AI agents can automate work that used to need humans at every step. But when agent projects fail, it’s rarely the technology’s fault. Almost always, someone picked the wrong process to start with. We see the same two traps over and over: projects that are too ambitious—open-ended tasks where nobody can agree if the agent got it right—and projects so trivial they aren’t worth the effort.
Here are the five questions we ask before we write a single line of code.
1. Is the process repeatable?
Agents thrive on work that happens often and follows a pattern. Cases handled every day, documents reviewed every week, reports compiled every month. If a task looks completely different every time you do it, it’s hard to evaluate and harder to improve. Repeatability is also how the math works out: an agent running a hundred times a week pays for itself in a way one that runs once a quarter just doesn’t.
2. Are the inputs and outputs clear?
What goes in, and what comes out? If you can describe it concretely—say, “in comes a PDF contract, out comes a filled review template”—you have a solid candidate. If the answer is “it depends” on both ends, you need to define the process before an agent can take it on. Honestly, that mapping exercise is rarely wasted. Just forcing a team to agree on what a process actually entails is valuable.
3. Can the quality be tested?
This is usually the dealbreaker. If no one can judge if the agent’s output is right or wrong, you can’t evaluate it. And without evaluation, you can’t run it in production responsibly. You want a process with a ground truth, where a senior colleague can glance at the output and say, “this is correct,” or, “I would have done it this way.” We can then bake that judgment into automated tests that run whenever the agent is updated.
4. Are there realistic integration points?
An agent does real work when it can act in your systems—reading tickets, updating records, preparing drafts. But that requires APIs, permissions, and a way into the infrastructure. An agent without integrations is basically a chatbot. Most of the business value vanishes right there. Figure out early which systems the agent needs to touch and what access looks like. That alone affects the architecture and timeline more than which model you pick.
5. What happens when the agent is uncertain?
No agent gets it right every single time. It doesn’t have to. You just need a process that accounts for it. The safest places to start are processes where mistakes are cheap to catch and correct. The agent does the heavy lifting. A human reviews the work before anything goes live. Uncertain cases get escalated automatically. Processes where a single mistake is costly or irreversible can absolutely be automated, but they definitely shouldn’t be your first project.
Start with a delivery, not a platform
If a process answers all five questions well, the next step isn’t to build a platform. It’s to build a single delivery. A tightly scoped rollout in production, with clear acceptance criteria, proves whether the solution actually holds up. It gives you real answers before you try to scale. That’s how we run every agent project: deliver to production, evaluate against the criteria, and only scale what works.
Want to talk through which process should be your first? Get in touch and we will look at it together.