Published
Four steps, not four products
AI at work is easier to plan as a progression than as a list of products. Each step adds one capability to the one before: an assistant answers, an agent acts, automation runs a process, and an AI employee owns a role. Knowing which one a piece of work needs is most of the design.
Assistants answer
An assistant responds when someone asks: a customer on your website or on WhatsApp, or a colleague looking for a policy. A good assistant answers from your own knowledge and says where each answer came from.
It changes nothing in your systems. Use one when the work is mostly questions that have known answers.
Agents act
An agent can do something about a request: look up an order, update a record, book a meeting. It understands the request, decides the next step, uses a tool and reports what it did.
Every tool it may use, and every limit on it, is defined in advance. Use one when answering is not enough and the next step is clear.
Automation runs the process
Automation connects steps across systems: a form arrives, its data is checked, a record is created, a person approves, a message goes out. AI adds judgement at the steps where fixed rules fall short, such as reading a document or sorting a request.
Use it when the same process, with several steps, runs again and again.
AI employees own a role
An AI employee combines the three. It talks to people, acts in your systems and runs its processes, within one defined role, such as a sales coordinator or a helpdesk.
It has its own permissions, its own activity record and a person it reports to. Use one when much of a role's work is repeatable.
Where to start
Start with the work, not the technology. List the requests that arrive every day, where they wait and who decides. Look for one workflow where an agent or a small automation could remove most of the waiting.
Start there, in Assist, and widen what the AI may do as you review its work.