By Guest commentary Friday, November 13th, 2026

By Eric Zackrison

With so much information out there and so much hype about AI and agents, it is easy to get caught up in the shiny newtool. 

I was just working with a client last week who told me they wanted to start using a new AI scheduling tool and asked me how I thought they could integrate it into their workflow. That is exactly what I keep seeing, and it is also exactly backward. 

In deciding what tools to integrate, leaders should always start with the work itself, not the tool. 

The first step is to look at the work; without a clear understanding of the workflows, AI agents will only muck things up. So, take the time to map who does what and what the process looks like. 

Pay particular attention to repetitive tasks or tasks that are seen as mundane or simple. Things like data entry, summarizing meetings, scheduling, drafting routine emails, matching invoices to orders or pulling together weekly reports. 

Once you have those, then you need to understand which should and should not be handed off to an agent. Just because an agent is capable does not mean it should handle the task. Divide tasks up into four categories.

1. Must stay human: These are mostly judgment-based, like hiring decisions, pricing exceptions, delivering bad news to a client or performance conversations.
2. Partnership: These are things that may go back and forth between a human and an agent, like first drafts of proposals a person reviews, research summaries or customer responses a person approves.
3. Can be handed off: These are the repetitive things like the ones we listed above.
4. Work that defines people: This one is trickier. It is personal to each role and each individual. These might surprise you, but there is work that people love, or that may even be part of their identity: the salesperson’s relationship calls, the designer’s early sketching, the analyst’s favorite part of building the model. The best way to find these is to talk to your team before rollout.  

Then identify the tools that work best for each task. That might be a specialized tool built for one job, like scheduling or invoicing, or a general-purpose AI like Claude, Gemini, or ChatGPT. One thing to always check here is which tools your organization approves or pays for. 

Then you have to treat the agent like an employee. If you don’t have a clear job description for an employee, you wouldn’t expect them to be able to do their job. The same goes for an agent. But there are some nuances with agents that are different from those involved with employees.: 

1. When to defer to a human. Agents need to know exactly when to hand off to a person, and those handoff rules need to be explicit and detailed. 
2. What it’s not allowed to touch. A new employee learns boundaries by asking or by reading the room. An agent needs them spelled out in explicit detail.
3. What good looks like. Employees learn through training, but also through seeing others work to see what good looks like. You need to offer examples of what good looks like for your agents. 
4. It won’t tell you it’s confused. A new hire is much more likely to tell you they’re confused. Most AI tools are built to be helpful and agreeable, so they would rather give you an answer than say they don’t know, and when agents are wrong, they will confidently give you incorrect information.. So someone has to be looking, particularly when agents are new.  

This gives you some tools, but it is up to you to use them. Even if you have already rolled out agents, it is not too late. 

Do a quick audit of your team and make sure you can answer four questions for every agent you use: What job was it given? What is it not allowed to touch? When does it hand off to a person? And did anyone ask your team what work they would hate to lose? 

• Eric Zackrison is a professor at UC Santa Barbara’s Department of Technology Management. This is part 2 of an 11-part series.