Building agentic intake applications in 2026 is straightforward. The model knows how to read documents. It knows how to extract structured fields. It knows how to call tools and route results. The first end-to-end demo runs in mere days.
Testing them can be absolutely brutal.
This is consistently turning out to be where our engagements get hard. The build is the straightforward part. The QA is not.
The structural reason is simple. Old-school enterprise workflows were linear. A document came in, a person extracted fields, another validated them, another routed the result. Each step had a clear input, a clear output and a domain owner. QA was deterministic and decomposable. You tested each step against its expected behavior. The chain held because each link was verifiable in isolation.
Agentic workflows collapse that chain. A single agent execution might read a document, classify intent, extract fields, validate against a database, decide what to do next, call a tool, transform the result and route to the next system, all in one non-deterministic chain. The same input often produces different paths. The same path produces different outputs. The agent is not deterministic, and the test cannot be either.
The industry is starting to admit this. Traditional software testing relies on systems behaving the same way given the same input. Traditional machine learning evaluation relies on a fixed input-output mapping. Agentic AI can break both assumptions at once. The evaluation tooling is fragmented, benchmarks are inconsistent, and there is often no consensus on what good even looks like. Gartner forecasts that by 2028, a third of enterprise software applications will include agentic AI. The QA problem is about to become a dominant operational problem in enterprise AI deployment.
QA for agentic systems requires a kind of judgment that QA for traditional systems did not. The reviewer has to understand the domain, the system, the operations and the failure modes simultaneously. That combination of domain knowledge, system understanding and operational fluency is the operator profile we keep writing about. It is also exactly what is in short supply in the firms most eager to deploy agentic AI.
The discipline we have adopted, by necessity, is often anchored in manual testing. Scenario-based testing rather than input-output matching. Trace review by operators with domain expertise. Human-in-the-loop confirmation at decision points where the cost of an agent error is meaningful. None of this scales the way traditional QA does. All of it is more expensive than the demos suggest.
For firms thinking about agentic deployments: the model is the cheap part. The build is the medium part. The QA is the part that will surprise you. Budget more process judgment for it than you think you need. The companies that ship agentic systems that work are the ones that paid for the operational judgment to validate them.