AI agents can automate work that used to require people at every step. But the most common reason agent initiatives fail is not the technology. It is that the wrong process was chosen from the start. We see two recurring failure modes: initiatives that are too ambitious (open-ended processes with no ground truth, where nobody can tell whether the agent got it right) and initiatives too trivial to be worth the investment.
Here are the five questions we ask before we build.
1. Is the process repeatable?
Agents are at their best on work that happens often and follows a recognisable pattern: cases handled every day, documents reviewed every week, reports compiled every month. A process that looks completely different every time is hard to evaluate and even harder to improve. Repeatability is also what makes the investment pay off: an agent that runs a hundred times a week pays for itself in a different way than one that runs once a quarter.
2. Are the inputs and outputs clear?
What goes into the process, and what should come out of it? If the answer can be described concretely (“in comes a contract as a PDF, out comes a completed review template”) you have a good candidate. If the answer is “it depends” at both ends, the process probably needs a clearer definition before an agent can take it over. That work is rarely wasted: being forced to describe what the process actually does is often valuable in itself.
3. Can the quality be tested?
This is the question that most often decides. If nobody can judge whether the agent’s output is right or wrong, you cannot build evaluation either, and without evaluation you cannot operate responsibly. The best processes have a ground truth or a clear definition of quality, where an experienced colleague can quickly say “this is correct” or “I would have done this differently”. That judgement can then be built into automated tests that run every time the agent changes.
4. Are there realistic integration points?
An agent creates value when it can act in your systems: read cases, update records, prepare material. That requires APIs, permissions and a way into the infrastructure. An agent without integrations is in practice a chatbot, and most of the value disappears with it. Map out early which systems the agent needs to reach and what access should look like. It affects both architecture and timeline more than the choice of model does.
5. What happens when the agent is uncertain?
No agent is right one hundred percent of the time. It does not need to be, if the process is designed for it. The best starting processes are those where mistakes are cheap to catch and correct: the agent does the heavy lifting, a person reviews before anything takes effect, and uncertain cases are escalated automatically. Processes where a single error is expensive or irreversible can absolutely be automated, but they should not be the first thing you build.
Start with a delivery, not a platform
Does the process answer all five questions well? Then do not build a platform. Build the first delivery. A well-scoped delivery in production, with explicit acceptance criteria, shows whether the solution holds and gives you the real answers to the questions above before you scale. That is how we work in every agent engagement: deliver to production, evaluate against the criteria, scale what has proven to work.
Want to talk through which process should be your first? Get in touch and we will look at it together.