How to Build Your Own Enterprise AI Benchmark — After Nadella’s Paradox
Satya Nadella says own private evals inside a trust boundary. Step-by-step: what enterprises should measure, task design, scoring, harness isolation, model-agnostic runners, and the compound learning loop — not vendor leaderboards.
"Create your private evals, because evals define what 'good' looks like inside the organization."
This guide is the derivative playbook — how to build that eval set, what to measure, and how to run it inside a trust boundary so intelligence exhaust does not become your competitor's training data.
TL;DR — enterprise benchmark vs public leaderboard
explainx.ai translation: You are not buying a model score. You are buying a measurement instrument you control — the enterprise equivalent of patents in Arrow's original paradox.
Anatomy of a private enterprise benchmark
snippet
┌─────────────────────────────────────────────────────────┐
│ CHARTER — accountability + workflow scope │
├─────────────────────────────────────────────────────────┤
│ GOLDEN TASKS — private scenarios (not public SWE ids) │
├─────────────────────────────────────────────────────────┤
│ SCORING — tests + rubrics + human slices │
├─────────────────────────────────────────────────────────┤
│ HARNESS — your agent loop, tools, MCP, tenant boundary │
├─────────────────────────────────────────────────────────┤
│ RUNNER — model-agnostic (API swap without rewrite) │
├─────────────────────────────────────────────────────────┤
│ REGISTRY — git-versioned evals + results artifacts │
├─────────────────────────────────────────────────────────┤
│ COMPOUND — failure → trace → new case → re-run │
└─────────────────────────────────────────────────────────┘
Step 1 — Charter one workflow (not "AI in general")
Pick one high-stakes loop first:
Good first targets
Avoid as v1
Tier-1 incident triage
"All of engineering"
KYC / compliance doc review
Generic chat quality
Internal API migration
Trivia Q&A
Sales engineering RFP slice
Copy-paste SWE-bench tasks
Charter template:
markdown
## Workflow: [name]- Accountability: [regulation, SLA, security class]
- Success: [measurable outcome in production terms]
- Failure modes: [what must never happen]
- Human gate: [when review is mandatory]
- Models in scope: [list candidates — min 2 vendors]
Tie success to outer alignment — intent your legal team would sign, not a leaderboard column.
Step 2 — Harvest golden tasks from real work
Public benchmarks fail enterprises because they lack particular knowledge (Hayek) — time, place, and circumstance only your org holds.
Senior SWE-Bench (Snorkel) validates the pattern: Slack-style prompts, runtime investigation, taste scoring — not 6,000-character specs. Your enterprise set should look more like Senior SWE-Bench or GeneBench-Pro (domain judgment) than generic HumanEval.
Target size:20–50 tasks for v1 — enough to discriminate, small enough to maintain.
Trust rule: Anonymize customer PII; keep structure of failure. Never paste raw tasks into ChatGPT/Claude web UIs with default retention.
Cursor reward-hacking lesson (SWE-Bench eval contamination): if the harness leaks hints or shares infra with training, scores lie. Freeze harness version per eval release.
# Pseudocode — same eval id, swap model via env
EVAL_SET=./evals/v2026-07-13 \
HARNESS=./agent-loop \
MODEL=gpt-5.6-sol ./run-eval.sh
MODEL=claude-fable-5 ./run-eval.sh
Procurement question Nadella implies: If either model vanishes tomorrow, do our evals still run?
Step 6 — Version and govern the registry
Artifact
Storage
Task definitions
Git (private repo)
Pass criteria
Same commit as tasks
Harness version
Tagged with eval release
Results
S3/GCS + dashboard you host
Corrections
Ticket → candidate task PR
Access: eval authors ≠ production deployers — separation like code review.
Cadence: re-baseline monthly or on major model upgrade — July's limit-reset week showed vendor behavior shifts overnight.
Step 7 — Compound loop (Nadella's hill climber)
snippet
Agent run → fail eval → trace review → root cause
↓
New regression task OR harness fix OR prompt/skill update
↓
Re-run full suite → compare to baseline tag
This is loop engineering applied to governance — not a one-off consultant deck.
Feed corrections back as tasks — Nadella's exhaust becomes your training signal, not OpenAI's.
What good looks like — sample enterprise scorecard
Metric
Q3 target (example)
Golden task pass@1
≥ 70% on Tier-1 workflow
Regression suite
100% on prior Sev-1 incidents
Human edit ratio
↓ 20% vs baseline
$/merged PR (agent-assisted)
Below Sonnet-only baseline
Compliance hard-fails
0
Model swap delta
< 5 pts pass rate between primary + backup
Report trends, not vanity peaks — JPMorgan's backtest warning applies to evals too: in-sample heroics ≠ live durability.
We teach this stack in workshops: charter → harness → skills → private evals → orchestration. Public leaderboards are R&D signal. Your benchmark is operating system.
Benchmark design reflects Nadella's July 2026 framework and explainx.ai enterprise eval practice as of publication. Adapt scoring weights to your regulatory context with legal and security review.
