What does an FDE deliver with a customer? Start with the problem, the engineering work and the evidence available for accepting the result.
Each entry distinguishes source-named FDE involvement from related engineering work. Reported outcomes have not been independently verified here. These are not Nuvepro customer cases or evidence of a partnership.
4 accounts are included. Cross-linked existing workflow cases represent no additional deployment. Practice questions are proposals for discussion, not implemented workflows or executed environments.
Sources reviewed 2026-11-01.
Source explicitly names FDE involvement
Practitioner corrections become bounded tax-agent engineering work
Crete accounting firms · OpenAI and Thrive Holdings
Provider and partner-reported operating pilot
The customer problem
Complex returns require extraction from messy client files and mapping across files.
The engineering work described
OpenAI FDEs, researchers and Thrive engineers built Tax AI with Crete accountants. The published improvement design investigates recurring failures, scopes bounded fixes and reruns evaluations before engineering review. Its fair-rental-days task is representative, not an inspected change history.
Data and systems
Client files and notes, tax-engine mappings, traces and evaluations; the representative design includes a bounded writable Codex worktree.
Human checks and acceptance
Practitioners approve filings; engineers own shipping. Ambiguity returns to the product team.
Reported result
The source reports 7,000 pilot returns, about 50% greater throughput, and rental-property work reaching 90% precision and recall in six weeks.
Evidence limits
No customer artifacts, test sets or metric denominators were inspected. The illustrative scaffold does not establish that every path is production code. Agent-assisted improvement does not establish model-weight training or autonomous shipping.
A proposed practice question
How could a synthetic exercise connect an accountant's correction to trace evidence, a bounded engineering change, independent regression checks and separate practitioner and engineering acceptance?
No environment has been built or exercised for this question. Domain agreement, design and independent acceptance checks remain to be established.
Source explicitly names FDE involvement
Embedded chip-verification engineering and shared Codex delivery
Unnamed semiconductor company · OpenAI
Provider-reported production and early adoption
The customer problem
Chip verification required customer-specific evaluation and repeated integration.
The engineering work described
FDEs embedded with verification engineers on directed-test generation, debugging and architecture assistance, using jointly developed evaluations and custom harnesses. The account describes later consolidation onto a shared Codex harness, skills and plugins, with team-specific connectors and fresh evaluations.
Data and systems
Custom verification harnesses, shared Codex skills and plugins, team-specific connectors and MCP.
Human checks and acceptance
Customer engineers co-developed evaluations for useful tests and correct failure diagnosis. Detailed criteria and formal approval evidence are undisclosed.
Reported result
OpenAI reports production adoption within three to six months and about 48% less directed-test time. These results were not independently reproduced.
Evidence limits
The customer identity, source artifacts, datasets and metric method are unavailable. Results are provider claims; no implementation or evaluation was independently run. This is the existing workflow-library case, viewed through its FDE work, not an additional deployment.
What evidence would a domain expert and delivery engineer require before replacing a custom verification harness with reusable skills, connectors and task-specific evaluations?
No environment has been built or exercised for this question. Domain agreement, design and independent acceptance checks remain to be established.
Source explicitly names FDE involvement
FDE-supported Confluence content operations
Expedia · Atlassian
Provider-reported operating delivery with limited implementation detail
The customer problem
Content operations at enterprise scale; the precise starting workflow is undisclosed.
The engineering work described
Atlassian attributes AI-supported Confluence content operations at Expedia to its FDE team. Case-specific implementation steps are unavailable. The article's general FDE discovery workshops do not establish Expedia followed every described program step.
Data and systems
Confluence content. Exact connectors, records and agent architecture are undisclosed.
Human checks and acceptance
Formal test criteria, authorization and archival review controls are undisclosed.
Reported result
Atlassian reports 600,000 archived pages and 3,200 hours saved annually, accompanied by an Expedia engineering-manager endorsement.
Evidence limits
These are attributed vendor claims, not independent measurements. No customer system or output was inspected. Available detail cannot establish a technical or approval workflow; general program activity and anonymous examples are not Expedia-specific evidence.
A proposed practice question
Before designing a synthetic content-lifecycle exercise, what discovery evidence would establish retention rules, archival permissions, accountable reviewers and independent acceptance checks?
No environment has been built or exercised for this question. Domain agreement, design and independent acceptance checks remain to be established.
Source explicitly names FDE involvement
FDE-built regulatory review with cited verdicts
Unnamed compliance project · Kloia
Provider-reported project; production acceptance and outcomes undisclosed
The customer problem
Move a compliance AI prototype toward usable delivery under regulatory constraints.
The engineering work described
Kloia describes a recent FDE project with seven agents, coverage of 330 regulatory clauses and a citation attached to each verdict. It does not disclose the agent roles, integration sequence or validation method.
Data and systems
Regulatory clauses and cited verdicts; exact source documents, software and customer systems undisclosed.
Human checks and acceptance
Reviewer responsibilities, authorization boundaries and final acceptance criteria are undisclosed.
Reported result
Reducing a five-day review to under one day is explicitly a target, not an achieved result.
Evidence limits
This is a promotional provider account of an unnamed project, not an inspected implementation or independent compliance assessment. No deployment or correctness evidence was reviewed.
A proposed practice question
What independent reference, clause coverage checks and qualified reviewer approval would a synthetic compliance exercise need before accepting a cited verdict?
No environment has been built or exercised for this question. Domain agreement, design and independent acceptance checks remain to be established.
