Assertions account for roughly 20% of the code a test automation engineer writes. On a team of 50 engineers at a fully loaded cost of $100K each, that is about $1M a year spent writing and maintaining UI validation logic — work that adds no new test coverage.

Most QA conversations right now are about fully autonomous, AI-generated test suites. That framing sets the bar too high. Moving to AI does not require an overnight overhaul, and the returns do not wait for one. Applying it to even 5% of your tasks produces measurable gains and builds the foundation for wider integration later.

A smarter approach to UI validation

One example.

Traditional UI automation depends on framework-specific locators that break whenever elements shift. Vision Language Models (VLMs) offer a more resilient path:

Framework agnostic — send a screenshot to the model and validate UI state across web, mobile, SAP, or desktop with one code pattern instead of four.

Self-healing — dynamic element IDs, minor styling updates, and DOM restructuring stop producing red builds.

Lower maintenance — engineer time shifts from repairing locators to expanding coverage.

In practice, this replaces a brittle chain of locator lookups with a single call: here is the screenshot, confirm the confirmation page shows a total of $49.99 and a success banner. The same assertion works against the mobile build.

Managing the transition safely

Visual assertions bring real trade-offs, and the honest ones are worth naming:

Non-determinism. A model can return different verdicts for the same screen. Keep exact-value checks (totals, IDs, dates) in deterministic code and use the model for layout, state, and rendering questions.

Cost and latency: every assertion is an API call. Reserve visual validation for checks that can fail, and batch or cache where possible. Use local LLMs to reduce costs when feasible.

Data exposure. Screenshots from production-like environments often contain customer data. Local or self-hosted models resolve this in regulated industries.

These are design decisions, not blockers.

The pattern I keep seeing in many enterprises is the inverse of this post’s title: a large, expensive AI programme that produces a small change in outcomes. Big change, tiny impact. The teams that get real value do the opposite — they find one narrow, repetitive, costly task and remove it.

That requires two things at once: knowing your own processes well enough to see where the cost actually sits, and knowing enough about how LLMs behave to judge where they can be trusted. Neither alone is enough.

Enterprise AI strategy is not about replacing human ingenuity or handing control to automated systems. It is about applying AI deliberately to eliminate low-value maintenance work and free engineering capacity for the work that matters.

UI assertions are one example. Most teams have several.

Where is the 20% in your workflow?