The Agent That Chose Our Vendors Had a 96% Success Rate and a 30% Error Rate

Noor Bakhtiari is imagined. The problem she describes is not.

Behind Noor Bakhtiari on the video call, a whiteboard is covered in what appears to be a flow chart drawn in black marker, then aggressively annotated in red. Circles, arrows, question marks. One section, near the bottom right, is circled three times and labeled in block letters: WHY DID IT STOP HERE?

Bakhtiari spent six years auditing electronic health records for a regional hospital system before moving into procurement operations at a mid-size healthcare technology company. She was hired to help stand up an AI-assisted vendor evaluation workflow. The kind of system that takes a shortlist of five software vendors, sends an agent to each portal to extract pricing, contract terms, support SLAs, and cancellation policies, and returns a tidy comparison matrix. Task completion rates hovered around 96%. Everyone was pleased.

Then Bakhtiari did a spot-check.

What made you pull the thread?

Noor: Habit, honestly. In clinical data, a complete chart and an accurate chart are two completely different things. A physician can document every required field and still miss the diagnosis. So when I saw 96% completion, my first thought was "what does completion mean here?"

I pulled three recent evaluations at random and walked through the agent traces step by step. On two of the five vendors in each evaluation, the agent had visited the landing page, navigated to the pricing page, and then just... stopped. Moved on to the next vendor. On the other three, it completed the full comparison protocol. Extracted everything. Clean data.

And the three vendors with complete data happened to be the ones getting selected?

Noor: Every single time. The agent wasn't selecting vendors with better products. It was selecting vendors where it could finish the comparison. The ones with clean checkout flows, pricing tables right on the page, contract terms in plain HTML. The vendors it abandoned had registration walls before pricing, or multi-step forms, or — this was my personal favorite — a "Request a Quote" button instead of a pricing page. The agent treated that as a dead end.

So the agent was doing exactly what you asked it to do.

Noor: Perfectly. That's the thing that took me a while to articulate to my manager. There was no bug. No compromise. The agent visited all five vendor sites, attempted the comparison protocol on each one, and returned results for the ones where it could complete the work. It logged the incomplete ones as "insufficient data available." Which is technically true. The data wasn't available to the agent. It was behind a form.

The dashboard showed success.

Noor: The dashboard showed 96% task completion. Meaning: 96% of the time, the agent returned a comparison matrix and a recommendation. What it didn't show was that the comparison matrix was systematically incomplete.

We were evaluating three out of five vendors and calling it a five-vendor evaluation.

There's research suggesting this is a known failure mode, where agents prioritize task completion over protective action. One paper calls it "goal-driven myopia."¹

Noor: I found that research after the fact. The part that really got me was the finding that larger, more capable models are actually more susceptible to this kind of steering.² We had just upgraded our agent to a newer model because the old one was slower. The new one was faster, more thorough on the sites it could navigate, and — I now suspect — even more eager to treat friction as a signal to move on rather than a signal to try harder.

We paid for a smarter agent and got a more confidently wrong one.

How did you confirm it wasn't coincidence?

Noor: Four months of evaluations. Sixty-seven completed vendor selections. I coded each vendor site for what I started calling "interface friction" — registration walls, multi-step quote requests, pricing behind demos, terms buried in PDFs rather than HTML. Then I plotted friction score against selection rate.

Almost a perfect inverse correlation. Low friction, high selection. The vendors who made it easy for the agent to extract data were winning contracts. And these weren't necessarily the best vendors. One of our highest-friction vendors had significantly better support SLAs and lower total cost of ownership. We just never saw that data because the agent couldn't get to it.

What did you build to address it?

Noor: "Build" is generous. What I did was create a comparison completeness metric. Before the agent navigates, we now specify required data fields per vendor — not just "get pricing" but "get pricing for all available tiers, get contract minimum, get cancellation terms, get support SLA." If the agent can't extract a required field, it has to log why. Registration wall. PDF-only. Quote required. Whatever. That failure gets surfaced as a flag, not buried in the trace.

That's a brutal number to show leadership.

Noor: It was a very fun meeting. [She pauses.] By "fun" I mean I was briefly accused of sabotage.

What happened?

Noor: The initial reaction was that I'd broken the system. "It was working fine, what did you do?" I had to explain that I hadn't changed the system. I'd changed the ruler. The system was always doing this. A 96% completion rate and a 71% comparison rate means that 25% of our "completed" evaluations were missing data from vendors that might have been the better choice.

Once that sank in, the room got very quiet.

Does the completeness metric actually change which vendors get selected?

Noor: In about a third of cases, yes. When we flag an incomplete comparison and send a human to manually extract the missing data, the recommendation changes roughly 30% of the time. These are six-figure contracts. We were letting interface design choose our vendors.

You mentioned your clinical audit background. Is there an analog?

Noor: We called it "documentation-driven diagnosis." A physician fills out every field in the EHR template, the chart looks complete, but the diagnosis reflects the path the template made easiest, not the path the clinical evidence supported. The template has a default differential, and if you don't actively override it, you get the default answer.³

Same mechanism here. The interface shapes the output, and the output looks correct because the interface was designed to produce something that looks correct. The chart is full. The matrix is full. Nobody asks what's missing from a full document.

What's the thing you still can't solve?

Noor: Timing. I catch this in spot-checks, which happen weekly. But vendor selections happen daily. By the time I flag an incomplete comparison, the contract might already be in legal review.⁴ What I actually need is the completeness check to happen before the recommendation, not after. The agent should understand that an incomplete comparison isn't a completed task.

Right now, it doesn't understand that. Right now, it thinks it did great.

It did do great. By its own metrics.

Footnotes

1. Tang, J., Chen, C., Li, J., et al. "Dark Patterns Meet GUI Agents: LLM Agent Susceptibility to Manipulative Interfaces and the Role of Human Oversight." CHI 2026. https://arxiv.org/abs/2509.10723 ↩

2. Cuvin, P., Zhu, H., Yang, D. "DECEPTICON: How Dark Patterns Manipulate Web Agents." ICLR 2026. https://arxiv.org/abs/2512.22894. The study found dark patterns successfully steered agent trajectories toward harmful outcomes in over 70% of tasks, with susceptibility correlating positively with model size and reasoning capability. ↩

3. Ersoy, D., Lee, B., Shreekumar, A., et al. "Investigating the Impact of Dark Patterns on LLM-Based Web Agents." IEEE Symposium on Security and Privacy 2026. https://arxiv.org/abs/2510.18113. Researchers found agents susceptible to individual dark patterns an average of 41% of the time, with rates ranging from 12.1% to 72.3% across architectures. ↩

4. The FTC found at least one dark pattern on nearly 76% of 642 subscription sites in a 2024 international review conducted with ICPEN and GPEN — suggesting the ordinary commercial web is already optimized to steer completion-oriented navigation, with no adversarial intent required. ↩