The short answer.

Companies have invested in AI in two broad ways: building new functions that AI runs, and pairing nearly every employee with an assistant such as ChatGPT, Claude, Copilot or Gemini. The first kind gets a project plan, a budget and an owner; the second kind holds most of the people and, in most companies under 200, most of the money, and it was switched on and left with no plan and no owner. Nearly everyone with a seat uses it as a chat box, which is the surface of what these assistants do.

Companies have put money into AI in two ways.

The two kinds of AI investment are different in size, in who touches them and in how they are run.

  • New AI-run functions.A company builds something that did not exist before: a system that reads every incoming order, a service that answers a class of customer question, a tool that drafts a report nobody had time to write. A few people work on it. It has a budget line and a deadline.
  • An assistant for every employee.A company buys seats on ChatGPT, Claude, Copilot or Gemini, or finds the assistant already arrived inside software it pays for. Nearly everyone has access. The cost sits on the software bill next to email.

Most of the attention goes to the first kind. Most of the people are in the second kind, and for most companies under 200 people, so is most of the money spent on AI.

The first kind arrives with a plan and an owner.

A new AI-run function is managed because it was started as a project. Someone proposed it, someone approved the budget, and one person, by name, answers for whether it works. It has a scope, a test, a launch and a review. When it goes wrong, people know whose desk the problem lands on.

The second kind was switched on and left.

The assistant on every desk was never started as a project, so it never got the things a project gets. It was bought as a utility. Seats were assigned, a login went round, perhaps a short session showed people how to open it, and then the company moved on.

An assistant on every desk with no plan and no owner looks like this:

  • No list of work it is for.Nobody has written down which pieces of each team’s work the assistant should prepare. Each person decides alone, each day.
  • No one who answers for it.The IT lead manages the seats. Nobody manages what the seats are used for.
  • No shared version of anything.When one person finds a good way to do a piece of work with the assistant, it lives in that person’s chat history and leaves when they do.
  • No review.The new AI-run function gets a monthly look. The assistants get looked at when the renewal invoice arrives.

This is the state Signs your AI use is undirected describes. Experimenting was the right way to start. After years of experimenting, the second kind of investment is still where it started.

The gap between basic and advanced use.

Nearly everyone with an AI assistant uses it as a chat box: type a question, read the answer, copy the useful part. Because that works, people believe they know how to use AI. Most people do not know the assistant can do far more, and nobody at their company has shown them.

Spreadsheets are the closest comparison. One person opens a spreadsheet, types a column of numbers and adds them up. Another builds a model that pulls in last month’s figures, runs the scenarios and produces the same report every month with one change of input, with macros doing the repeated steps. Both people “use spreadsheets”. The work they get from it is not comparable.

AI assistants have the same gap, and the good news is that the advanced end is far easier to reach than an advanced spreadsheet. The assistants your company already pays for can:

  • Keep standing instructions.The assistant remembers how your company writes, what your customers are called and what a good answer looks like, so nobody types it again.
  • Hold a project.A team keeps its files, its reference documents and its instructions in one place, and everyone in the team works from the same set.
  • Be set up for one job.A purpose-built assistant does one piece of work the same way every time, for anyone in the team who opens it.
  • Work from your records.The assistant reads a spreadsheet export, a set of contracts or a month of tickets, and works through all of it.
  • Follow a written set of steps.Given the steps for a piece of work, the assistant runs them in order and hands back the result for a person to check.
  • Connect to the systems you run.Set up with guardrails and a person approving, the assistant reads from your accounting system or your ticket queue and writes back.

Put those together around one piece of repeating work and you have a workflow: a written way of doing that work that the assistant prepares and a person decides on. Almost no company has crossed from chatting to building on purpose. What is the difference between using AI and deploying it? draws the same line from the team’s side.

Why nobody manages the second kind.

Nobody manages the assistant on every desk because every reason a company normally starts managing something is missing.

  • It was cheap per person.A seat costs about what other software costs, so it was approved like other software: by the person who manages the bill, not by the person who runs the work.
  • It belongs to everyone.A tool used by every team has no natural owner. Each team lead assumes it is someone else’s job.
  • The chat box looks finished.The assistant opens to an empty box and a blinking cursor. Nothing on screen suggests there is more to set up, so nobody sets it up.
  • Advanced use is invisible.When one person builds a project with standing instructions, colleagues see nothing. Good practice stays private and never spreads.
  • Training stopped at the login.People were shown how to open the assistant, not how to give it a piece of their team’s work to run.
  • Nothing breaks.A managed project fails loudly. Unmanaged assistants fail quietly: the work simply stays as it was, and the bill keeps arriving.

What the second kind looks like in one company.

Take a 120-person home health agency. Its leaders approved one AI project: a system that drafts visit schedules from caregiver availability. It has a project lead, a vendor and a monthly review. Separately, every office and clinical coordinator has an assistant seat.

Illustrative example.

  • Intake.Two coordinators ask the assistant to tidy their notes after referral calls. A third pastes in the referral and asks for a summary in her own format. None of them knows the others do it.
  • Billing.The billing lead tried the assistant once on a denied claim, got a vague answer and has not opened it since.
  • Caregiver onboarding.The onboarding manager built a project holding the agency’s training checklists and uses it every week. Nobody else in the agency knows the project exists.
  • Leadership.The director of operations knows the scheduling project’s status in detail. Asked what the assistants are used for, she says “emails, mostly”.

The agency manages the investment that touches a few people and leaves alone the one that touches everyone. That is the common pattern, and it is not a failing of the agency’s people. Nobody was asked to own it.

A self-check to copy.

Answer these about the assistants your company pays for. Each “no” or “I don’t know” marks the second kind of investment running without management.

  1. Can you name the person who answers for what the assistants are used for?Not the person who manages the seats.
  2. Does each team have a written list of the work the assistant prepares?One list per team, kept in one place.
  3. When someone finds a good way to use the assistant, does the rest of the team get it?As a shared version, not a story at lunch.
  4. Has anyone set up standing instructions or a project for a team?Or does every person start from an empty chat box every day?
  5. Is any piece of repeating work run through the assistant the same way every week?By whoever is doing it that week.
  6. Do the assistants get a review that is not the renewal invoice?Monthly, with the same person each time.

How to take stock of the AI your company already pays for turns the same questions into a three-column list.

The way out is a plan.

The way out is to give the assistant on every desk what the AI project already has: a plan and an owner. Plan it, team by team: which pieces of work the assistant prepares and who decides. Develop the workflows for those pieces of work. Deploy them to the team, so everyone in the role runs the same version. FullyPowered Planner is where that plan lives: you set out your teams, their roles and the work between them, choose what to build first, and your colleagues add their own teams. How organizations plan their use of AI covers the method.