Business Process Automation With AI: Write the Job

Business process automation with AI starts with a written job. We mean the steps a person already follows, in order. We also mark the spot where they use their own judgment. An AI model can’t follow a job that lives in someone’s head. It will make up the missing steps, and the made-up steps will sound very sure of themselves. Anthropic’s December 2024 essay says to start with the simplest thing that works. It also says to keep a fixed list of steps until the next step truly depends on what the last one found.¹ We treat that as everyday advice. The tool comes after the page, and if the page is blank, the tool is just guessing.

Cited in this note

75%

If the steps are not written down, stop spending money. The process guide is the matching fix for that blank page. The automation note helps you pick which job to start with once you can see more than one. The agents note covers a later question. It asks what to do when a fixed list is not enough, because the next step depends on what the last step found. An AI agent is a tool that picks its own next step. Most of the rooms we sit in are not ready for that yet. People are still telling the work from memory, and the story changes each time we ask.

The process is the product

The map is the thing you are making. Software is only a way to run that map the same way on a busy Tuesday as on a quiet Monday. Microsoft’s 2024 index says 75% of knowledge workers use AI, and 46% of those users started less than six months earlier.² Using AI is already normal, but having a shared plan is not. Leaders worry about this, since 60% of them say their company lacks a plan and vision. Another 55% of leaders were worried about finding talent. Only 25% of companies were planning to teach people how to use AI. People are already in the tools, and almost nobody is showing them a shared way to work.

Your staff are probably using AI already, with or without a plan. Among AI users, 78% bring their own tools to work, and that number is 80% at small and midsize businesses. At a small firm that often means a personal login the owner likes and a different tab the office manager likes. The work lands in the shared inbox, but the method stays in a browser history. When that person is out, the method is out too. That is not a process, because it is only a habit with a password. If you ban the login without writing down the path, the habit just moves somewhere else.

The place where the hours go should decide which job you write first. People spent 60% of their time in the work apps on email, chat, and meetings, and 40% on making new things. Owners often want to automate the part that feels like making something, and they leave the inbox alone. But the inbox is the job, and so is the chat thread. So is the meeting that ends without a note. Making new things is the smaller share of the day. So we sort the daily mess before we polish a rare piece of writing.

Gartner’s January 2026 article says at least 50% of generative AI projects were dropped after the first test run by the end of the year before. It names poor data, weak risk checks, rising costs, and unclear value as the causes.³ We don’t read that as a reason to fear the tools. We read it as a picture of demos that never had a job attached. Poor data shows up when nobody decided which record is the real one. Weak risk checks show up when nobody owns the stop button. Rising costs show up when the test keeps growing because the goal was never one written path. Unclear value shows up when you can’t point at the step that got easier to finish. Then the project is dropped, and the report blames the model. The missing page never makes it into the report.

Write down the steps a person already follows. Mark the step that waits for a judgment, and automate the steps that don’t. If every step needs a judgment, you don’t have a process yet, you have a craft. A craft can still get help at the edges, such as drafts, reminders, and a place for the file to land. But it should not be sold a machine that pretends to replace the judgment. We would rather say that on the first call. Then nobody pays for a pilot that looks just like the pile Gartner described.

A worked example

Take a trades desk as the example we walk through. This is not a measured result and it is not a trophy story. It is the shape of the afternoon we keep seeing when we ask someone to tell us about yesterday. The owner prices the work and the office manager answers the phone. A lead tech visits the site and comes back with a note that is half measurements and half memory. The bookkeeper sends the invoice only after someone says the job is done.

Yesterday’s quote starts as a voicemail. The office manager jots down the name, the street, the room, and a mention of a leak. The lead tech visits and takes photos on a personal phone. Then he texts a few of them to the owner with a line saying the job isn’t bad. The owner opens an old quote, changes a few lines, and sets the price from a similar job they remember. The bookkeeper never sees any of it until the customer says yes. Sometimes the yes arrives as a text to the owner, who means to forward it and forgets.

When we sit there and write the path, it becomes ordinary in the best way. The office manager logs the address and the room before the visit is booked. The lead tech files the measurements and the photos before leaving the driveway. The owner prices the job from that note, and the price is the step that waits for a person. The office manager sends the quote only after the price is on it. The bookkeeper opens the invoice only after the yes is written down. We would not hand the price to a model. The rest is copying, ordering, and waiting for a mark. A simple system can carry the note, remind the tech about a missing measurement, and hold the quote. A person still sends it. What the bookkeeper wants is usually small. It is a written yes, a matching price, and the same spelling of the name on the invoice. That is a handoff.

A clinic has the same shape with different words. The front desk takes the call, and the practitioner decides whether the visit is something they can help with. Billing waits for both. Once those steps are written, the front desk can draft the confirmation while a person still sends it. If nobody writes them, someone buys a tool that claims to handle intake. Meanwhile the practitioner keeps a side pad that becomes the real process. A shop counter looks simpler, which is why owners skip the page. The Saturday clerk and the owner tell two different stories when we ask each of them how a reorder really happens. We write both versions down before anyone opens a settings screen.

How we’d run it

First, we ask the person who did the job yesterday to walk through it without making it better. We want the version that happened, including the text message and the sticky note. We write their words where they can see them. We don’t clean them up yet, because cleaning is how wishes sneak in.

Second, we ask another person who touches the same job to walk through it too. The office manager and the owner almost never tell the same story. Where they disagree, we don’t blend the two into a polite middle. We ask which version the customer actually got, and that version wins because the customer is the only one who paid.

Third, we write the steps as short sentences that a new hire could follow on a tired morning. Each sentence names a person and the thing they make. The lead tech files the measurements before leaving the driveway, for example. If a sentence has no result, it is not a step but a mood, and moods can’t be automated.

Fourth, we mark the step that waits for a judgment. On the trades desk that step is the price. At the clinic it might be whether the visit is something they can take on. At the shop it might be whether a regular customer gets a refund after the posted window. If we can’t find the mark, we are not ready to automate anything. We don’t yet know what has to stay with a person.

Fifth, we automate only the steps that don’t wait. We move the note, remind the person, hold the draft, and file the photo. We use the simplest thing that makes each move. A fixed list of steps is enough when the next step doesn’t depend on a surprise in the last one. The quote either has a price or it doesn’t, so that is a gate and not a thinking problem. We stay with the gate even when a vendor calls it an agent.

Sixth, we run the written path on live jobs while the old path is still allowed. Then we watch where people step around it. The workaround is the missing step, so we add it to the page. We don’t scold the person who stepped around, since they are showing us the job we failed to write. We keep a named person on the send button. The system can prepare the draft. But the words that reach the customer go out because someone with a name looked at them. Without that, you just have a mailbox that talks.

Common mistakes

The first mistake we see is buying a tool because a neighbor bought one. The neighbor’s job is not yours, and their judgment may sit in a different place. We once sat with an owner who bought a stack of seats. Nothing was connected to a real quote, and the owner called the quiet that followed a failed pilot. The pilot never had a page, and the bill for the seats still arrived.

The next mistake is writing down the process you wish you had. The wished-for version has no voicemail and no promise to remember, but the real version is made of those things. If you automate the wish, the staff will keep the real path in a group text while the new system sits empty. Then people blame the staff for not using it. Really, the page described a company that doesn’t exist. We write the messy path first, and then we remove one mess at a time once it is visible.

Another mistake is automating the judgment because it hurts, which has the best sales pitch of all. Pricing hurts, and so does deciding whether a visit is in scope. Telling a customer no hurts as well. But pain is not a sign that a model should own the step. We move the chores around it and leave the painful step with a named person. If the owner wants that step gone, we ask a question. What will you do when the model is wrong in front of a customer? If the answer is a shrug, we don’t build it.

It also wastes time to treat personal tools as a moral failure. The share of people who bring their own tools is just a fact. At small and midsize businesses that share is higher. If you ban the personal login without a shared path, people will find another login. So give them the path, and then show it on a real job. Teaching, which few companies were even planning, is how the path survives the first busy stretch.

Calling a first test run a strategy is how budgets disappear. A test run checks a written step, and with no step you only have a demo. When the value stays unclear, the project joins the dropped pile. Leaving out the person who lives in the path is the same mistake in a quieter voice. If the office manager was not in the room, the map won’t match the phone. If they stumble while reading a sentence aloud, the sentence is wrong.

What we’d do

We would say no to the model until the steps fit on one page. A person who does the job should nod at each line. We hold that view even when the owner is eager and a vendor is already waiting in the hallway. Being eager is not a process, and a demo in the hallway is not a process either. A pile of licenses is not a process either. The page is the process, and everything else is a way to run it.

We would pick the job that already happens every day in email, chat, or the meeting note. That is where the hours already go. We would mark the judgment and keep a named person on it. A fixed list of steps can carry the rest. We would not start with an agent or a custom build. If the next step truly depends on what the last step found, we can talk about something smarter. First, someone has to show us that branch on a real job from last week.

We would also say the quiet part about teaching. People are using these tools, and few companies were planning to show them a shared way. A path that nobody is shown on a real quote is like a folder nobody opens. If you want a test before you spend anything, ask yesterday’s doer to write the steps from memory. Then ask the person who receives the result what they wish they had received. If those two pages don’t match, you have found the work, and you can name a product after that, not before.