Chapter 1 · How can AI help my practice and day-to-day work?
00 Introduction 3 lessons
01 How can AI help my practice and day-to-day work? 5 lessons
02 Can I trust it? 5 lessons
03 Am I allowed to use it? 7 lessons
04 How do I choose the best tools? 6 lessons
05 How do we make it work across the firm or legal team? 7 lessons
06 How do I get better results? 3 lessons
07 What does this mean for my career and my firm? 4 lessons
08 Where and how do I start now? 4 lessons
Lesson 1.2 How can AI help my practice and day-to-day work? 7 min read
Asking a Question vs. Delegating a Task
Ask a colleague a quick question, and you have an answer in a minute. Hand them forty receipts and say “turn these into an expense report and flag anything odd”, and you get a finished piece of work at the end of the afternoon. Both draw on what the same person knows. What changes is who does the steps.
AI assistants now work both ways. In a chat, you ask and it answers. An agent takes a goal and works toward it on its own: opening files, searching, drafting and checking until the job is done. The model underneath is the same kind described in lesson 1.1. The difference is what it is allowed to do between your messages.
That difference matters because delegating changes where your attention goes: less on doing the steps, more on the brief at the start and the review at the end.
A chat answers, then waits
A chat works one exchange at a time. You write, the model replies, and nothing happens until you write again. It sees only what you put in front of it, and it can't act on its own answer: it can't open the next file, check a source or save the table. You decide what happens next.
For many tasks, that is exactly right. Explaining a concept, rewording an email, summarising a document you have pasted in, testing an idea: one question, one answer, and you judge it on the spot.
It gets tedious when the task has many steps. You become the go-between: fetching a document, pasting it in, copying the answer out, fetching the next one.
An agent works in a loop
An agent is the same kind of model, given two more things by its harness: tools and permission to keep going. Tools are actions it can take, such as open a file, search a database, write a document or send an email. Permission to keep going means it doesn't stop after one reply.
It then works in a loop: decide the next step, take it with a tool, look at the result, decide again. It stops when the goal is met, when it gets stuck or when it needs you. Anthropic, which makes Claude, describes agents as systems where the model “dynamically directs its own processes and tool usage” Anthropic Building effective agents, Dec 2024 (opens in a new tab). Nobody scripts the steps in advance; the model chooses them as it goes.
Try it
Same task, two ways.
Task: Turn 40 receipts into an expense report.
- You
- 0
- Assistant
- 0
Press play to watch the task unfold.
An illustrative example. Counts are actions, not minutes: one agent step can take seconds, one of yours several minutes.
What an agent is allowed to do
An agent can only act through the tools you give it, so the list of tools is the list of permissions. Three levels matter: what it can see, what it can change, and what it can send. Reading a shared folder is seeing. Editing the files in that folder is changing. Emailing someone outside the firm is sending, and that is the hardest to take back.
Good products keep these levels separate. They ask before any step that can't be undone, such as sending a message or deleting a file, and they let you choose how far the agent goes on its own before it checks back with you. When setting one up, give it the narrowest set of tools that gets the job done and widen from there.
- See Open files, search sources Nothing to undo
- Change Edit or create files in a folder you choose Undo from a backup
- Send Email, submit, delete Hard to take back
Start narrow. Widen as trust grows.
Developers got them first. Legal is next
Software developers were the first to use agents at scale, through coding tools such as Claude Code and Codex. These read a codebase, make changes, run tests and fix what breaks, with the developer reviewing the result rather than every step. That pattern is now moving into other professions.
In a 2026 survey of legal, tax and other professionals, 15% said their organisation already used agentic AI tools, and a further 53% said they were planning or considering it Thomson Reuters AI in Professional Services, Feb 2026 (opens in a new tab). The major legal research platforms now ship agent features, and general AI assistants and office suites have agent modes too. Switch one on and it can work across your files, your browser and your documents, taking many steps in a row without you directing each one.
How long can they work without supervision?
METR, an independent testing group, measures this by asking: how long a task can an agent finish on its own about half the time? Task length is measured by how long it would take a skilled person to do it.
Over six years, that figure doubled roughly every 7 months METR Measuring long tasks, Mar 2025 (opens in a new tab). By May 2026, METR said its tests could no longer reliably measure beyond 16 hours of human work METR Time horizons, May 2026 (opens in a new tab).
One caution: those tests are mostly software tasks where the agent works alone, with no clients, colleagues or real-world documents involved. Treat the numbers as the direction of travel, not a promise about how an agent will handle your work.
More steps, more to check
Delegating moves the risk. In a chat, you see every answer before you use it. With an agent, you see the result, not each step, and a mistake early on carries through everything built on it.
The arithmetic is unforgiving. If each step is right 95% of the time and nothing catches the errors, a 20-step task comes out entirely right only about a third of the time.
Try it
How small errors add up.
Long tasks need checks along the way, not only at the end.
Illustrative: assumes each step fails independently and nothing catches the error. Agents that check their own work do better than this. Lawve analysis.
Good agents catch many of their own errors: they re-read, test and correct, as the agent in the example above did. Two risks remain yours. The agent acts on what it reads, so a document with hidden instructions can steer it, a risk called prompt injection that Chapter 3 covers. And the work still goes out under your name.
So ask for the trail: what it opened, what it changed and what it was unsure about. Read that before you rely on the result.
What this means for your work
Choose the mode by the shape of the task. Ask when you want to think; delegate when you want something done.
In practice
Ask to think. Delegate to get it done.
Ask in a chat
- Explain a concept or a rule in plain words
- Reword an email or a paragraph
- Ideas and counter-arguments
- Summarise one document you attach
Delegate to an agent
- The same points pulled from many documents
- Every change between two versions of a document
- Gathering sources for a note, with links
- Filling a template from a file
Keep for yourself
- The decision and the advice you give
- Anything sent or signed in your name
- The final check of sources and figures
You can delegate the steps, not the responsibility.
Brief an agent as you would a capable new colleague: say what done looks like, give it the right materials, set the limits and tell it when to stop and ask. Chapter 6 covers writing that brief. Next, lesson 1.3 shows how skills, MCP and plugins give an assistant your methods and connect it to your tools, which is what turns a general agent into one that works the way your team does.