Agentic AI evidence library

Case studies are useful only when their claims can survive diligence. A customer logo, a polished quote and an unqualified percentage do not explain the system boundary, measurement method or risk accepted. Trunnion does not publish placeholder customer stories. Named deployment pages will be added only after the evidence and publication rights exist.

What should every case study disclose?

Evidence field	Minimum disclosure	Why it matters
Workflow	Trigger, decisions, tools and side effects	Defines what was actually automated
Boundary	Included systems, data and users	Prevents broad claims from a narrow pilot
Stage	Evaluation, pilot or production	Separates learning from operating proof
Release	Version, date and configuration	Makes the claim time-bound
Controls	Policy, approval, identity and audit design	Shows how authority was bounded
Tests	Ideal, denial, tamper and failure paths	Demonstrates behavior beyond a demo
Measures	Definition, denominator and observation window	Makes outcomes interpretable
Limitations	Known gaps, exclusions and residual risk	Prevents evidence from becoming advertising
Approval	Customer and Trunnion publication owner	Establishes permission and accountability

If a field cannot be disclosed, the study should say so and narrow the conclusion.

Which deployment stages should not be conflated?

An internal prototype demonstrates feasibility. An evaluation measures behavior under a defined test plan. A pilot introduces representative users or data within a limited boundary. Limited production carries real consequence under restrictions. Scaled production adds operational history, support and change control.

Each stage supports different claims. A prototype can show an interaction pattern. It cannot establish production reliability, compliance or return on investment.

How should outcomes be measured?

Define the baseline and unit before collecting data. "Faster" is not a measure. "Median analyst handling time for eligible requests during a defined observation window" can be, if the exclusions and sampling method are retained.

Use a balanced evidence set:

• task success and quality under a reviewed rubric;
• denied and escaped unauthorized actions;
• percent of actions sent for approval and reviewer handling time;
• exceptions, retries, duplicate attempts and reconciliation work;
• audit-event coverage and verification status;
• human rework and escalation rate;
• infrastructure, model and operator consumption;
• incidents, near misses and unresolved limitations.

Report counts with rates. A zero-event claim without the number of opportunities and observation period is weak evidence.

What negative evidence belongs in the story?

Readers need to know what failed. Include denial tests, dependency outages, policy misconfiguration, reviewer backlog, model drift, malformed tool responses and evidence-store failure. Explain the expected safe state and what the system actually did.

A useful case study can conclude that a control or workflow was not ready. Publishing the limitation makes the evidence more trustworthy and shows where engineering effort is going.

How should customer confidentiality be protected?

Use an approved disclosure plan before drafting. Separate facts that are public, facts that can be generalized and facts that must remain private. Remove operational detail that could reveal mission, vulnerability, identity, geography or system topology.

An anonymized study still needs an evidence owner. "A federal customer" is not permission to imply authorization, certification, fielding or endorsement.

What will Trunnion publish when evidence is ready?

Future named studies will follow a stable structure: context, bounded workflow, architecture, controls, test design, observed outcomes, limitations, customer-approved statement, evidence date and related technical artifacts. Every metric will name its measurement basis.

Until then, technical evaluators can use the public reference audit specification, interactive verifier and comparison framework to build a diligence plan without relying on invented social proof.