Asking a Question vs Delegating a Task
Same model, different harness. A chat answers and waits; an agent is given tools and permission to keep going. Delegation moves the steps, not the duty - and a 95%-reliable step is not good enough twenty times in a row.
- وقت القراءة: 13 min
ما يغطيه هذا الفصل
- A Chat Answers, Then Waits
- An Agent Works a Loop
- The Same Task, Run Two Ways
- Tools Are Permissions
- More Steps, More to Check
- Ask the Agent for the Trail
فصول الدورة مكتوبة بالإنجليزية. أما بقية الأكاديمية فمترجمة.
Ask a colleague a question and you have an answer in a minute. Hand the same colleague a bundle and say "turn this into a chronology and flag anything that contradicts the witness statement", and you get a finished piece of work at the end of the afternoon. Both draw on what one person knows. What changes is who does the steps.
AI assistants now work both ways, and the difference is not the model. The model is the same kind of next-token predictor described in Legal AI Fundamentals. The difference is what it is allowed to do between your messages.
A chat answers, then waits
One exchange at a time. You write, it replies, and nothing happens until you write again. It cannot open the next file, check its own citation or save the table. You are the thing that moves the task forward, which is exactly right when the task is one question, and tedious when the task is forty.
An agent works a loop
The same model, given two more things by its harness: tools it may call, and permission to keep going. It decides a step, takes it, reads the result, decides again. It stops when the goal is met, when it is stuck, or when it needs you. Nobody scripts the steps in advance.
You still sign it
The work goes out under your name and your practising certificate, whichever mode produced it. Delegation moves the steps. It does not move the duty, and no tool setting can be configured to move it for you.
The same task, run two ways
Anthropic, which builds Claude, defines an agent as a system where the model "dynamically directs its own processes and tool usage" Anthropic · Building effective agents, Dec 2024 (opens in a new tab). That is a precise claim and worth reading twice: the steps are chosen as the work goes, by the model, not laid out in advance by a developer. Watch where your own attention goes in each column.
The task, both columns: Produce a one-page briefing on the new DIFC Employment Law amendments, with sources.
Step 0 / 10
Ask · chat
- Press Play. Nothing starts on its own.
- Your actions
- 0
- Its actions
- 0
Delegate · agent
- Same task, one brief, then it runs the loop itself.
- Your actions
- 0
- Its actions
- 0
Delegating does not change the model. It is the same reasoning on the same law. What changes is who takes the steps in the middle - and therefore where your attention goes: more of it on the brief at the start and the review at the end, much less of it on the ten exchanges that used to sit between them. The two counters are the whole argument.
An illustrative run, scripted to teach the shape of the loop rather than measured from a live one. The counters count actions, not minutes: one agent step can be seconds, one of yours - reading 94 pages, writing the next instruction - can be several. The ten chat turns are a flattering simplification; a real version of that conversation is longer.
Count the left column again. In a chat you are the integration layer, fetching the document, pasting it in, copying the answer out, fetching the next one. Delegation does not make the model cleverer. It moves your attention to the two places where judgment actually lives: the brief at the start, and the review at the end.
Tools are permissions
An agent can only act through the tools you give it, so the list of tools is the list of permissions, and it is worth reading it that way rather than as a feature list. Three levels matter, and they are not the same risk.
Permission ladder
Three levels, ordered by how hard the damage is to undo
- See
Open files, search sources, read the matter folder. It looks, it reports back.
Nothing to undo - Change
Edit or create files inside a folder you choose. Drafts, not dispatches.
Undo from a backup - Send
Email, file, submit, pay, delete. The action leaves your control the moment it runs.
Hard to take back
Start narrow, widen as trust grows. Nobody earns Send on day one - not a new associate, not an agent.
Good products keep the levels separate, ask before anything that cannot be undone, and let you set how far the agent goes alone before it checks back. When you set one up, give it the narrowest set of tools that finishes the job, and widen from there. A firm-wide version of this decision is what Rolling AI Into a Firm writes down as policy.
More steps, more to check
Delegating moves the risk as well as the work. In a chat you see every answer before you use it. With an agent you see the result, not each step, and an error early on is carried by everything built on top of it.
The arithmetic is unforgiving, and it is the reason "the model is 95% accurate" is a sentence that should never be allowed to stand on its own. Drag the sliders until you find the point where you would be comfortable signing the output.
A checkpoint catches what a reviewer would notice, not everything, and it buys one redo of that stretch rather than an unlimited supply.
36% chance that every step is right with no checks. On average, 1.0 of 20 steps go wrong. Checked no checks: 36% chance the delivered work is right.
- right
- caught at a checkpoint
- wrong, uncaught
- built on that error
- checkpoint
One run, not the average. The percentages above are the whole distribution; this strip is a single draw from it, with the same luck at every accuracy so you can see which failures the slider removes.
20 steps at 95% each come out entirely right 36% of the time, and nothing in the run would tell you which step was the bad one.
An illustration, not a measurement. The model assumes every step fails independently at the same rate, which real work does not: steps vary in difficulty and errors cluster. The checkpoint column assumes a review catches the error and allows exactly one redo of that segment, and it charges nothing for the review itself. Agents that check their own work as they go do better than these numbers; nothing here is a measured HAQQ result.
How long can an agent work unsupervised? METR, an independent evaluation group, measures the length of task an agent can finish on its own about half the time, where length means how long a skilled person would take. Over six years that figure doubled roughly every seven months METR · Measuring AI Ability to Complete Long Tasks, Mar 2025 (opens in a new tab). One caution that matters more for lawyers than for anyone else: those tasks are mostly software, worked alone, with no client, no counterparty and no privilege. Read the trend as a direction of travel, not as a promise about your matter.
Two risks stay yours whatever the agent catches. It acts on what it reads, so a document carrying hidden instructions can steer it. That attack is called prompt injection, and it is taken up in Privilege & Confidentiality with AI. And the output still goes out under your name. So ask for the trail: what it opened, what it changed, what it was unsure about. Read that before you rely on the result.
In practice
Ask when you want to think. Delegate when you want something done.
Ask in a chat
- Explain a rule, a concept or a clause in plain words
- Reword a paragraph, an email or a letter
- Arguments the other side might run
- Summarise one document you have attached
Delegate to an agent
- The same point pulled from across many documents
- Every change between two versions of an agreement
- Gathering authorities for a note, with links to each
- Filling a house template from a matter file
Keep for yourself
- The decision, and the advice you give on it
- Anything sent, filed or signed in your name
- The final check of every authority and every figure
You can delegate the steps. You cannot delegate the responsibility.
Brief an agent the way you would brief a capable new joiner: say what done looks like, give it the right materials, set the limits, and tell it when to stop and ask. The next chapter is about the part that makes that briefing reusable instead of retyped: skills, MCP and plugins.