We gave everyone ChatGPT or Claude. Why hasn't it automated any process?
A chat assistant makes each person faster. A process changes only when one piece of work is redesigned and given to an agent with context, tools and an owner.
Because a chat assistant makes each person a bit faster at their own tasks, and nothing in a rollout changes the process itself. Nobody picked a piece of work, redesigned it, gave an agent the company context and system access that work needs, or made someone own the result. To automate a process, start from one piece of work, not from the tool, and set the standard, the context, the tools and the approvals for that work.
Why the question comes up
We hear it from companies that did the sensible first step. On one recent call, a software company had rolled an assistant out to all staff. People were faster on their own work, but no process had been automated. On another, an advisory firm had a company wide AI subscription that 3 of its roughly 10 people used. Building a small tool took them 15 minutes. Making it stable and hosted was the hard part.
The pattern is common. In MIT's 2025 study, over 80% of organisations had explored or piloted tools like ChatGPT and Copilot and nearly 40% reported deploying them, but these tools mainly improved individual productivity, not P&L results. In McKinsey's 2025 survey, 88% of companies use AI in at least one function and 39% report any EBIT impact.
Personal speed is not process change
A process is the chain of steps, handoffs, checks and systems that turns an input into a result, such as an order into a confirmed delivery date. A chat rollout leaves that chain exactly as it was. Each person now does their step a little faster, in their own way, when they remember to ask. The minutes saved are spread across many people and mostly absorbed by other work. The bottleneck, the re-keying between systems and the waiting for someone to check, are all still there.
| In practice | A chat assistant for everyone | An automated process |
|---|---|---|
| Starts when | A person writes a prompt | An event or a schedule, such as a new order or a weekly report |
| Knows | What the person pastes in, plus the files it can search | The company's map of the work: its entities, rules, sources and exceptions |
| Does | Gives one person an answer or a draft | Works in the systems: updates records, drafts the reply, flags the exception |
| Checked against | The person's judgement in the moment | A standard agreed on real examples |
| Owned by | Each user | One named owner for the work |
| Measured by | Seats and usage | Quality, cost, time and human effort per piece of work |
Five reasons the rollout stops at the person
- Nobody chose a process. A rollout is a tool decision. Automation is a decision about one piece of work: which one, who owns it, and what result counts.
- The work was not redesigned. Putting AI on top of today's steps keeps today's handoffs. Most of the gain comes from changing the work first, as in redesign the work before you automate it.
- The assistant doesn't know the company. It knows what each person tells it. A process needs a maintained picture of customers, products, rules and exceptions, or the answers are confident and wrong. See why AI agents give confident wrong answers.
- It can't act where the work runs. Acting in a CRM, an ERP or an inbox needs scoped access, approvals for anything that leaves the company, and a record of what was done. A tool built in an afternoon rarely has those, which is why it stays a personal shortcut.
- Nobody owns or measures it. Without an owner, a standard and a baseline, nobody can say whether it works, so nobody can decide to rely on it. Who owns the AI platform after launch covers how to split that job.
What to do with the rollout you already have
Keep it. The assistant is where people ask questions and get help with their own tasks. Agents are where whole processes run. The two work together, as explained in agentic vs generative AI.
Then use it as a lead:
- Look at what people already use it for. Repeated personal use points at work worth looking at. MIT's study notes that some companies now learn from this employee use before they buy anything bigger.
- Pick one piece of work. It runs often, one person owns it, and the result can be checked. Keep it small enough to finish in one build cycle.
- Map it and agree the standard. Real examples, including the messy ones, and what a good result looks like for each, plus a baseline. See agree the standard before you build.
- Give an agent only what the job needs. The context for that work, tools granted for it, and a person's approval for anything that leaves the company.
- Measure it per piece of work. Quality, cost, time and human effort, before and after. The method is in how to measure AI agent ROI.
How Ortelian approaches it
Ortelian is a platform plus a forward deployed team. Every engagement starts with an audit, about two weeks on site, that finds where the work is held up and redesigns the process with your team. Then we build, test and deploy in two-week cycles until the first job meets the standard agreed on real examples. Agents are read only until a write is granted, and actions that leave the company, such as sending an email or an invoice, wait for a person. Your own AI tools, such as ChatGPT, Claude and Copilot, can connect over MCP. We set that up inside an engagement, so people keep the assistant they know while it draws on the same world model the agents use. The platform is hosted in the EU. More on how we work and the platform.
Questions people ask
Is ChatGPT Enterprise or Copilot enough to automate a process?
They help people with their own tasks, and they have added features that run multi-step tasks. A process also needs the company context, scoped access to the systems where the work runs, approvals and an owner. Without those, the assistant speeds up a person, not the process.
Why do so few employees use the AI tool we bought?
Usually because it adds a step instead of removing one. People use AI when it takes work off their plate inside the tools they already use. More in how to get staff to adopt AI agents.
Should we stop paying for the chat assistant?
Not because of this. It does a different job. Check what it costs per active user and what people use it for, and use that to choose the first process. The AI costs note covers how spend grows after a rollout.
Can employees build their own automations with it?
They can build them quickly. The hard part is keeping them stable, hosted, safe in your systems and owned by someone after the person who built it moves on. That is where most home built automations stop.
Sources
- MIT NANDA, The GenAI Divide: State of AI in Business 2025, 2025
- McKinsey, The state of AI in 2025, 2025
- Ortelian, How we work and Platform
