Series 3 of 4 · AI Evals · Level 1 · Topic 09

The Tool Landscape

The Ground Truth
IN THIS NOTE YOU WILL LEARN
  • 01.The four categories of eval tooling and when each matters — from open-source frameworks to enterprise governance platforms.
  • 02.How to avoid the three procurement traps: the PII leak, the APM illusion, and the ecosystem lock-in.
  • 03.Why PMs and domain experts — not engineers — should lead tool procurement, and what happens when they don't.

The Story

An enterprise legal drafting platform built their entire evaluation pipeline using a single model provider's ecosystem. Six months of calibrated evaluation data. Engineering was proud of the infrastructure.

Then a competitor released a model that significantly outperformed on the platform's specific legal reasoning tasks. The PM wanted to switch. The team assumed they could swap the API endpoint. In reality, their entire evaluation infrastructure was hardcoded to one provider.

Three months of engineering velocity lost. A rival platform that had built vendor-neutral evaluation made the same model switch in two days.

The Core Idea

The eval and observability tooling ecosystem in 2026 is no longer neatly separated. The products are converging. But four buying postures still matter.

Category 1: Open-source frameworks. DeepEval, Ragas, Phoenix. Maximum flexibility and zero license cost. They run in your infrastructure — your data never leaves.

Category 2: Managed platforms. Braintrust, LangSmith, Arize AX. Faster time-to-value — dashboards, alerting, collaboration, managed infrastructure.

Category 3: Model-provider eval tools. OpenAI's Evals API. The fastest path from "I use this model" to "I evaluate this model." The tradeoff: locked to one provider's ecosystem.

Category 4: Enterprise governance platforms. ModelOp, Credo AI. They sit above eval tools — tracking which AI surfaces exist in your organization, who owns them, and whether their eval evidence is current.

The Three Procurement Traps

Trap 1: The PII leak. A team integrates managed observability without asking the deployment question first. Traces containing unredacted patient transcripts flow to a multi-tenant SaaS vendor without a BAA. Self-hosted Phoenix or Langfuse would have kept traces inside their infrastructure.

Trap 2: The APM illusion. A Fortune 500 mandates their existing Datadog contract for LLM observability. Everything is green. Latency normal. Error rate low. But Datadog tracks infrastructure health, not semantic quality.

Trap 3: The ecosystem lock-in. The legal tech story from the opening. The fix: OpenTelemetry for instrumentation and a platform-agnostic eval framework for scoring.

Fig 1. The Tool Landscape
THE DEFINITION

Four ways to buy eval tools.
The feature lists converged. The postures didn't.

Pick the posture, not the checklist.
OPEN SOURCE

Maximum control. You build it.

MANAGED

Faster setup. They run it.

PROVIDER-NATIVE

Tightest integration. Locked in.

GOVERNANCE

Tracks everything. For compliance.

Kitchen equipment: buy ingredients and cook, use a meal kit, order from one restaurant,
or hire a food safety inspector. Different needs. Different answers.

Where This Hits in Production

Enterprise procurement eliminates tools before features matter. A CISO's checklist: SOC 2 Type II, data residency guarantees, SSO/SAML integration, role-based access control. A technically superior tool may be disqualified because it can't guarantee EU data residency.

PMs must lead procurement, not engineers. When engineers lead tool selection, they optimize for trace visualizations and SDK quality. The resulting tool often lacks what PMs need: a no-code annotation queue, readable prompt playgrounds, and accessible experiment comparison. The procurement test: if the non-technical quality owner can't comfortably review traces in the tool's UI, don't buy it.

!

Common Mistake

Buying a category, not a workflow.

First: buying governance before you have eval plumbing. You can inventory use cases and pass audits, but you still can't answer "which prompt version improved resolution rate?"

Second: engineers leading procurement — they pick the tool with the best waterfall charts. The PM retreats to spreadsheets.

Choose based on the operating loop you need to accelerate next — CI testing, experiment comparison, production debugging, or enterprise governance.

In Practice: The Three-Layer Enterprise Stack

The Maturity Progression
Layer 1: Open-Source
You build it. You run it. You own it.
DeepEval for CI/CD eval gates. Ragas for RAG-specific metrics. Phoenix for tracing + eval.
$0 costFull controlDIY effort
$0/month
Layer 2: Managed Platform
Faster time-to-value. They build it.
LangSmith for experiment tracking. Braintrust for dataset + scoring. Arize AX for production monitoring.
DashboardsAlertingVendor dependency
$500-5K/mo
Layer 3: Governance
Registry, risk tiers, audit trails.
ModelOp for AI system registry. Credo AI for compliance workflows. Metadata only — no prompts/outputs.
ComplianceAudit trailsEnterprise only
$50K-200K/yr
Connected by OpenTelemetry

One instrumentation. Switch platforms = 2 config lines, not your codebase.

Remember This

1. Products are converging on features. What still differs is buying posture: open-source vs managed, provider-native vs provider-agnostic, quality tooling vs governance. Pick the posture that matches your maturity stage.

2. The metric that matters for tool selection isn't feature count — it's time to first trusted regression signal: how long from a code change to a cross-functional team seeing whether quality improved or regressed.

3. Engineers and PMs hunt for different things in the same traces. The tool must serve both — or the PM retreats to spreadsheets and the eval flywheel dies.

References

1. LangSmith: Deployment Options — LangChain

2. Braintrust: Self-Hosting — Braintrust

3. Arize Phoenix: Self-Hosted Privacy — Arize

4. Getting Started with Evals — OpenAI

5. How to Choose an AI Evaluation Service — Kili Technology

Previous Topic Back to the Deep Dive