The unglamorous truth
If you believed the launch videos, an agent would run your business while you sat on a beach. If you believed the sceptics, it cannot reliably count to three.
Both are wrong in the same way: they are arguing about agents in general when the only useful question is about specific jobs. Some jobs are a genuinely good fit right now. Others are not, and will not be for a while. Knowing which is which is the most valuable thing in this course.
Here is the honest list.
One: research and summarising
This is the strongest use, and it is not close. An agent can look at twenty sources, thirty emails or a folder of PDFs and hand you back the shape of it. Not a verdict, a shape.
- Read the last 40 support tickets, group the complaints by theme
- Summarise a 60 page supplier contract into the 10 things that
affect us, with the page number for each
- Check what five competitors currently say about pricing, in a table
- Go through this year's board packs and pull out every commitment
we made with a date attached
- Read these 25 survey responses and tell me what the quiet
minority is saying, not just the loud majority
Why this works: the output is text you read, judgement stays with you, and if the agent misses something the cost is that you go and look yourself.
Two: sorting and routing
Anything shaped like putting things in the right pile. Support tickets by topic and urgency. Job applications against stated criteria. Expenses into categories. Inbound enquiries to the right person.
Agents are good at this because the rule is usually written down somewhere and the volume is what makes it miserable for a human. Sorting three hundred items is not hard, it is just long, and being long is precisely what makes people careless by item ninety.
Sorting is also where you get the clearest measure of whether it is working. Have it sort a hundred things you have already sorted yourself, then compare. You will know within twenty minutes whether to trust it, and you will know exactly which categories it fumbles.
Three: drafting from several sources at once
Not writing from nothing, which is where AI gets bland, but assembling. The monthly report that pulls from four systems. The client update that needs last month's notes, this month's numbers and the open action list. The handover document.
An agent that can read all four places and produce a first draft with the numbers already in the right cells saves an hour of copying, and copying is where humans introduce most of their errors anyway.
In plain English
- Triage:
- Sorting incoming work by type and urgency so it reaches the right person quickly.
- Multi-step task:
- A job with several stages where each stage depends on what the last one found.
- Human in the loop:
- A person reviews or approves before anything irreversible happens.
- Reconciliation:
- Checking two lists against each other and flagging what does not match.
Four: repetitive multi-step admin
The jobs made of eight small steps, none of them difficult, all of them tedious. New starter setup. Chasing five suppliers for the same missing document. Reconciling a delivery list against an invoice list and flagging the gaps.
TASK Weekly delivery reconciliation
- Open this week's delivery log
- Open the invoices received folder
- Match each delivery to an invoice by reference number
- Flag any delivery with no matching invoice
- Flag any invoice with no matching delivery
- Flag any pair where the amounts differ by more than 2 percent
- Write the three lists into a summary document
- Stop. Send nothing. Tell me what you found.
Time for a person: most of a morning, every week.
Time for an agent: minutes, and it never gets bored on row 200.
Five: coding assistance
Worth mentioning even though you may never write a line of code, because it is where agents are furthest along. Software people now routinely hand over whole chunks of work.
The reason is instructive: code can be run, and running it tells you immediately whether it worked. That instant feedback is exactly what agents thrive on, and it explains why the rest of us are further behind. Most office work has no equivalent of press run and see if it breaks.
Checkpoint
Agents do best on work that is high volume, low judgement, and easy to check. When you can verify the output quickly, the occasional mistake costs you nothing much. When you cannot, every mistake is a surprise later.
The pattern underneath all five
HIGH VOLUME Enough repetition that a person gets bored and sloppy
LOW JUDGEMENT The rule is written down, or could be
EASY TO CHECK You can spot a wrong answer in seconds, not weeks
CHEAP TO UNDO A mistake means redoing work, not phoning a client
Score your task out of four.
Four: hand it over today.
Two or three: hand it over with a human check before anything final.
One or zero: keep it. Genuinely.
โ Weak prompt
Prompt
Research our competitors and tell me what you find.
Output
The competitive landscape is dynamic and evolving. Several players are positioning around value and innovation. Key themes include customer experience and digital transformation...
No named competitors, no specific question, no format. You get consultancy vapour. The agent had nowhere to go, so it went nowhere expensively.
โ Good prompt
Prompt
Check the pricing pages of these four companies today: [names]. For each, give me the entry price, what the cheapest plan includes, and anything they charge extra for. Put it in a table. Include the page link for each row so I can verify it. If a price is not published, write not published rather than guessing.
Output
Table of 4 companies with entry price, plan inclusions, add-on charges and source links. Two prices not published, marked as such.
Named sources, specific fields, a format, and explicit permission to say I do not know. That last instruction is what keeps the table honest.
Notice what is missing from the whole list above. Nothing here involves an agent deciding something about a person, spending money on its own, or speaking to a customer unsupervised. Those are not on the list because they are not ready, and the next lesson explains exactly why.
๐ Quiz
Question 1 of 4Which kind of work do agents handle most reliably today?