A customer wants an earlier delivery. You need an answer from operations, operations needs a clarification from a supplier, and an employee is waiting to know which job to prioritise. In a small management team, much of the day can disappear into finding out what has been agreed and who needs to answer next.
An AI agent could help prepare those exchanges and follow them up. But when it contacts someone in your name, another question appears: what can the recipient assume you have actually decided?
My view is that a leader should delegate the coordination they can describe clearly, while retaining the decisions that change other people’s commitments or priorities. The test is whether employees and customers get a clearer answer with less work. Saving the leader a few minutes is a poor result if everyone else must spend longer interpreting the agent.
Your agent also borrows your position
A message from the boss carries weight. An employee may rearrange the day to answer it, even if the question was generated without the leader reading it. A customer may interpret a suggested date as a promise. Writing in your usual tone can make those assumptions stronger.
I would make the status of a request visible. Is this an agent gathering information within an agreed task, a proposal awaiting your decision, or a message you approved? These distinctions should be clear to the person receiving it, including someone who uses no AI assistant.
That is a different question from whether a message sounds convincing. In my article on email and trust, I discuss why plausible wording says little about authority. Inside a company, a familiar name can create a similar problem: people may comply before checking what was authorised.
Follow one customer request through the business
Consider a hypothetical Norwegian service company with twelve employees. A customer asks whether an installation can move from Thursday to Tuesday. The general manager is away from the office. A proposed agent workflow might look like this:
- Prepare the situation. The agent reads the permitted order information and planning calendar, identifies what is missing, and prepares a short summary.
- Ask a bounded question. With permission to send this type of internal request, it asks the operations lead whether Tuesday is feasible. The request states that the date has not been approved and that no one should change the schedule yet.
- Bring back the decision. Operations says Tuesday is possible only if another job moves. The agent presents the conflict and drafts a customer reply. The manager and the employee responsible for planning resolve the priority.
- Communicate the approved outcome. The manager checks the recipient and exact wording before the reply is sent. A material change to the date or conditions requires another review.
The useful result is a prepared decision with the relevant facts. It is not the number of messages exchanged. If the answer from operations is uncertain or out of date, the agent should preserve that uncertainty instead of turning it into a confident customer promise.
This is a workflow to evaluate, not a claim that a particular product can already run it across all the company’s systems. For the product background, see how Dots and Grok Bot use cloud computers.
Give employees a say in how the agent approaches them
Before testing, ask the people who receive these requests what would make them useful. Agree on the channel, the response window and when reminders should stop. Silence should leave a request unresolved. It should not be treated as agreement. Give employees an easy way to correct a misunderstanding or take the matter directly to a person.
Employees may also have assistants of their own. A scheduling exchange between two agents should carry the same limits as the human task. An employee’s assistant cannot acquire the manager’s permissions simply by receiving a message from the manager’s agent. Nor should an agent get access to private conversations because they might contain useful context.
Explain what information is recorded, who can see it and when it is deleted. Datatilsynet’s privacy principles require defined purposes and personal data limited to what is necessary. My recommendation is to keep trial measurements about the workflow and avoid ranking employees by response speed or message activity.
If the arrangement involves control or monitoring measures, Arbeidstilsynet describes requirements for justification, proportionality, discussion with employee representatives, information and evaluation. An AI tool is not automatically a control measure; assess how it is used. The article on practical employee AI rules covers the broader workplace policy.
Agree on one communication task before connecting accounts
Write down a small mandate with the team. For the hypothetical company, I would start with the following:
Example for discussion
Clarify a proposed delivery change
Purpose: prepare a reliable answer to an existing customer’s request.
Sources and recipients: the relevant order, the approved planning calendar and the named operations lead. No access to unrelated employee or customer records.
Permitted communication: one internal availability question and one reminder within agreed working hours. Clearly identify the agent and the task it is acting on.
Decisions retained by people: moving another job, changing a price, approving overtime and sending the final customer reply.
Stop and hand over: when sources conflict, an employee disputes the request or no answer arrives by the agreed deadline. The named manager can pause the workflow and withdraw its access.
This is a proposed operating agreement, not a legal template. Support it with access restrictions and approval controls in the systems involved. A written instruction alone does not establish that an agent is technically prevented from doing something else. Test the pause, permissions and approval step before using live customer information.
Keep personal judgement in the relationship
Some messages matter because you took the time to send them. I would keep difficult feedback, a sensitive personnel conversation and an important apology as direct human communication. An assistant may help you prepare within appropriate data rules. Deciding how to listen and respond remains part of being a leader.
Routine external coordination is a more suitable starting point: prepare an introduction, clarify an agreed meeting or follow up a specific outstanding question. Agree on the purpose and contact limits. Making it cheap to generate messages does not give you a reason to send more of them.
Keep the recipient’s route to a person open. If a customer questions an answer, someone should be able to see what was sent, correct it and take responsibility. The customer should not have to diagnose which assistant made the mistake.
Start in Norway, preserve the choice of assistant
I want to explore this first with Norwegian businesses and ordinary Norwegian-language work. My proposed design principle is that leaders, employees and business partners should be able to choose different AI providers while retaining a way to communicate. Changing an assistant should not mean losing your relationships or the rules for contacting you.
That is an ambition to develop and test. It is not a claim that assistants from OpenAI, Anthropic, xAI or other providers already share a working communication system. People should also be able to participate without an agent.
Cooperation with customers and suppliers elsewhere in Europe would be a natural next step. European digital identity could eventually help with verification, but identity, company affiliation and permission to act remain separate questions. Digdir describes the separate process for introducing eIDAS 2.0 in Norway. Neither a future identity connection nor European hosting establishes compliance for a whole workflow.
Find out whether the whole team saves work
A small company can start by walking through a fictional case with the employees involved. If that exposes a useful task, evaluate the actual tools, data handling and controls before a limited live trial. Compare the total time spent preparing, replying, reviewing and correcting with the current process. Record missed questions, mistaken commitments and whether recipients understood who had approved what.
Continue if the team gets reliable answers with less effort. Narrow or stop the trial if it creates more chasing or uncertainty. Leaving all coordination manual also has a cost, particularly when one unavailable manager holds up the work. That makes a limited experiment worthwhile, without making automation the predetermined answer.
Management owns the workflow and the employee involvement. Where the trial creates material exposure, the board should ask management to explain the limits and results. The board should not manage individual messages.
Invitation
Help shape a working group for Norwegian leaders
I would like to gather a small working group of leaders in Norway who want to explore this with input from their employees. Start with one task: something that repeatedly stalls between you, your team, a customer or a supplier.
I am developing a solution in this area, and the discussions may inform its development. The first step is to understand the needs and assess a limited workflow. This invitation is an expression of interest, not registration for an operational service.
Email me about the working group with your role, company and one situation you would like to improve. Leave out confidential information and personal details about employees or customers. See how contact enquiries are handled.
Sources checked on 12 October 2026. The service company, workflow and mandate are hypothetical. Recommendations and the proposed communication model are my analysis, not documented product results or a legal assessment of a particular business.