The 3 Modes of AI Use and Which One to Avoid
How you work with AI matters more than how much you know about AI. The mode to watch out for is the one where you review the AI’s output closely but look at the wrong things.
Overview
A July 2026 Harvard Business Review article, Shaping Early-Career Success in the Age of AI, by Agarwal, Barua, Puvvada, Song, and Wen describes a KPMG and UT Austin field study that I think is particularly relevant to how law firms are rolling out AI. The researchers asked 523 early-career professionals to complete realistic client work with an AI agent, had experts grade the results, and then compared each person’s work against the AI’s output.
The authors identified three groups in that comparison: amplifiers, delegators, and apprentices. Amplifiers (50.1%) outperformed the AI working alone by framing the problem, grounding the AI in the right materials, interrogating the output, and then deciding and recording their reasoning. Delegators (25.8%) roughly matched AI by handing over the inputs and accepting whatever looks right. Apprentices (24.1%) actually performed worse than the AI on its own, mostly because they would engage with and critique the output, but they tended to focus on the wrong layer, such as polishing the prose without addressing latent defects.
The surprising part for me was that the Apprentices were not the least skilled group. The Apprentices generally matched the Amplifiers on foundational skills and scored higher than the Delegators, yet still performed below the AI baseline. In other words, what separated the groups was mostly how each person worked with AI. The good news is that the amplifier interaction method can be taught and fostered at a law firm. The bad news is that delegation can look a lot like success, which makes it a real risk for any law firm. I suspect that all three profiles exist at most law firms and at every level of seniority.
AI Issues at All Levels of Experience
These profiles do not seem to decrease with increasing experience, but each profile does look different at each level. For example, a clerk may generate AI-drafted claims from a disclosure without ever developing foundational drafting skills. An early-career attorney may accept an AI-generated office action argument because it sounds right. A mid-career attorney may submit one brief prompt to AI and then proceed with the first AI response. A senior attorney may review AI-assisted work as if it were a junior associate’s draft, but then miss the fluent-but-wrong parts that a junior associate is unlikely to have produced.
Junior attorneys may not yet have the foundation needed to evaluate the output, since amplification requires something to amplify. A clerk probably cannot evaluate a claim set that the clerk could not have written with their level of experience. If AI is introduced too early, you may end up with a delegator who has no clear path out of delegation, along with a “learning debt” that leaves that attorney behind similarly situated peers.
Experienced practitioners have a different problem. Experienced practitioners often have good judgment regarding patent strategies, but that experience may not automatically translate to working with AI. Importantly, the failure modes of AI differ from those of an associate. AI may fabricate citations, misinterpret rules or caselaw, or quietly amend claim language while smoothing out the prose. For junior attorneys, the priority should be building the foundation first. For experienced practitioners, the priority should be recalibrating how they direct and review AI-assisted work, rather than remediating their legal skills.
A Phased Roadmap
The role of AI should evolve as an attorney’s capabilities change, always keeping in mind our professional responsibilities. I propose a career training framework with four training objectives. The first objective is to learn patent law fundamentals in general without AI, which is where all law clerks should start. The next objective is to evaluate AI that may be used by early-career attorneys in parallel with their ongoing training. The following objective is to
collaborate with AI, which mid-career attorneys may use to test and amplify their experience. The final objective would be to supervise AI-assisted work, which may be used by more senior attorneys in a manner similar to working with a junior associate. The common goal across all four objectives is for the attorney to produce work that exceeds the AI tool baseline. Importantly, the attorney should be able to explain why their work is better than the AI tool baseline, for example by identifying and documenting areas where their strategy diverged from the AI tool output.
One Method for Everyone
For clerks and early-career attorneys, the roadmap may be separated into four phases. The junior attorney may progress between stages based on demonstrated capability rather than months or years of experience.
- Phase 0 (Foundation): AI stays off while the clerk or attorney practices drafting claims, building office action strategies, writing specification sections, and reading references unaided.
- Phase 1 (Guided): AI is introduced under supervision, and attorneys use the five-step recording method described below and keep a decision log starting with the very first session.
- Phase 2 (Practicing): Routine AI use begins, alongside “beat-the-baseline” assignments where the review covers the attorney’s method and the substance of their work product.
- Phase 3 (Coaching): Attorneys use AI fully, review the AI methods of other attorneys, and run clinics. This phase is intended to be an ongoing firm role rather than a graduation.
A five-step recording method may be useful in Phase 1 or at any stage of training attorneys with AI. This five-step recording method is tool-agnostic, so it should work across Arty (our in-house AI), Claude, Cowork, Copilot, and whatever comes next. Here is the five-step recording method:
- Frame: Before you prompt, define the problem, define constraints, and define what a good answer looks like.
- Ground: Place the right materials for the task into the AI context window, such as the disclosure, the cited art, the file history, client standards, or governing legal framework.
- Interrogate: Treat the output as a hypothesis. Challenge the hypothesis, ask for reasoning, and verify any cited authority.
- Decide: Accept, change, or reject the output based on articulated criteria.
- Record: Write down the articulated criteria and any judgment calls where you overrode the AI, caught a gap, or redirected scope or strategy, along with your reasoning.
As practitioners transition into experienced practitioners, additional practices may be helpful. One practice may be to hold practice-group recalibration sessions that cover the three profiles (amplifier, delegator, apprentice), the delegation trap, and real failure modes. Experienced practitioners may find it helpful to identify which types of preparation and prosecution tasks are best suited to AI assistance, and to use task playbooks to map where AI helps, how to avoid AI traps, and which tasks should generally be performed without AI. Practitioners may use decision logs to capture two or three lines on the decisions that matter when interacting with AI. Some practitioners may also benefit from peer method reviews, which may include a fifteen-minute walkthrough of how a practitioner directed an AI tool on a particular task. The idea is that we are not teaching prosecution to experienced practitioners, but instead teaching them how AI fails and how to avoid those failure modes.
So Which Mode Should You Avoid?
The study sorted people into three groups, but we have found that AI experience and the particular task should decide which mode to use. For example, depending on the particular task, you may be able to generate high-quality work product using the amplifier or delegator mode.
Amplifier mode is the right choice for office action strategy, claim scope, complex inventor comments, and foreign filing decisions, where the judgment is the actual work product. For those tasks, you structure the task, direct the AI, set the criteria, and refine the work product over several rounds.
Delegator mode is the right choice for tasks like filing attachments, indexing folders, first-pass inbox triage, and formatting. In each of these tasks, the work is reliable, low-stakes, and easy to check. Delegation itself is not the problem, but delegating the wrong task can be a problem.
The apprentice mode is the mode to avoid. Unfortunately, reverting to apprentice mode often happens by accident. The apprentice mode often shows up whenever you focus on what is easy to see rather than what matters. I have seen this often when an AI tool provides a primary recommended strategy for an Office Action response, but that response includes amendments that change the claims substantially beyond what is needed to overcome the rejections. The best way to avoid this is to start by deciding what you are checking for in the AI output before you start reading the output.
Webinars
My colleague Nathan Elder and I presented a webinar, “AI Skills: A Phased Roadmap for Junior and Senior Attorneys.” In that webinar, we walked through choosing the right tool for each task, deciding when to delegate and when to amplify, grading the method and not just the document, and launching a contained training pilot to help your team turn AI into a lasting advantage. If you want to see how this framework applies specifically to patent attorney training, my colleague Jeffrey Cobia and I also presented “Training Patent Attorneys in the Age of AI” on October 27th. The recordings and slides for both webinars are available on demand:
- Training Patent Attorneys in the Age of AI
- AI Skills: A Phased Roadmap for Junior and Senior Attorneys
Which mode do you find yourself in most often? Has your team found a good way to catch Apprentice mode before the work reaches the client?
This article is provided for general informational purposes and does not constitute legal advice.