The AI iceberg: four levels of AI use in a company
Most teams use AI at one level: the chat window. That gives back minutes per task. Below the waterline sit three more levels, and they give back hours per week and days per month.
The AI iceberg is a way to see which level your team works at today. The four levels are not a ranking of companies. They exist to help you pick the next sensible step.
| Level | What it is | What it gives back |
|---|---|---|
| 01 · AI queries | One-off tasks in a chat window. | Minutes back per task |
| 02 · AI assistants | A helper that holds context. | Hours back per week |
| 03 · Automations and agents | Systems that connect your tools. | Days of process time |
| 04 · Agentic OS | Company-wide order for agents. | Company-wide |
A team can own the first three levels itself. Those are the levels ChangeAI takes teams through, from 01 into 03. The fourth is the direction we prepare you for.
01 · AI queries
One-off tasks: find, understand, draft, summarise. What most people already do with ChatGPT.
The gain is immediate. A task takes less time, and the person starts to learn where AI genuinely helps them. This level matters at the start, because the team sees examples from its own work.
The limit shows when everyone re-explains the same context, pastes the same documents and keeps their best prompts to themselves. That is the signal a task is ready to become a shared team tool.
02 · AI assistants
Helpers that hold context: email, document drafting, data consolidation, role-specific assistants.
An assistant works from agreed sources, keeps to the team's format and produces a first result a person reviews. The value shows within a week: less starting from a blank page, more consistent answers, a faster first draft.
The Kärcher customer-care team built an assistant that answers from their own knowledge base. Their reported result: 10 or more hours a week. Customer stories.
The next level is worth considering once the assistant needs to pull information from another system, or react to a recurring event.
03 · Automations and agents
Systems that connect tools and run without a human in the loop: SharePoint, Teams, knowledge bases.
The main question here is no longer how a prompt is worded. The work has to be described: what starts it, which data it uses, what must be checked, who receives the result and who handles the exceptions.
The Orbio World finance team reports that 3,000 invoices, which took three people a full week, now take about two hours. Debt reconciliation runs while the team does other work. Customer stories.
This is the level where time comes back in days. It is also where we take teams to.
04 · Agentic OS
One environment where a company's agents work: the knowledge base, skills, connections and access permissions.
It becomes relevant once the first three levels are working and the new systems have settled. Different companies will pick different tools, so this is not one vendor's product name.
The fourth level rests on people who work well with queries, teams that look after their own assistants, and processes with a clear owner. We present it as the direction and help you choose the right foundation when the time comes.
Finding your level in ten minutes
Pick one recurring task and talk it through with the people who do it.
- Does a person use AI only occasionally, to get this task done faster?
- Does the team share an assistant, a set of sources or a written way of working?
- Does information move between people or tools in a set order?
- Is it clear who owns the result, the input data and the exceptions?
- Would several teams benefit from the same knowledge, permissions and skills?
The first yes shows your current level. The first unanswered question shows what to sort out next.
What this model does not tell you
The iceberg does not count tools and does not rank companies. A team can work well with assistants and have no reason to automate a low-volume task. Another company can run a sophisticated agent while the wider team still needs the basic habits.
Start with work people already want to improve. Describe the process clearly, test it on a real task, and leave a named person accountable for the result.
To see your own processes at all four levels, that is where the AI Opportunities session begins.