Stop Thinking Chatbot. Start Thinking Workflow.

AI adoption is creating an unusual dynamic inside businesses. Instead of starting with an operational problem and looking for the right technology to solve it, organisations are increasingly starting with the technology itself: Where can we use AI? Do we need an agent? Could we automate this with an LLM?

There is a more useful place to start: how does the work actually get done today?

Consider a business receiving hundreds of customer enquiries through email, WhatsApp and website forms. Employees read each message, identify the customer, understand what they need, determine its priority, assign it to the appropriate person and respond. From a distance, this can look like an obvious opportunity for an AI agent. Break the workflow down, however, and a different picture emerges.

Bringing multiple communication channels together is an integration problem. Looking up an existing customer is a data problem. Applying known priority rules and routing an enquiry to an account owner are conventional automation problems. None inherently requires AI.

Understanding an unstructured customer message is different. A customer might write that half their delivery arrived damaged and nobody has responded for three days. Turning that language into structured information - complaint, damaged goods, existing customer, high urgency - is a problem where AI can add considerable value.

This suggests a useful principle: use AI for the uncertainty and software for the certainty.

Once AI has converted unstructured information into something structured, conventional software can often take over again. The CRM provides known facts. Business rules determine priority. Workflow logic determines routing. Automation executes predictable actions. Humans remain involved where risk, ambiguity or accountability requires them.

Agents introduce another level of capability, but they should not be the default destination for every AI workflow. If the sequence of actions is already known, a controlled workflow containing AI may be sufficient. Agency becomes more valuable when the desired outcome is known but the system needs discretion over how to achieve it - choosing tools, gathering information, evaluating results and determining what to do next.

A useful way to approach the problem is therefore:

Process → Automation → AI → Agency

Start by questioning the process itself. Then automate what is predictable. Introduce AI where interpretation or reasoning genuinely adds value. Introduce agency where the system genuinely benefits from autonomy.

Importantly, this is not a maturity ladder. A business should not aspire to move every workflow towards agents. Sometimes conventional automation is the right final architecture. Sometimes AI should perform one small step. Sometimes the right answer is still a human. And sometimes the process should simply be removed.

As AI becomes more capable, can AI do this? will become an increasingly uninteresting question. In many cases, the answer will be yes.

The more valuable question will be:

Should AI do this, or is there a simpler and more reliable way to get the work done?