Good people, doing work that never needed a person.

Copying the same details from one system into another. Retyping what a customer already told you. Confirming the same appointment three times. Chasing the same document from the same supplier every single month. Writing the fourth version of an email you have written four hundred times.

None of it needs judgment. All of it needs somebody, so somebody does it, and that somebody is usually one of the people you would rather have talking to customers or running the floor.

And because it is repetitive, it goes wrong. Not dramatically. A number typed one digit out. A confirmation that went to the wrong person. Something one person assumed another had handled, so nobody did. Repetition is where errors come from, and the errors are almost never in the interesting work. They are in the part everybody is doing on autopilot at four in the afternoon.

Then it happens again next week, in exactly the same shape, because nothing about the process changed.

Here is what that looks like on one ordinary job. Same seven things happen either way. What changes is how many of them land on a person.

The same thing again, step by step, if you would rather read it than look at it.

That example is an inquiry, because it is the easiest one to picture. It is not the only shape. Most repetitive work in a business is something arriving, something being worked out, and something being prepared as a result.

Every one of those is the same job with different nouns. Something triggers it, the agent does the middle, and a person still decides the part worth deciding.

Most business software asks you to translate how you work into its screens, its fields, and its idea of a workflow. You spend the first month bending your process to fit the tool, and the tool never quite fits.

An agent works the other way round. There is no settings screen to learn. There is a written description of a job, in ordinary English, and the agent does that job. When the job changes, you change the words.

Which means the hard part of this is not technical. The hard part is saying out loud how something actually gets done in your business, including the parts that live in one person’s head. That is the work we do together, and it is why this starts as a conversation rather than an installation.

That is the whole of it. There is no fifth hidden part, no model to choose, nothing to tune. If you ever want to know why an agent did something, the answer is in those four things, and you can go and read them.

A playbook is not code and it is not a flowchart. It reads like onboarding notes for a new hire, because that is exactly what it is. Here is a piece of a real one, for an agent that chases unpaid invoices.

Notice what is in there. Who the people are. What the thresholds are. How you sound to your own customers. And, most importantly, the things it must not touch. That last section is usually the one that takes the longest to write, and it is the one that makes the whole thing safe.

We write the first draft with you, because most people have never had to say their process out loud before and it takes a couple of passes. After that it is yours.

Connecting a system means going to that system and granting permission there, on its own screen. Fluid never asks for a password and never has one. You are approving a named permission, the same way you would for any other app you trust.

Those permissions are deliberately narrow. An agent that drafts replies can read what comes in and write a draft. It cannot delete anything, and it cannot change mail that already exists, because we do not ask for the ability to do either.

Connections are then attached to agents one at a time. An agent can only touch what is explicitly linked to it, so an agent built for one job cannot wander into another.

The dial is enforced in our software, not left to the agent’s judgment. At the first two settings the ability to send does not exist for that agent at all. It is not that it has been told not to. There is nothing there to use.

The limits at the top of the dial hold no matter what: an agent reaches only the people and places the work already involved, or the ones you wrote down when it was set up, and it has no ability to delete anything at any level. Those are ours, not yours, and they are not adjustable. An agent cannot be talked into writing to somebody new, because picking who to write to was never something it does.

1. It wakes up

   On a schedule you set, or when something happens. Twice a weekday is common. First it checks whether you have paused it, and stands down if you have.

2. It checks it can still get in

   Passwords change, permissions get withdrawn, people leave. An agent that has quietly lost access looks exactly like a quiet week, so it proves it can reach your systems before doing anything else. If it cannot, it stops and tells you.

3. It looks only at what is new

   It remembers where it got to last time. It does not re-read your history, and it does not act twice on the same thing.

4. It checks whether a person already handled it

   Somebody may have picked up the phone an hour ago. Two answers to one question is worse than none, so it looks before it writes.

5. It decides, using the playbook

   Every item gets sorted, and each one carries a plain sentence saying why. Anything the playbook says to leave alone is left alone.

6. It does the work, and stops short of the last click

   A reply written but not sent. An order filled in but not submitted. A record updated and flagged for a look. Whatever the job is, it is left where you already work so you finish it the way you always have.

7. It writes down what it did

   Every item, every decision, and a link to the actual work. Nothing happens that you cannot go and look at.

8. It hands back what it will not decide

   Anything ambiguous, anything about money it was not told about, anything that smells like a complaint. It says what it saw and stops.

Most runs are quiet. That is the point. The interesting output of a good agent is a short list of things that need you, and a much longer list of things it dealt with that you never have to think about.

Most of the time there is nothing to tell you, because the work turns up where you already work. The reply is sitting in your drafts when you open your mail in the morning. The order is in the system you were going to open anyway. No new place to check, no new habit.

When something genuinely needs a person, how loudly you hear about it depends on what kind of thing it is. A hot inquiry at eight at night is not the same as a newsletter signup, and they should not arrive the same way.

Two rules make this work rather than become noise. Nothing shouts outside your working hours, except an agent that has got stuck or lost access, which is the one case where silence is genuinely dangerous. And a new agent starts quiet: everything lands in the summary until the record shows the sorting is right, and then you turn up the ones worth turning up.

Underneath all of it, Fluid holds the full account of every run: what came in, what it decided about each item, why, and a link straight to the real work. That is where you go to look into something, not where you find out.

You can also tell it when it got something right or wrong, one item at a time. That is not a rating for its own sake. It is the evidence that decides whether an agent has earned more rope, and it is the thing we use to sharpen the playbook.

It misreads something. At Prepares only, you notice when you read the draft, and you fix it in ten seconds or bin it. That is the entire cost. This is why agents spend their early weeks there, and stay until the drafts are consistently right.

It is not sure. The playbook tells it to stop rather than guess. An agent handing something back with a sentence about why is doing its job, not failing at it.

It loses access. Someone changes a password, an administrator withdraws a permission. The agent checks before every run and stops with a plain explanation instead of failing silently for a fortnight.

Something goes wrong that nobody predicted. Every action is written down, so we can go and look at exactly what happened rather than guess. Then the playbook changes so it cannot happen the same way twice.

The temptation is to start with the biggest headache. That is usually the wrong choice, because the biggest headache is normally the one needing the most judgment. Start with something dull and frequent. It builds a record, it teaches us how you work, and the second agent is faster because of it.

Most businesses have five or six candidates once they start looking. Finding them is what the first conversation is for.

The time comes back. Count what the repetitive part of a job costs you in a week, across everybody who touches it. That is the obvious number, and it is usually the smallest of the three.

The errors stop. Harder to price and usually worth more. A number typed wrong, a confirmation that never went, two people each assuming the other had it. Those cost you rework, and sometimes a customer, and they happen because a person is doing something repetitive rather than because anybody is careless.

Your good people get their attention back. The argument that actually matters. The hours you free up are not generic hours, they belong to the people you least want spending an afternoon copying details between two screens.

The bill has three parts: one cost to get the agent built, a flat monthly fee for the agent, and a monthly fee for each person it works for. No metering, no per-message charge, no bill that moves because you had a busy week. You should be hoping it does more work, not watching a counter. The numbers are on the pricing page.

Week one. We talk through the job in detail and write the playbook together. You connect the one system it needs. This is the part that takes your time, and it is a matter of hours, not days.

Weeks two to four. It runs, preparing only. You check what it writes and tell it when it is off. It gets noticeably better over this period, because the playbook gets sharper every time you correct something.

After that. You look at the record when you feel like it, and deal with the short list of things it flagged. If the record says it has earned more rope, you give it more. If it has not, it keeps drafting, which is still most of the value.

Then we do it again for the next job, and that one is quicker, because the hard part was never the technology.