Which workflows should you not automate with AI?

Do not automate work that nobody owns, work you cannot yet describe consistently, customer-facing actions with no approval step, or decisions whose success you cannot measure. Some of that work should stay with people for good. Much of it is simply not ready yet, and automating it early makes the underlying problem harder to see.

Why does it matter what you leave out?

Deciding what not to automate protects your first project as much as deciding what to build. A workflow that fails in front of customers or the board makes the next one harder to approve.

Anthropic's engineers make a related point in Building effective agents: start with the simplest solution that works and add complexity only when it is needed. The same discipline applies to choosing where not to start.

Which workflows are not ready to automate yet?

These are readiness problems. Fix the underlying issue first, then look at the workflow again.

• Nobody owns it. If no one is responsible for the work today, no one will own the exceptions, the monitoring or the changes once it is automated.
• You cannot describe it. If two people would explain the steps differently, the workflow will encode one side of that disagreement.
• The data is unreliable. Automation moves poor data faster. Agree a system of record and clean the fields the workflow depends on first.
• Success has no measure. Without a baseline you cannot tell whether the change helped, so you cannot defend keeping it.
• It seldom happens. A task that comes up a few times a year rarely repays the effort of building and maintaining a workflow.

Which work should stay with people?

Some work is better kept human even when the process is clear, because the judgement, the relationship or the accountability is the point of it.

• Commercial commitments. Pricing, discounts, contract terms and anything else that binds the business.
• Sensitive customer moments. Complaints, cancellations, bad news and conversations where trust is at stake.
• Decisions about people. Hiring, performance and anything with legal or ethical weight.
• Final approval of high-stakes messages. Board papers, investor updates and public statements.

AI can still help around these moments by preparing context, summarising history or drafting options for a person to review. Where decisions about individuals are involved, the ICO guidance on AI and data protection explains the extra obligations that apply. For a practical test of how much control to hand over, read what you should never hand an agent, and why.

What are the warning signs in a proposed workflow?

When someone proposes a workflow, check the design for these signs before anyone starts building.

• It sends customer-facing messages with no approval step.
• It asks for broad access to systems it does not strictly need.
• It keeps no log, so nobody can say later what it did and why.
• It has no pause switch and no named person to handle failures.
• Its running costs, including AI usage, have not been estimated. Our guide to controlling AI agent costs covers how to track and cap them.

Most of these can be designed out. If a workflow still cannot pass these checks after a redesign, it is usually one to leave with people for now.

What should you automate instead?

Start where a mistake is visible and cheap to correct: research, preparation, routing, reporting and record updates that a person confirms. Our guide to what to automate first with AI ranks ten B2B workflows and gives a six-question scoring method. If you are unsure whether you need an agent or a simpler automation, the Bowtie Academy lesson on agents, chatbots and automations explains the difference in plain English.

When you are ready to build, see how we approach AI automation, or send a short brief through find your starting point.

Knowing what to leave with people is part of the design, and the teams that get automation right are clear about both lists from the first workflow.