The short answer.
Typing a prompt into ChatGPT, Claude, Copilot or Gemini and getting a fluent answer feels like knowing AI, so nearly everyone in your company believes they already do. Chatting is the surface of these tools: the same assistant, on the same account you already pay for, can hold a team’s instructions, read its files, work from its records and run a piece of work the same way every time. Most people never see that part, and a company where everyone chats can go years without anyone noticing what it is missing.
A fluent answer feels like competence.
A fluent answer feels like competence because the assistant does the hard part of looking good. You type a loose question, and a well-written reply comes back in seconds. Nothing in that moment shows how much more the tool could have done.
Chatting was the right way to start. It costs nothing to try and it teaches people what the assistant is good and bad at. The problem is not that people chat. The problem is that chatting feels finished, so most people stop there.
The chat box also hides its own depth. A spreadsheet shows you its menus and functions even when you never use them. A chat window shows you one empty line and a cursor. Everything past that line has to be known about before it can be used.
The same gap as basic spreadsheets and an advanced modeller.
The gap between chatting with AI and knowing AI is the gap between someone who does basic spreadsheet work and an advanced modeller. Both open the same program. One types numbers into cells, adds a column and formats a table. The other builds a model that takes this month’s figures, runs them through formulas and macros written once, and produces the same finished report every month without being rebuilt.
Nobody would say the person who can add a column “knows Excel” in the sense a finance director means it. Yet in most companies, the person who can ask ChatGPT for a draft email is counted as someone who “uses AI”, and the question stops there.
The good news is that the advanced end of AI is far less complicated than advanced spreadsheet work. There are no formulas to learn and no macro language. The advanced user of an assistant writes down, in plain sentences, how a piece of work is done, gives the assistant the files and records it needs, and checks what comes back. The difference is not talent. It is knowing that this way of working exists, and deciding to use it on the work that repeats.
What sits past the chat box.
Past the chat box, the paid versions of ChatGPT, Claude, Copilot and Gemini each offer, in their own form, the parts that turn a conversation into a workflow. Most people in most companies have never opened them.
- Saved instructions.Written once, applied every time: how your company writes, what a good answer contains, what never goes out. A chat starts from nothing; saved instructions start from your standards.
- Shared workspaces with files.A place where a team keeps its reference material, its templates and its past examples, so every conversation begins with the context instead of without it.
- An assistant set up for one job.A version of the assistant built for a single piece of work, such as preparing a weekly summary or a first reply to a type of enquiry, that anyone on the team can open and run.
- Working with real files.Reading a spreadsheet export, comparing two documents, pulling figures from a stack of reports, and producing a finished document in the shape your team uses.
- Connections to business systems.Reading from the systems your company already runs, such as your accounting system or your customer records, and, with a person approving, writing back to them.
- Longer, deeper work.Research across many sources, and tasks that run through several steps in order instead of one reply at a time.
Put together, those parts are a workflow: a piece of the business written down once and prepared by AI the same way every time, with a person deciding what happens next.
What your company is paying for.
Companies put money into AI in two broad ways. The first is building something new: a new function, a new service, a system made for the business. The second is pairing employees with AI tools, by buying seats of ChatGPT, Claude, Copilot or Gemini for the people who already do the work.
The second kind is where most companies of 10 to 200 people have spent, and it is the kind that stays at the surface. The seat you pay for each month carries everything in the list above. When the people holding those seats only chat, the company is paying for a full spreadsheet program and using it as a calculator. How do we get more value from the AI we already pay for? takes the money side of that question further.
How a whole company stays at the surface for years.
A whole company stays at the surface for years because every signal it can see says AI is going well. People use the assistant daily. The answers read well. Nobody complains. The things that would show the gap are the things nobody looks at.
- Everyone measures by use, not by work.“Do people use it?” gets a yes. “Which piece of work does AI prepare the same way every week?” rarely gets asked, and at most companies the answer is none.
- The advanced user is invisible.One person in a team may have built something far better than chatting. It lives on that person’s own account, the rest of the team never sees it, and it leaves when that person moves on.
- Nobody owns the question.Buying seats feels like the decision. After that, each person decides alone what to do with the assistant, which is how use stays undirected.
None of that is anyone’s failing. The stages of AI at work describes where this leaves most companies: switched on and experimenting, with the step to a plan still ahead.
How to recognise the surface in your own company.
You can recognise the surface in your own company by asking five plain questions of each team lead. Copy them and use them this week.
- Name one piece of work AI prepares the same way every week.If the answer is a description of chatting (“we use it for emails”), the team is at the surface.
- Where are the instructions written down?If they live in each person’s head or in old chats, nobody else can run them.
- Does the assistant start from your files, or from a blank line?A team that pastes context into every new chat has never set up a shared workspace.
- Who in the team uses AI best, and does anyone else use what that person built?If the answer is a name and a no, the best work in the team is stuck on one desk.
- Who decides what moves onto AI next?If nobody does, nothing will.
Take a 45-person civil engineering firm where every engineer uses an assistant daily. Asked the first question, each team lead describes drafting emails and tidying reports. One project manager has a saved set of instructions that turns site notes into the firm’s standard inspection report, and nobody else knows it exists. The firm has been using AI for a long time, and it has not deployed a single workflow. What is the difference between using AI and deploying it? sets out that line in full.
The way out is a plan.
The way out of the surface is a plan: decide which work in which team moves onto AI first, develop the workflows that do it, and deploy them so every person in the role runs the same current version. FullyPowered Planner is where that plan lives: you set out your teams, their roles and the work between them, and choose what to build first. How organizations plan their use of AI covers the method step by step.