$6.2B
Consolidated revenue
Average agent performance from 53% to 89% at a large financial services group
The journey worked for people. Tested against the agents already arriving, it completed 2 times in 5. Webzero measured a credit score journey inside a large diversified financial services group, rebuilt it around four properties, then re-ran the identical probe. Average agent performance moved from 53% to 89%, and successful runs from 2 in 5 to 4 in 4.
Average score across simulation runs of the same credit score journey.
Percent of runs returning a successful outcome. 2 of 5 became 4 of 4.
Minutes to complete the journey, roughly 71% faster.
Points between best and worst run. The journey became predictable.
Comparative AI agent simulations of one credit score journey, run before and after optimisation. Nothing in the probe changed between the two sets, which is what makes the two columns comparable.
A large, diversified financial services group operating across lending, investments, insurance and payments, and one of the largest non-bank financial ecosystems in its market. The scale is the reason a structural weakness in one journey matters far beyond that journey.
FY2025-26 consolidated. Local-currency figures converted to US dollars at prevailing rates and rounded.
The team picked a credit score journey: the first stage of the lending funnel and a high volume acquisition journey. Webzero evaluated it with the Webzero Agent Readiness Index and live agent simulation, across the full execution lifecycle.
Can an agent find the right journey at all?
Can it tell what the journey does and what it needs?
Can it reliably drive the forms, controls and navigation?
Can it get through verification without losing state?
Can it understand a failure and continue?
Can it actually obtain the intended outcome?
Can it determine the task successfully finished?
A site can be highly visible to AI and still fail the moment an agent tries to perform a task on it.
The index scores a journey on a 100 point scale across five pillars, 10 categories and 50 sub-pillars. The opening reading recorded 35 critical issues, 5 warnings and 12 healthy signals.
Can agents find and trust the site?
Can agents understand the content?
Do agents have the right tools?
Can agents navigate and recover?
How efficiently do agents succeed?
Average 53%. Only 2 of 5 runs returned a successful outcome. One run ran 888.5 seconds, another 662.6.
The agent could sometimes complete the journey. It could not do so reliably or predictably.
Scores ranged from 25% to 88% on the same task, with step counts from 3 to 14. Variance at that width is not a tuning problem in the agent. It is ambiguity in the journey.
In one representative run the agent found the right page, then talked itself out of the task. It read a mobile app prompt as a hard requirement, followed the primary call to action to an app store listing and terminated web execution. The task was completable on the web.
Nothing in the analytics of that session would have flagged it. No error fired, no form was abandoned mid-way, no support ticket followed. This is the shape most agent failures take.
Ambiguous controls, duplicate components, fragile sessions, poor machine readability, unclear states, weak recovery. These are familiar digital infrastructure problems. Agents simply made them measurable.
Natural language discovery was tested across three major assistants. All three surfaced competing bureau and marketplace providers. The client’s own journey was surfaced by none of them.
Why it matters If an agent cannot find the right journey, improving that journey has limited value. Readiness starts at discovery.
One journey contained four forms all reporting as visible in the DOM (display: block, visibility: visible, opacity: 1), three sharing identical HTML IDs. The only distinction between the live form and the duplicates existed at the rendering layer, not in the markup.
Why it matters A human uses visual context to tell them apart. An agent has to reason about which target is real, which adds retries, time and failure probability. One task should have one machine-interpretable path.
The credit score itself was rendered inside a canvas gauge. A DOM-based agent could read the axis labels (300 to 900) but not the number. A dropdown search using loose substring matching added a second layer of ambiguity.
Why it matters The information existed, presented for human perception rather than reliable machine interpretation. Human-readable is not agent-readable.
Session state was carried in a URL parameter, so Back, Home or refresh invalidated it and forced a full restart with no recovery path. One-time password validity was roughly 60 seconds, with no way to distinguish an expired code from an incorrect one.
Why it matters For a person, restarting is annoying. For an agent, losing session context means losing the entire execution state.
The experience did not consistently state what failed, why, whether the previous action succeeded or what to do next. An incorrect one-time password required the agent to detect the failure, start recovery and re-establish context unaided.
Why it matters Every ambiguous state adds a decision, and every decision adds variance. Agents perform when the system tells them what happened and what to do next.
The same patterns recurred across every journey assessed, which reframed the objective. Not fix this one form, but make the underlying journey deterministic enough for an agent to execute.
Nothing was bolted on top of the site. The journey underneath the agent changed.
The agent always knows which action is available to it.
The same task leads to the same execution path.
Important information is available in a form an agent can interpret.
Errors and interruptions do not terminate the journey.
The journey was consolidated around one deterministic path. Competing interaction patterns were removed, so the agent had fewer choices to interpret.
Duplicate forms and ambiguous elements were eliminated. One task, one intended interaction, one predictable next step.
The optimised experience shipped a structured definition of the journey alongside the page itself, purpose-built for LLM-based navigation.
One-time password handling, structured error states, recovery paths, journey continuity and progress visibility, so a recoverable hiccup stops becoming a total failure.
Results like the score itself were exposed through structured content and JSON-LD rather than visual presentation alone.
One task, one intended interaction, one predictable next step, and a definition of the journey an agent can read rather than infer.
The average matters less than the narrowing. A journey that returns 85 to 93 on every run is one an enterprise can plan around. A journey that returns 25 to 88 is not, whatever it averages.
2 of 5 successful. Median run time 8.3 minutes.
4 of 4 successful. Median run time 2.4 minutes, three runs under three minutes.
Successful runs went from 40% to 100%. Arguably the more important number: the point of an agentic journey is not a good score, it is a completed customer task.
Median execution fell roughly 71%, from 8.3 minutes to 2.4. The longest observed run halved, from 889 seconds to 426.
The gap between best and worst run closed from 63 points to 8. That narrowing is the clearest evidence the journey became deterministic, not just better on average.
Before optimisation the agent had to answer six questions the journey never told it. After optimisation the journey answered them.
That is the difference between a website an agent can read and a journey an agent can execute.
The most valuable result was not one optimised credit score journey. It was a repeatable model: measure, find the break, optimise the journey, prove the improvement, then run the same loop across the estate.
Two weeks, a permitted environment, synthetic identities, and a w0 score with the trajectory evidence behind it. No change to your stack.
The organisation's name, URLs and other identifying details have been intentionally redacted to protect confidentiality. All scores, run data and figures are reproduced as originally measured.