IN 1 MINUTE
You receive a perfectly written email: correct, structured, polite and perhaps considerably longer than necessary. As soon as you read the first few lines, you sense that it has been generated with artificial intelligence.
You have little time and too many things on your plate, so you pass it through another AI to summarise what matters. Then you use that same tool to help you prepare the reply. Exactly the same thing could be happening on the other side.
The scene is somewhat absurd, but it raises an increasingly relevant question for companies. As AI gains capability and autonomy, we are no longer talking only about using a tool to perform a task faster. We are starting to delegate parts of the work: analysis, decisions, interactions and, progressively, complete sequences within a process.
To decide how far we want to go, we need to understand something that does not always appear in a procedure: what is really happening around that work. Because while someone is replying to a client, they may be doing much more than replying. They may be interpreting a situation, using knowledge accumulated over years, detecting an exception or building a relationship.
Before deciding what we delegate to AI, we need to understand what else we are delegating along with it.
The question we should be asking ourselves
When we delegate part of the work to AI, do we really know everything we are delegating along with it?
When one AI writes and another AI replies
In a recent post, a deliberately provocative expression was used to describe a similar situation: “Meat Proxy”. The human almost becomes an intermediary between artificial intelligences: one AI drafts, someone reviews and sends, and at the other end another AI summarises, interprets and prepares a reply. It is a caricature, but it is becoming increasingly recognisable.
We use artificial intelligence every day to research, analyse information, prepare meetings, programme, write, test hypotheses or explore alternatives. We do too. And it would make no sense to go back.
However, the conversation is changing. During the first years of generative AI, we focused heavily on what tasks it could perform and how much time it could save us. Now we are starting to ask something different: what parts of the work it can take on, what decisions it can make and with what degree of autonomy we want it to act.
The change is important. One thing is to ask an AI to prepare a draft email. Another is to allow it to interpret the situation, decide what response is appropriate, send it and then determine what the next step should be. In both cases we are using AI, but we are not delegating the same thing.
AI no longer only helps us do the work
Productivity has rightly been one of the main arguments for bringing AI into companies, and there are already organisations achieving significant improvements. But the results are not uniform.
One of the differences seems to lie in how the technology is introduced. Redesigning the work and then deciding where AI can add value is very different from adding tools to existing processes and expecting them, on their own, to make those processes better.
It is also worth being careful when we try to turn work into percentages. The fact that an AI can perform a certain proportion of the tasks associated with a role does not necessarily mean that the same proportion of the work can disappear. Real work includes tasks, but also exceptions, coordination, context, decisions, supervision, relationships and accumulated knowledge.
With AI agents, this difference will become even more relevant. We are moving from systems that primarily help us produce an output to others capable of executing several stages of a process. They can consult information, interpret it, use tools, decide the next step and act within defined limits.
That is why the question is no longer simply what AI can do. We need to decide what we want to delegate to it and with what degree of autonomy. And to answer that well, we need to understand much better the work we are placing in its hands.
What does not appear in a presentation
There is one part of a company that is relatively easy to know. It is in its presentations, its website, its procedures, its data, its organisation charts or its manuals. And then there is everything else.
At LVS2, we have always spent a great deal of time on that second part, even when the project in front of us could apparently be solved without doing so. We need to understand how a company really works: how it makes decisions, what concerns its clients, why something is done in a certain way even though no one remembers where it was decided, or what someone knows after ten years working there that we would struggle to find in a presentation.
It is what lies between the lines. It is not always said, but you perceive it when you have been working with an organisation for some time. With AI, this way of getting to know a company takes on another dimension.
We can document a process, measure its timings, identify tasks and locate inefficiencies. All of that is necessary. But the map will remain incomplete if we do not understand what is really happening while that work is being carried out.
Someone knows that a certain client needs a phone call rather than another email. Someone else remembers why, years ago, it was decided to stop working in a certain way. A salesperson understands that a “we’ll see” can mean completely different things depending on the client, the person they are speaking to or the moment. None of that necessarily appears in the procedure.
Part of an organisation’s knowledge is transmitted precisely in this way. A new person observes how someone with more experience makes decisions, listens to conversations, asks why something was done in a certain way and also learns from their own mistakes. Over time, all of that becomes part of what the company knows.
It is difficult to measure in hours or include in a productivity calculation. Sometimes it is even difficult to identify until it is no longer available. That is why this is not about preserving inefficient processes or roles simply because they have always existed. It is about understanding what we are transforming before we transform it.
We can delegate part of the work to AI in a matter of weeks and discover later that around it there was knowledge that had taken years to build.
Having all the data about a client does not mean knowing the client
Perhaps one of the places where this difficult-to-document value is easiest to understand is in client relationships. A company may have years of perfectly recorded information about a client and still not know them particularly well. The opposite can also happen.
Someone who has worked with them for years knows when they need a detailed document and when they prefer a two-line reply. They know when to call, when to wait and how to raise a disagreement. They know concerns that may never have appeared explicitly in a briefing.
AI can help us enormously here. It can retrieve background information, structure scattered information, prepare a meeting, detect patterns or prevent all that knowledge from depending exclusively on one person’s memory.
But having all the information about a client and truly knowing the client are still not exactly the same thing. If we do not pay attention to that difference, we can introduce layers of intermediation where there was previously a direct relationship, generate formally better but less relevant responses, or standardise a relationship that worked precisely because it was not standard.
And then we are no longer talking only about productivity. We are also talking about trust, customer experience and brand.
What happens to us in the meantime?
There is another part of this transformation that is more difficult to measure. We are increasingly using AI to remember, synthesise, formulate, compare, argue, write, review and help us decide.
This opens up some questions for which we should not invent answers. Is it cognitively equivalent to build a line of reasoning and to review one that has been generated and decide that it is correct? What happens to learning when we reduce part of the effort required to reach an answer? Could we confuse our ability to obtain an excellent answer with our actual ability to understand what we are answering about?
We wanted to move beyond intuition. That is why we have asked a psychology professional to help us distinguish what we currently know, what research is beginning to suggest and what it would still be premature to claim.
We are also interested in what happens to relationships. If we know that a message has been generated by AI, does our perception of the person sending it change? Does the effort we believe another person has put into communicating with us matter?
And, above all, what happens when that intermediation stops affecting only clients or suppliers and starts becoming a regular part of conversations within the company itself?
These are difficult questions to translate into a productivity metric, but they may end up having very concrete consequences. Also for the brand.
The brand starts inside
When we talk about brand, we tend to look outwards: clients, market, candidates, partners, investors. However, an important part of the brand is built much earlier.
It lies in how decisions are made, in what an organisation considers acceptable and in what “we don’t do here”. It lies in how a new person is onboarded, how an exception is resolved or how someone decides to respond when the procedure is no longer useful.
A brand is not only what a company says about itself either. It is also in many of those things that do not appear in a presentation, but that a client ultimately perceives after working with the company.
Customer experience does not come exclusively from a brand manual. It also comes from hundreds of small decisions made by people who understand what company they represent. That is why decisions about AI may also end up being decisions about culture and brand, even if they were initially framed as questions of productivity or technology.
Not because everything should continue to be done by a person. That would be absurd. There is work that AI can do better, faster or with fewer errors. There are tasks that we should probably stop doing manually. And there are new capabilities that simply did not exist before.
The question is knowing what we want to gain and what we do not want to lose along the way. A conversation can transmit knowledge. A discussion can build judgement. An explanation helps us understand a decision. An ongoing relationship builds trust. Even a mistake can eventually become part of collective learning.
Not all of those frictions add value. But not everything that appears inefficient is without value either. And there is an important difference between delegating part of the work and, without realising it, delegating everything that used to happen around that work.
The LVS2 strategic perspective
At LVS2, we have spent years trying to understand precisely that part of companies that does not always appear in a briefing, a presentation or a procedure. The arrival of AI does not change that way of working. It makes it even more necessary.
To decide what role technology should play, we first need to understand what visible and invisible value exists around the work we want to transform. That means looking at efficiency, of course, but also at the knowledge the organisation needs to retain, the relationships that may be affected, the judgement being built, the impact on culture and brand, and the risks that appear when we grant a system greater autonomy.
At LVS2, we analyse that visible and invisible value before deciding what we delegate to AI, what role we want it to play and with what level of autonomy.
It is a human-centric approach, but not because it assumes that a person must necessarily remain in every process. It starts from something different: understanding what result we want to achieve and designing from there the right relationship between people, technology and organisation.
In some situations AI will assist a person. In others it may propose a decision or execute actions within limits. And there will be processes in which it can take on a much greater degree of autonomy.
There is no single correct answer for everyone. That is precisely why judgement matters.
Why should this be on the radar?
Because the capabilities of these systems are increasing much faster than our organisational experience of using them.
Until now, much of AI adoption has happened with a person in the middle: someone asks, reviews an output, corrects it and decides what to do next. With agentic systems, that intervention can shift.
AI can interpret a situation, use different sources of information, decide the next step, communicate with other systems or people and execute actions. That opens up enormous possibilities, but it also changes the nature of the business decision.
It is no longer enough to know whether something can technically be done. We need to know what knowledge that work requires, what exceptions may arise, what relationships are involved, what consequences an incorrect decision could have, where we need supervision and who is accountable when something does not work as expected.
The more autonomy we are able to grant, the more important it becomes to decide where it makes sense to grant it.
LVS2 Framework: checklist before deciding what we delegate to AI
Not all processes need the same degree of human intervention or technological autonomy. Before deciding, we suggest looking at six dimensions.
What does the organisation gain?
Does delegation provide time, cost savings, speed, capacity, quality, error reduction or the removal of friction?
In addition, it is worth carrying out an extra check: are we eliminating work or simply shifting it to another part of the process?
What does it need to preserve?
We need to identify the knowledge, context, skills and judgement that the organisation needs to retain, especially what currently resides in people, is not sufficiently documented and would be difficult to rebuild once lost.
What relationships can it alter?
Clients, employees, suppliers, partners and other stakeholders may be affected. It is not enough to know whether an interaction can be delegated. We need to understand what changes when we do so.
What can it change about who we are?
Culture, trust, perceived expertise, customer experience and brand are also part of the decision. The way a company uses AI can end up influencing how it thinks, decides and relates to others.
What risk does it introduce?
Accountability, traceability, privacy, intellectual property, security and compliance should be part of the analysis.
The level of control required should not depend only on what a system is capable of doing, but also on the consequences if it gets something wrong.
What role do we want AI to have?
AI can assist a person, propose a response or decision, execute actions within defined limits, or operate parts of a process with greater autonomy.
In each case, we will need to decide where we supervise, when we intervene and who remains accountable.
This is not a debate between human or machine. It is a design decision.
Conclusion
Over the past few years, we have spent a great deal of time discovering what artificial intelligence can do. It was necessary and we are probably still only at the beginning.
But the more capable these systems become, the less useful it will be to respond only with a list of tasks they can perform. The question that comes next is more complex: what we want to delegate, under what conditions and with what degree of autonomy.
To answer it, we need to understand very well the work we want to transform. Not only the procedure, but also its exceptions, the relationships it supports, the knowledge it transmits, the judgement it helps to build and the consequences that modifying it may have.
Because a company is not only the sum of its processes. It is also what it knows, how it learns, how it decides and how it relates to others. Many of those assets have taken years to build.
AI can help us use them better, share them, reinforce them and free people from work that no longer makes sense to continue doing in the same way. But before deciding what we delegate, it is worth knowing exactly what we are putting at stake.
Because when we delegate part of the work to AI, we may be delegating considerably more than the work we can see.
The conversation continues
Perhaps one of the paradoxes of this new stage of AI is that the more capable systems become of taking on parts of our work, the better we will need to understand how that work really functions.
Not only what is written down, but also what a company has learned over the years, what its people know without having documented it, and what clients and teams perceive even though no one has ever put it in a presentation.
For a long time, understanding that invisible part of a company has been important in order to understand the company itself. Now it may also be important in deciding how much we want to delegate to AI.
Because before asking what it can do for us, perhaps we need to be much clearer about what we do not want to delegate without realising it.
Frequently asked questions
What should a company delegate to AI?
It does not depend only on what a system can technically do. We need to assess the expected benefit, the knowledge involved, the relationships it may affect, the consequences of an error and the appropriate degree of autonomy for that process.
What is the difference between automating and delegating work to AI?
Automation usually describes the execution of a predefined task or sequence. With today’s AI, we can go further: interpret information, generate alternatives, propose decisions or execute actions depending on the context. That is why talking about delegation also allows us to consider what decision-making capacity and autonomy we are granting.
What is the difference between an AI assistant and an AI agent?
In simplified terms, an assistant helps a person carry out their work. An agent can pursue an objective and execute several actions to achieve it with a certain degree of autonomy. The greater that autonomy, the more important limits, supervision, traceability and accountability become.
Why can AI affect a company’s brand?
Because a brand is not built only through communication. It also emerges from how an organisation decides, responds and relates to others. If AI starts intervening in those interactions, it can also modify the experience that employees, clients and other stakeholders have of the company.
What does adopting AI with a human-centric approach mean?
It means designing the relationship between people and technology based on the result we want to achieve and the value we need to preserve. It does not assume that a person must always intervene, nor that everything that is technically automatable should be automated.
