idest@lausanne ~ % cat readme.txt
→ A practitioner’s view, not a market survey.
→ Examples are fictitious and built for discussion.
→ Forecasts are hypotheses, flagged as such.
→ AI prepares. A lawyer decides and signs.
Status: open for discussion_
“i.e.” stands for id est, “that is to say”: the two letters used whenever something needs to be made clearer. This paper tries to do exactly that for a question many legal teams are asking at once, sometimes without realising they are asking different things. It was written alongside a course on integrating AI into legal teams, given as part of a CAS at the University of Neuchâtel, and it can be read on its own or as background to that course. It sets out how we think about the subject, where our convictions come from and where they stop.
Our starting point is simple. At id est avocats, adopting AI was never in question. The difficulties described in the second part of this paper are those we hear from peers, in-house teams and public administrations. They are not the story of our own roll-out, and they do not claim to describe the whole market.
01 · Meet Manju
A trainee with a name
Manju is our AI trainee at id est avocats: part of how we organise the work, not a lawyer or an independent decision-maker. Manju has a name, a profile page and a visual identity on our website, alongside the lawyers: Manju | id est avocats. We made that choice mainly for transparency: AI is part of how we work, and we would rather say so openly than act as if it were not. The role is clear: Manju prepares and challenges; a lawyer supervises, decides and remains responsible.
The word “trainee” is deliberate. Manju needs supervision, makes mistakes, and a lawyer always signs off. The slogan on Manju’s profile, “You’re absolutely right”, is self-mockery: it pokes fun at how eager to please these tools tend to be. We find that a little irony helps keep the right distance. A colleague who agrees with everything is pleasant company and a poor reviewer.
idest@lausanne ~ % whois manju
role .............. AI trainee
supervision ....... always
may sign .......... no
may email the team yes (internally)
best at ........... first drafts, research leads, contract mark-ups,
data rooms, internal know-how, playing devil’s advocate
known weakness .... telling you that you are right
What Manju actually does
Manju prepares first drafts and initial legal research, reviews and marks up contracts, and helps with due diligence and data rooms. Manju also supports our internal knowledge, our templates and some of our business development content.
The use that has changed our work most is less obvious. Manju is our sounding board, a kind of legal red team. Before a negotiation, or in the course of litigation, we ask Manju to attack our position: find the weak arguments, play the other side, point to the piece of information that would make us change our mind. The objections are rarely decisive, and often predictable. They nonetheless make our reasoning more robust and our blind spots easier to see.
The work isn’t necessarily faster. It’s tested earlier.
What Manju does not do
The rule is simple: AI can prepare, but a lawyer decides and signs. Accountability is never handed over, and neither is professional secrecy, which shapes which tools we use and how we use them. We delegate tasks, never responsibility. When a client has a question about a piece of work, the answer comes from a lawyer.
Giving Manju a visible place in the team also starts a conversation, with clients and internally, about the question that matters most to us: where does the value of our service actually lie? We return to that question in the fourth part of this paper.
02 · Our questions
What we keep hearing
In our conversations with other firms, in-house legal teams and administrations, a recurring picture emerges. Many of the teams we speak with have taken the first steps: a licence, a usage policy, some training. These are useful and often indispensable. They still leave open the questions that decide whether work actually changes. On the Monday after the training, who works differently? Who checks? Who learns? And what has improved for the people who receive the legal work?
Three questions that sound like one
Discussions about legal AI often feel like a single debate. In practice, three questions coexist. The first concerns the principle: should we adopt AI at all? The second concerns the conditions: how should use be organised, with which rules and which responsibilities? The third concerns the purpose: for which results, and for whose benefit?
These questions do not follow each other along a maturity ladder. Different people in the same organisation can ask them at the same time. One practice group is still debating the principle, management wants rules, and users are looking for something concrete that helps them on Tuesday. Part of the frustration we hear comes from there: two people who seem to disagree are often answering different questions.
Everyone is in favour, nobody has the mandate
Between agreement in principle and a first useful collective use, the path runs through a few decisions that look simple: choose a use case, name a pilot, involve the business, IT, security and procurement where needed, and set aside time and budget. Each step carries its own question: what is the objective, who has the mandate, who is the counterpart, who is available?
This is where many projects stall. Two loops are frequent. In the first, the decision is postponed “to the next meeting” and the discussion keeps returning to the mandate. In the second, individual tests bypass the collective decision. Those tests are valuable, since they build curiosity and experience, but they do not automatically turn into shared practice. They can even create the impression that the matter is settled because a few people are moving ahead on their own.
Rules of use, and then what?
Usage policies set conditions: approved tools, admissible data, access, confidentiality, review. They rarely answer the practical question of the person who wants to start: for this task, with which documents, who checks, and whom do I ask when in doubt? A policy can authorise a use without saying where to begin. We find it helpful to add, for each authorised use, a short operating note: the task, the documents that may be used, the person who reviews and the escalation route.
The cost nobody budgets
The subscription is the visible part of the effort. Below the waterline sit other resources that rarely appear in a budget: time to learn, preparation of documents and sources, review, and the upkeep of methods once they exist. For each of them, three questions deserve an answer: what kind of effort is it, who carries it, and who decides on it? When that time is not planned, it is taken from the evenings of the most motivated people, or it is not taken at all.
A grid we find useful
Structure. The size of the team, the balance between partners, senior lawyers and juniors, the number of layers and the support available. Different pyramids raise delegation, supervision and training in different ways. In-house teams have their own logic, shaped by their place in the organisation.
Culture. Curiosity, appetite for risk, the right to try and to get things wrong. Faced with the same pilot, one team says “let’s test it on a case”, another “what do we need to check?”, a third “who will have the time?”. All three are legitimate. Curiosity has no birth date, and mistaking age for attitude leads to poor decisions in both directions.
Shared project. The ability to agree on priorities, to decide and invest together, and to integrate practices. A stated vision does not guarantee integration, and sharing a tool does not by itself make a project.
03 · Our approach
Three workstreams, one file
We organise our thinking around three workstreams that condition each other. The team: skills, supervision and mandate. The work: sources, hand-offs, review and shared methods. The model and structure: value, resources, pricing and careers. Underneath sit the prerequisites: tools, data, rules and security conditions.
The links between them matter more than any one of them. Skills make review possible, since nobody can organise a check that nobody knows how to perform. Following the file from end to end reveals where the effort really sits. And the business model is what pays for learning: without an economic decision, training and supervision remain good intentions.
The team: learning to check before learning to supervise
AI changes what we expect from lawyers early in their careers. They have to manage machine output, control its quality and develop the instinct to say “this looks right but isn’t”. In a sense they become supervisors much sooner than we did, which makes the job harder. A trainee who can tell a plausible answer from a correct one is worth more than one who produces volume.
You can’t check what you never learned to do.
Our first principle for training
The fundamentals therefore come first. Trainees still need to research, draft and get things wrong themselves, or they will not spot it when the machine gets it wrong. We think it is worth protecting stretches of work without AI, as a deliberate choice: they build the judgement that makes supervision possible later on. A simple exercise works well. Take a fictitious review note that contains three reasonable-sounding statements: the draft received allows us to recommend signature, an annex confirms that the service meets the need, the timetable fits the client’s objective. Each calls for a different check: the basis of the recommendation, the documents behind it (which annex, which version?) and the assumption nobody stated. Fluency in the interface is no substitute for legal judgement. A smooth text can even hide its weaknesses, because it inspires more confidence than it deserves.
Time saved raises its own questions. When a task gets shorter, it is not obvious whether the freed time should go to learning, to the client relationship, to breathing space or to more output. If nobody decides, output usually absorbs it. The question “what skills will be expected of me tomorrow?” is a legitimate question about the future of a profession, and treating it as resistance to change means losing useful information.
Governance, finally, means naming people. A review that is “planned” but assigned to nobody is a familiar risk, and AI makes it more likely, because the output looks finished earlier and responsibility thins out along the chain. We distinguish a few roles: steering the uses, preparing sources and access, producing, validating the result. One person may hold several roles, especially in a small team; each role still needs a named holder. Two further points matter to us. There must be an explicit escalation route, so that an unresolved doubt reaches the person responsible for the file. And it must be safe to say “I don’t know how to check this answer”. One trap is common: making the junior who is most comfortable with AI the implicit owner of everything related to it. That person needs a mandate, time and a senior counterpart.
The work: redesigning the file, not only the prompt
Once roles are clear, the question becomes where AI enters the file, with which sources and which hand-offs. Following the file from intake to delivery shows where AI genuinely improves the work and where it only moves the waiting time elsewhere. A draft produced faster does not shorten the matter if documents are missing at the preparation stage, versions bounce between drafting and review, or sign-off is pending before delivery. The measure that counts is total lead time, review included.
A contract review before negotiation illustrates the division of labour we aim for. The lawyer frames the objective, the priority risks and the scope. Preparation gathers the reference version, the annexes and the instructions, and AI can flag what is missing. In drafting and challenge, AI proposes changes and objections and lists the points that still need checking. Validation and the advice to the client remain with the lawyer: which positions to take, which risks to accept, which recommendation to give. This is an organisational target to adapt case by case, and it says nothing about guaranteed technical capability.
Quality deserves the same discipline. In our experience, framed, supervised and iterative use can improve the result, and not only make it faster. We present this as a practitioner’s conviction, not a statistic. It is not automatic either: without framing and selection, the main effect is more documents and more proofreading. A useful iteration addresses an identified defect: context clarified, objection checked, inconsistency fixed, recommendation sharpened. A further rewording with no identified defect usually adds review work and little else.
That observation led to a small piece of fiction we like to share. A colleague keeps asking Manju for “another version”, then “not quite, try again”, then “one more”, without saying what is wrong. Manju replies:
“Three versions, and still no defect identified. I’m part of the team, not the merry-go-round. Your call: what needs to change, and why?”
Manju, in a fictitious dialogue
The comedy lies in the reversal. The serious point is that if we treat AI as a member of the team, we can also ask it to tell us when we are going round in circles. Building that into working instructions is a rule to design and test; we do not claim it as an autonomous capability. The signal does not decide on quality or on when the work is finished. It prompts a human decision: freeze a reference version if the need is met and the checks are done, correct and re-check if a precise defect has been found, or take the work back, complete the sources and revisit the scope with the person responsible. No necessary check disappears along the way.
Last, knowledge. A method that lives only in one person’s private chat history leaves with that person on Friday afternoon. We try to share methods rather than prompts alone: a short method sheet stating the use and its scope, the sources and reference version, the expected result, the checks and limits, and the person responsible with a review date. Start with a frequent, well-bounded use. Before the pilot, define what would count as an improvement and record how comparable work is currently done. Compare total lead time, human preparation and review effort, rework, and the nature and seriousness of defects found, rather than the volume of output. Faster drafting alone is not success if review takes longer or important errors go unnoticed. Set a date for the person responsible to decide whether to continue, adjust or stop, and record the reasons. Only then should a promising test become a shared method.
04 · Model, pricing, structure
Where does the value lie?
This is the part of the subject we find most interesting and least settled. Changes to people and processes have economic and structural consequences, for law firms and in-house teams alike. We set out below what we observe, what we are trying and where we remain unsure.
Time, value and the price of a matter
For a given mandate, three situations should be kept apart. In the first, AI lets us deliver the expected result with less effort. The need is met and the checks must stay in place; the open question is whether, and how, the benefit is shared with the client. In the second, AI helps deliver a better result, for example a decision that is better informed, at the cost of additional, targeted effort. That effort needs to be agreed and the scope adjusted. In the third, AI produces more versions with no identified benefit. The usefulness is nil, review time grows, and someone has to call a halt.
| Situation | Value for the client | Human time and review | Pricing question |
|---|---|---|---|
| Expected result, less effort | Need covered | Checks preserved | How is the gain shared? |
| Better result, justified effort | Better decision | Targeted effort to agree | Which scope and which price? |
| More versions, no identified gain | None identified | Extra review | Who decides to stop? |
A discussion grid with no figures: it qualifies trade-offs and does not measure them.
Hourly billing sits awkwardly with all three. When fees are fixed or capped and internal costs rise without compensation, the margin shrinks. When fees are hourly, efficiency gains reduce the bill, which is good for the client and raises a fair question for the firm about how to fund the investment that made the gain possible. Tool costs add a further layer. They depend on the contract with the provider (subscription, quotas or pay-per-use) and do not behave like human time, so they cannot simply be billed in the same way.
Our current approach rests on a few principles rather than a formula. We discuss the expected result and the scope at the outset. For larger matters we set an effort budget with a decision point, so that additional work is a choice and not a drift. We explain what the price covers and what the client gains. And we are open about our use of AI. Clients now expect us to use it, and to be good at it. What they want to know is that there are proper guardrails and real quality control, and that we remain fully responsible for the result.
Clients expect us to use AI, and to be good at it. They also expect a lawyer to answer their questions.
What we hear from clients
What we are unsure about
We do not yet know how far clients will pay for better quality, as opposed to faster delivery. Many clients press on price; fewer press on how the work is delivered. We also do not know which share of the gains should go to clients, which to the team and which to reinvestment in methods and training. Initial work on a reusable method is an investment, an improvement to a file is a service, and extra versions with no identified gain are a cost. Telling the three apart in real time is harder than it sounds.
A flatter pyramid, a deliberate apprenticeship
Our working hypothesis is that the traditional leverage pyramid may flatten: fewer juniors per partner, with more responsibility, sooner. A firm of our size may be able to redesign its structure and training with fewer coordination layers. Larger firms face different constraints, but may also have more resources to invest in change. Size alone tells us little about the willingness to adapt or the expectations of clients. This remains a practitioner’s hypothesis, not a forecast of employment in the profession as a whole.
A flatter structure does not solve the learning question by itself. Juniors have traditionally learned through repeated exposure to reviews, research and preparation for negotiations. If AI takes over part of that work, the learning opportunities must be deliberately replaced, with time set aside for supervision. Juniors should be encouraged to ask for the hard file, the client call and the negotiation. But access to those opportunities cannot depend on confidence or persistence alone. The firm must organise it, assign supervisors and check that learning actually happens.
If AI takes over the repetition, the firm must redesign the apprenticeship.
Our hypothesis on careers
We find it more useful to reason in skills than in headcount. For each skill a lawyer must acquire, we ask three questions: in which real situation can it be learned, who supervises, and what shows that it has been mastered? Access to the file, to the client and to the negotiation table, independent analysis and critical feedback count for much more than familiarity with an interface. In practice this means choosing which fundamentals to protect, organising exposure to formative situations, recognising supervision as real work, and adapting roles before fixing headcount.
Differentiation when everyone has the same tool
When access to the tools is no longer scarce, a subscription does not create a lasting difference. What clients look for is an informed decision, trust and reliable execution. Two files handled with the same tool can look very different depending on what surrounds it: knowledge of the client’s context, the way options are framed, the quality of review, and support in implementing the decision. A recommendation that fits the situation requires knowledge of the context and shows in clearly stated options. Trust requires real responsibility and control, and shows in a decision someone stands behind. Reliable execution requires shared methods and knowledge, and shows in clean hand-offs and follow-up.
So-called AI-native law firms are part of that landscape. They feature in our own SWOT analysis, and more will come. Being designed around AI does not make a firm automatically better, any more than a long history makes it automatically wise. We expect the winners to be neither the firms with the best rankings, nor those with the most fee earners or the biggest budgets for impressive tools, but those that rethink how they operate and what their people need to be good at, especially management and quality control. Choosing a difference also means choosing what we do not want to become.
05 · What lies ahead
Our bets, and our doubts
We were asked to say where we think this is going. We will try, with two caveats. These are bets made by a small firm on the basis of its own practice, and they may prove wrong. And the pace of change makes any forecast fragile, including ours.
Five bets we are willing to make
-
Review becomes a core professional skill, and a selling point. Producing a first draft will matter less than knowing how to check one, explain the checks and say when something cannot be verified. Firms will increasingly be chosen for the quality of their control.
-
Pricing will split along the nature of the work. For repeatable, well-bounded work, we expect prices to move towards scope and outcome. For work dominated by judgement and uncertainty, time-based billing will probably remain, with more transparency about what the time is spent on.
-
Teams will become smaller and more senior-leaning, with apprenticeship by design. The firms that keep training lawyers well will be those that organise learning explicitly, protect formative work and treat supervision as part of the job.
-
Methods and knowledge become firm assets. Shared, maintained methods will count for more than individual prompting talent. This also raises a question every firm should ask: what stays with the firm when a key person leaves?
-
Clients will keep asking for a person. They will expect AI to be used, and used well, and they will keep wanting a lawyer who answers for the result and explains it in their context.
The first four bets are also about people. Done well, they can give lawyers more interesting work earlier and more time for clients. Done badly, they can produce teams that supervise without having learned, and partners who review more and decide less. Which way it goes depends on choices made inside each organisation.
What we do not know
We do not know how the rules on professional secrecy, data protection and the use of external tools will evolve, or how quickly. We do not know how dependent firms will become on a small number of providers, or what that will cost over time. We cannot yet quantify the environmental footprint of our own use in a meaningful way, and we prefer to say so rather than quote numbers without a source. We do not know how the next generation of senior lawyers will have learned their craft. And we do not know whether clients will reward better quality, or only faster delivery. These doubts do not stop us from moving. They are the reason we try to move with a method, and to say what we are doing.
06 · Monday morning
Three decisions to take together
Giving AI a place in a legal team only makes sense if the team, the file and the organisation change together. In our experience, three decisions carry most of the weight, and each organisation will take them differently depending on its structure, its culture and its shared project.
-
A skill to protect. Identify one fundamental skill that AI could erode, such as writing a first analysis or checking a source, and organise the situations in which it is learned and kept alive.
-
A process to redesign. Pick one frequent, well-bounded use, follow it from intake to delivery, name the roles and the checks, and turn it into a shared method.
-
A trade-off to own. Decide explicitly how freed time is used, how gains are shared with clients or internal stakeholders, and what resources go into learning.
For the first Monday, three verbs are enough. Stop one habit that no longer helps, such as an iteration with no objective or a review that nobody owns. Test one framed use, with a mandate and feedback. Protect one skill, one learning moment or one client relationship that AI should not absorb. Decisions of that kind, taken locally and concretely, do more than a general commitment to innovation. The question of the organisation we want to build stays open, and we would be glad to discuss it.
Acknowledgements
We thank the participants of the CAS at the University of Neuchâtel for the discussions that shaped this paper. We also thank Manju, our AI trainee, for first drafts, critical re-reads and a steady supply of objections, delivered with the eagerness we have come to know. In keeping with Manju’s status, Manju signed nothing. The choices made here, and any errors, are ours.
To go further
-
Professional guidance. Ordre des avocats vaudois (OAV), Guide sur l’utilisation de l’IA générative, adopted by its Commission de modernisation in June 2026: oav.ch. A practical starting point for understanding and using generative AI in legal practice. Accessed 19 November 2026.
-
A foreign perspective. The Law Society of England and Wales, Using artificial intelligence in legal practice, practice note, 21 October 2026: lawsociety.org.uk. A comparative checklist covering AI governance, training, supervision and confidentiality, written for solicitors in England and Wales rather than Swiss practitioners. Accessed 19 November 2026.
-
Research and analysis. Harvard Law School, Center on the Legal Profession, AI & the Legal Profession, a continuing essay-and-visualization series led by Anthea Roberts and David B. Wilkins: clp.law.harvard.edu. Useful for examining the links between professional structure, training and judgement; an evolving analytical programme, not a single empirical study. Accessed 19 November 2026.
-
An investor’s view. Ben Horowitz and Michael McGuiness, Databricks & the Future of the Firm, Andreessen Horowitz, 27 October 2026: a16z.com. A technology-investor perspective on shared organisational knowledge, delegation to agents and the changing work of management. A useful provocation, not a neutral account or evidence specific to law firms. Accessed 19 November 2026.
This paper is provided for general information. It is not legal advice and does not replace advice tailored to a specific situation. Examples are fictitious. © id est avocats Sàrl, Lausanne, 2026. Except where otherwise indicated, this paper is licensed under the Creative Commons Attribution–NonCommercial–NoDerivatives 4.0 International licence (CC BY-NC-ND 4.0): https://creativecommons.org/licenses/by-nc-nd/4.0/. Third-party materials, logos and trademarks are not covered by this licence.
