Case Study · B2B SaaS · Germany

How a German AI team cut
debugging time by 90%
while staying GDPR-compliant

Industry: Legal Tech  ·  Team size: 12 people  ·  Stack: OpenAI GPT-4o + Python  ·  Use case: AI-powered contract analysis

90%
faster root cause analysis for agent failures
€1,200
in API costs prevented by budget alerts
3 days → 90 min
time to detect quality regressions
0
findings in GDPR audit after deployment

Three problems they couldn't ignore

The team had built an AI assistant that analyzes legal contracts for German SMEs — extracting clauses, flagging risks, and summarizing obligations. The core product worked. But operating it in production was another story.

Silent quality regressions

A prompt change on a Friday afternoon caused a 22% drop in output quality. The team found out three days later — via support tickets from angry customers. By then, hundreds of contracts had been processed with degraded outputs.

Runaway API costs

A bug in their multi-step agent caused an infinite retry loop. It burned €1,200 in GPT-4o API costs over a single weekend before anyone noticed. There were no alerts, no spend limits, no visibility.

GDPR exposure from cloud-only tools

They evaluated LangSmith and Helicone. Both required sending prompts and outputs — containing sensitive contract data — to US servers. Their legal team immediately said no. They were stuck monitoring nothing.

Self-hosted observability in one afternoon

The team deployed AgentLens on their own infrastructure — a single Railway instance in the EU region. No data leaves their environment. The integration with their existing OpenAI setup took two lines of code:

import agentlens
 
# Added to their existing app startup. That's it.
agentlens.init("https://agentlens.their-infra.eu")
agentlens.patch_openai() # every GPT-4o call tracked automatically

For their multi-step contract analysis agent, they added trace instrumentation to get a full waterfall view of every step:

from agentlens import trace_agent, span
 
with trace_agent("contract-analyzer", input=contract_text) as trace:
    with trace.span("extract-clauses", span_type="llm") as s:
        clauses = extract(contract_text)
        s.set_output(clauses); s.set_cost(0.008)
    with trace.span("flag-risks", span_type="llm") as s:
        risks = flag(clauses)
        s.set_output(risks); s.set_cost(0.012)
    trace.set_output(risks)

They also configured a budget alert at €50/day and enabled automatic quality scoring on every response — no manual setup, no evaluation pipeline to maintain.

Four concrete outcomes in the first 30 days

Quality regressions caught in minutes, not days

The automatic quality scoring runs on every response. When a prompt change caused a quality drop, AgentLens fired an alert within 87 minutes. The team rolled back before any customers were affected.

Before: 3 days via support tickets After: 87 minutes via alert
💸

€1,200 in wasted API spend prevented

A budget alert at €50/day caught an agent loop within 40 minutes. The team fixed the retry logic the same day. Previously, the same bug had cost €1,200 over a weekend undetected.

Before: €1,200 lost in 48h After: caught at €38, fixed same day
🔍

Agent debugging: 45 minutes → 4 minutes

The waterfall debugger shows every span — which step ran, how long it took, what it cost, and what it returned. Finding the failing step in a 5-step agent went from log-diving to a single glance.

Before: 45 min of log parsing After: 4 min with waterfall view
🔒

GDPR audit passed with zero findings

Because AgentLens runs entirely within their own infrastructure, contract data never leaves their EU environment. The auditor reviewed the setup and closed the audit without a single corrective action.

Before: blocked by legal on all cloud tools After: GDPR audit — 0 findings

"We had a prompt change tank quality by 22% on a Friday afternoon. We found out on Monday morning via support tickets. With AgentLens, we would have caught it in under 90 minutes — automatically, without any manual evaluation setup. And since everything runs on our own server, our legal team actually approved it."

CTO · German Legal Tech SaaS · 12-person team

Ready to see this in your stack?

20-minute demo. We connect AgentLens to your LLM app and show you live what it finds.
No data leaves your infrastructure. No long-term commitment.

Or start free: github.com/Soufianeazz/agentlens