questions is described in talking to your restaurant: gimmick or new interface?.

One source, two uses

The two layers feed each other. What the job assistant helps people do every day (a task ticked off, a check completed, a report filed) becomes the material the analytical layer can connect for the manager. Without a record there is nothing to analyze, and a record nobody looks at ends up forgotten in a binder.

But it must be clear what gets connected: dated facts about the work. A checklist done or not, a reading, an incident, an action. Not the tone of a voice, not a mood, not an employee score.

Who does what between a chatbot, a job assistant and analytical AI

CriterionGeneral-purpose chatbotJob assistant (San)Analytical AI for managers
Who it is forAnyoneThe person on shiftThe manager, the owner
What it knowsGeneralitiesThe location: roles, validated procedures, checklists, equipmentThe history of facts at a location or across a group
Who is talkingIt does not knowIt knows, each person has a sessionIt knows, based on role
PermissionsNo frameworkThe person's, nothing moreThe manager's, nothing more
Shape of the answerLong, fluentShort, one step at a timeShort, with the facts behind it
Missing dataFilled inFlagged, points to a personFlagged as missing
ActionsNone, or outside any controlPrepared, confirmed by the person, recordedPrepared, confirmed by the manager, recorded
RecordIn a personal chat historyIn the work: task, reminder, reportIn the follow-up of decisions
Critical topicsAnswers like anything elseShows the validated sheet or stays quietPresents the facts, does not conclude alone
Its value is measured byThe quality of the textWork moved forward, without interruptionsDecisions made earlier, based on facts

Where the red lines are

The closer an assistant gets to real work, the more its limits need to be written down. Here are ours.

  1. It does not judge people. The AI neither scores nor ranks employees, and San evaluates no one. Tsuno does show a monthly leaderboard (points earned through training, punctuality and completed tasks), but it is never passed to the assistant: the AI does not use it. An evaluation remains a human act, based on facts. See AI should not judge your employees.
  2. It does not interpret emotions. Inferring a person's stress, fatigue or mood from their voice or behavior in the workplace has been prohibited in the European Union since February 2, 2025 (Regulation (EU) 2024/1689 on artificial intelligence, Article 5). We would not do it anyway.
  3. It does not act without confirmation. Every action is prepared, summarized and then confirmed by a person. Payroll, invoicing, permanent deletion and legal acts are excluded from what the AI can do.
  4. It does not record what should not be recorded. Tsuno refuses to record an employee's health, pregnancy, disability, religion or opinions (special category data under Article 9 of the GDPR). Instead, it suggests recording dated, observed facts.
  5. It does not answer critical questions from memory. Allergens, temperatures, safety: the validated sheet, or "I don't know, ask the person in charge."

What if the team already uses ChatGPT, Claude or Mistral?

Many managers already use a general-purpose assistant to write or think things through. The line is not about the assistant's brand. It is about what it is connected to and the permissions it acts with.

With Tsuno, you can query your data from Tsuno itself or from your usual assistant (ChatGPT, Claude, Mistral), with the same permissions. The general-purpose assistant then becomes one more way into the location's memory, and Tsuno's rules still apply: same permissions, same two-step actions, same record. What changes is where you ask the question from, not what you are allowed to see or do.

Common mistakes

  • Judging an assistant on the quality of its conversation. An elegant answer that moves nothing forward is still just an elegant answer.
  • Pasting your procedures wholesale into a chatbot. With no validated version, no date and no permissions, you create a second, parallel truth that nobody can check.
  • Measuring use by time spent. In hands-on work, long sessions are often a sign of a poor answer.
  • Confusing analysis with judgment. Connecting facts about the work is not evaluating a person. The day that line blurs, the team's trust is gone. On that risk, read management software or surveillance: where is the line?
  • Opening everything to everyone "so the AI is smarter". An assistant does not need to see more than the person talking to it.

What to measure

  • The share of requests that end in a confirmed action or an answer backed by a sheet, over a week. That is the measure of work moved forward.
  • How many times the manager is interrupted for a procedure question, before and after.
  • The time between a reported fact and a manager's decision, tracked on three real cases.

Where Tsuno comes in

Tsuno San is the job assistant: the intelligent voice assistant for frontline teams, connected to the location's procedures, tasks, checklists, routine checks, equipment and information. It acts within the person's permissions, prepares actions and waits for their confirmation. San Voice, voice interaction, is available in demo; San Audio (hands-free) and San 360 (a broader vision of operational support) are under exploration and testing. Details are on the Tsuno San page.

For managers, Tsuno answers questions from the location's memory, with the facts behind each answer. What we do with your data is described on the trust and data page.

If you are looking for a tool to write your menus or social media posts, a general-purpose assistant will do that very well. Tsuno is not built for it. It is built so that real work moves forward, and what was done can be found again.

Going further

To see San as a whole, read the pillar page Tsuno San, the work companion for frontline teams, then what San never does. On the management side, software recalls the facts, people decide sets out the doctrine, and talking to your restaurant: gimmick or new interface? offers a simple test for judging an AI on your own questions. This debate is part of a wider shift, described in the future of work in hands-on jobs. On the voice side, the same line runs between a general voice assistant and a companion that knows your workplace, and on the action side, between AI that analyses and AI that acts.

Key takeaway

A conversational assistant talks to everyone; a job assistant knows who it is talking to, what it is allowed to do and when to stay quiet. For frontline teams, its value is the work moved forward and the record left behind; for managers, it is connected facts, and the decision stays human.

Frequently asked questions

What is the difference between a chatbot and a job assistant?

A general-purpose chatbot answers anyone about anything, without knowing your business or the permissions of the person talking to it. A job assistant knows the roles, the validated procedures and each person's rights, prepares actions without carrying them out on its own, and can say it does not know.

Can we use ChatGPT to give procedures to kitchen staff?

It is useful for drafting or rewording a procedure. To share it with the team, the essentials are missing: the validated, dated version, each person's permissions and a record of what was done. With Tsuno, however, you can query your data from ChatGPT, Claude or Mistral, with the same permissions as in Tsuno.

Can an AI assistant at work act without approval?

It should not. In Tsuno, every action happens in two steps: the assistant prepares and summarizes, the person confirms, and it is recorded. Payroll, invoicing, permanent deletion and legal acts are excluded from what the AI can do.

Can AI help a manager without judging the team?

Yes, if it connects dated facts about the work (a checklist done or not, an incident that keeps coming back, an action left open) and leaves the decision to the manager. It should neither score nor rank employees, nor interpret their emotions.

Is a chattier assistant more useful in a restaurant?

No. In hands-on work, time spent talking is taken from the task. A good job assistant answers briefly, moves a task forward, leaves a useful record and then goes quiet.