You get the best out of an AI agent the way you'd get the best out of a new hire: give it a narrow job described in one sentence, your real documents and exceptions, clear rules on what needs your approval, and a weekly review for the first month. Better documents and examples improve an agent far more than a newer AI model does.
None of this needs technical skill. It's groundwork that people tend to skip.
How do you brief an AI agent?
Start with a single sentence. If you can't describe the job in one, the agent won't understand it either.
A clear job:
"Every Friday at 4pm, go through this week's meeting notes in the shared folder, list each action I agreed to with its deadline, and draft the follow-up emails for me to check and send."
Too vague:
"Help with admin."
The vague version gives the agent no idea where its job starts or stops, so it fills the gaps with guesses, and you may not notice until it's doing things you never asked for.
Then add two lines: what "done" looks like, and what it must never do. Those three lines are the agent's brief.
- The job: the sentence above.
- Done means: by 5pm on Friday, every meeting from the week has an action list and a draft follow-up waiting for me.
- Never: send anything itself, or commit me to a date I didn't agree in the meeting.
What does the agent need to know about your firm?
AI models know a great deal about the world and nothing about your firm until you tell them. Most of your setup effort belongs here.
- Your documents. Old letters, templates, standard clauses, past proposals. Use the versions you really send, imperfect ones included.
- Examples. Two or three pieces of work you're happy with, and one you're not, with a note on what's wrong with it.
- Your vocabulary. What you call things inside the firm, and what clients call the same things.
- Your tone. Formal or warm, first names or titles, short or detailed, British or American spelling. Write three sentences that sound like you.
- Your exceptions. The client who must always be copied in, the type of matter you never take on, the fee question that always goes to a partner.
Exceptions are where your judgement lives, and they're the part people forget. An agent that doesn't know them will do the wrong thing with complete confidence.
Which job should it start with?
Pick the safest job that still saves real time, rather than the most impressive one. A good first job:
- repeats at least weekly
- has a clear end point, so you can tell when it's finished
- is cheap to fix if it goes wrong (a wrong draft is fine; a wrong payment isn't)
- is something someone on your team dislikes doing
That last point pays off more than you'd expect. The person who hated the task will tell you, in detail, what's working and what isn't, so you learn quickly.
Leave these out of a first job: anything involving money, anything that reaches a client without a person checking it, and anything whose rules exist only in someone's head.
What should need your approval?
Draw the line between what the agent does alone and what it prepares for you before it starts. The simplest test is how easy the action is to undo.
- Easy to undo (drafts, summaries, tags, internal notes): it does these alone, and you check afterwards.
- Hard to undo (sending to a client, filing a document, changing a client record): it prepares, and you approve and send.
- Can't be undone (payments, deletions, legal filings): it never does these alone.
Our guide to AI agent safety uses the same three levels.
Write the rules down on day one. Firms that only decide after something goes wrong usually overreact, require approval for everything, and end up with an agent that saves no time at all.
How do you know it's working?
Many firms measure nothing and then argue about whether the agent is any good. Track two numbers instead:
- Hours saved per week, after subtracting the time you spend checking its work.
- Redo rate: out of every ten things it produces, how many did you have to change significantly?
Hours saved tells you whether it's worth keeping. The redo rate tells you whether it's getting better.
Our rule of thumb is to aim for fewer than 2 in 10 needing big changes by the end of the first month. If the rate is higher, the cause is nearly always missing knowledge: a document, an example or an exception the agent was never given. Fix that before you blame the model.
The first month, week by week
Setting up an agent is closer to training a new team member than installing software.
- Week 1: let it run, check everything, and note every correction you make.
- Week 2: give it the missing information those corrections revealed, usually a document or an exception.
- Week 3: review the approval steps and remove any that no longer earn their place.
- Week 4: stop checking everything and check a sample instead.
After that, a monthly review is enough. The most common mistake is to stop improving after week one and call the agent "fine" or "not working" on the strength of five examples.
Where should the agent stop?
Some decisions should stay with people, and spelling that out for the agent and the team makes it more useful, because everyone knows where it will hand back.
- pricing and fee discussions
- deciding whether to take on a client
- anything that carries a professional or legal obligation
- apologies and bad news
- the final check on anything that goes out under the firm's name
Agents will also meet cases they can't handle. Decide in advance which of these applies:
- Go ahead when the case is clear and the action is easy to undo.
- Prepare and flag when the case is clear but the action is hard to undo, so a person approves it.
- Stop and ask when something is unclear, information is missing, or two rules contradict each other.
"Stop and ask" needs a named person, or a shared queue someone checks every day. A question nobody sees is worse than having no agent, because everyone assumes the matter is in hand.
Mistakes that sink first agents
- Automating a process nobody has written down. The agent inherits all the existing confusion.
- Starting with the hardest task. It struggles, and the team loses confidence early.
- Not showing it what good work looks like. It produces work that sounds right and isn't.
- Leaving out the exceptions. It handles the routine cases well and gets the important ones wrong.
- Requiring approval for everything. The approver becomes the bottleneck, and nothing gets faster.
- Requiring approval for nothing. Something wrong goes out, and the team stops trusting it.
- Measuring nothing. Every discussion about whether it works comes down to opinion.
A one-page setup checklist
- The job is written in one sentence
- "Done" is defined
- "Never do this" is defined
- It has your real documents, not tidied-up versions
- It has at least two good examples and one bad one
- Your tone is described in three sentences
- There's an up-to-date list of exceptions
- Every action is sorted into: do alone / prepare for approval / never
- A named person handles anything it's unsure about
- Hours saved and the redo rate are being tracked
- A weekly review is booked for the first month
Common questions
What should I do when the agent keeps making the same mistake?
Add the missing piece to its standing knowledge rather than correcting it in the moment. Repeated mistakes usually mean it lacks an example, an exception or a clear rule. Write that down once, where the agent will use it every time; correcting your agent covers how.
Will a better AI model fix a weak agent?
Rarely. A stronger model writes more fluently, but it still doesn't know your templates, your clients or your exceptions. When an agent underperforms, check what it was given before you switch models, and use choosing an AI model when you do need to compare.
Can one agent handle several jobs?
Yes, but add them one at a time. Get the first job stable, with a redo rate below about 2 in 10, before you give it a second. Adding several at once makes it hard to tell which brief is causing which problem.
How much of my time does an agent need after the first month?
Mostly a short monthly review plus spot checks on a sample of its work. Expect extra time whenever your process changes, a new type of work appears, or you add a job. Supervision falls as the setup improves, but it doesn't reach zero.
Part 2 of 3. Read next: How AI agent safety works. Previously: AI usage vs an AI agent.
We set up your first agent with you, using your own documents, and stay involved for the first month of improvements. Book a 30-minute diagnostic.