An agent is a step inside a workflow, not a chat box in the corner of the screen. The model prepares a draft, a person approves it or rewrites it. The agent proposes, the decision stays with the employee. That is how the bot answering customer reviews and questions works in production: it takes the routine off people, and it sends nothing on its own.
Where an agent earns its place
An LLM covers a narrow class of work well. The signs are simple, and they are worth checking before any development starts.
The task repeats
The same job dozens of times a week. A one-off is cheaper to do by hand than to automate.
The wording is routine
Answers differ in detail, not in meaning. There is a tone to hold, and the tone can be written down as rules.
A bad draft is cheap
A weak draft is seen by a person and fixed in a minute. Nothing leaves the warehouse and no money moves.
Somebody is there to approve
The process already has a person who reviews the output. The agent goes into their screen, not in place of them.
Where a model gets in the way
Some tasks look like a fit and fall apart on the first read-through. They need a reproducible answer, and ordinary code is what gives you one.
Numbers and money
Cost of goods, commission, payroll, tax. The figure has to add up and come out the same on every run, and a model offers no such guarantee.
Legal text
Contracts, terms, a reply to a formal claim. Only a lawyer can sign off on the wording, and they will rewrite it themselves anyway.
Same input, same output
If the result must repeat, it belongs in a rule with a test next to it. A rule can be read and repaired. A prompt can only be rewritten and tried again.
Decisions without a human
Discounts, cancellations, refunds, account blocks. An agent can prepare the draft, but it does not press the button.
What the work includes
Scope depends on the task, but the skeleton is the same. Below is what an agent needs to survive its first months in production.
Picking the entry point
We look for the step where text is typed by hand every day. Some candidates end up as plain code after the teardown, and that is a fine outcome.
Draft and approval
The answer waits in the interface with its topic and a confidence score attached. The send button is pressed by a person.
Rules before sending
Business limits are enforced before the text leaves, not after: no promised discounts, no competitor names, no empty or bloated reply.
Cost and latency
Price per request and response time are worked out up front on real volumes. Repeats come from a cache, simple steps go to a cheaper model.
Logging
The input, the draft, the human edit and the final text are all written down. Without that record a complaint cannot be investigated, only believed.
Honest failure
When the model API is unavailable, the interface says so and opens a manual field. The customer never receives a silent fallback template.
How the work runs
- Teardown of the task. I look at where the time actually goes and which text is typed by hand. Cost per request and monthly volume are estimated here, so the conversation is about money rather than impressions.
- Rules and reference answers. We collect the hard business limits and a dozen answers that are good enough to copy. That is cheaper than training a model of your own and far easier to maintain: the rules sit in plain text and can be read.
- Drafts in the interface. The agent goes into the screen the team already works in. Every correction is stored and fed back into the prompt, so drafts drift towards what the person would have written anyway.
- Watching it after launch. Spend, response time and the share of edited drafts are all visible. If the edit rate does not fall, the entry point was wrong, and saying so early beats polishing a prompt for months.
What it looks like in practice
Each system has a detailed teardown and a clickable demo on fictional data. Client names are withheld, the numbers are real.
Common questions
What is an AI agent in a business workflow
It is a step inside a process you already run, not a chat window bolted onto the side. The agent reads data from the system, prepares a draft and hands it to a person for approval. Every run is logged: the input, the draft, the edit and the text that finally went out.
Which tasks are worth handing to an LLM
Tasks where the wording is routine, the work repeats, and a weak draft costs a minute to fix. Automated review replies fit almost perfectly. Calculations, money and legal wording do not fit, because they need a reproducible result, and that is a job for ordinary code.
How much does running the model cost
Cost per request and response time are estimated before development, on your real volumes. A cheap model handles the simple steps, repeats come from a cache, and a long prompt is compressed into rules plus a few reference answers. Skip that work and the provider invoice arrives as a surprise at the end of the month.
How do you know the agent did not write nonsense
The draft is checked against business rules before it is sent, not after. Low confidence drafts move to the top of the queue and are excluded from any bulk send. The log closes the loop: for any complaint you can see the input and the exact text the customer received.
Can an agent be added to a system we already have
Usually yes, if there is a place to read the data from and a screen where somebody approves the result. The agent belongs inside that screen rather than in a separate service with its own login. If there is nobody to approve the output, the task is better left alone.
If there is a repetitive writing job somewhere in the company, write a couple of lines about it: LinkedIn or Telegram. What comes back is questions about the work, not a pitch deck.