Two job titles, two business models, one operational rubric that tells them apart.

It depends on the firm. The Big Four are now hiring under the FDE title; EY publicly launched 45 to 50 FDE roles for the UK and Ireland on 2026-04-28, focused on AI in production, governance, and risk. The role at Palantir, where the term originated in the early 2010s, was different: an engineer with commit rights, on the customer's stack, shipping code as the primary deliverable. The honest test in 2026 is operational, not titular. Ask the firm to put two specific lines in the SOW: a named-engineer commit-rights clause for week 1, and a numeric eval threshold wired to a CI gate that pauses billing on miss. A real FDE engagement signs both. A repackaged consulting engagement adds a meeting to discuss them.

Six files in a PIAS engagement: rubric.yaml (the numeric thresholds and gate config), .github/workflows/pilot-gate.yml (the week-2 required check), .github/workflows/production-gate.yml (the week-6 required check on the promotion PR), eval/cases.yaml (30+ rows of test cases that grow weekly), runbook/<agent>.md (the on-call playbook with five canonical sections, 300+ words), and flags/<agent>.yaml (the rollout_percent feature flag the promotion PR flips from 10 to 100). All six are version-controlled, all six are in your repo, none are hosted by us. If the engagement ends and we never come back, the agent keeps shipping because the gates, the rubric, and the runbook all live on your main.

pilot-gate.yml has a refund-signal job that runs only on the scheduled Monday 09:00 UTC snapshot, and only when the eval scoring step failed. The job posts the failing snapshot to a billing webhook configured in rubric.yaml, which pauses the open invoice. It then opens a GitHub issue against the engagement owner, labeled refund-triggered and billing-paused, linking the failing workflow run. No human clicks anything. If the next Monday snapshot is green, billing resumes and a 7-day prorated credit is applied. If it misses three Mondays in a row, the engagement ends per the MSA exit clause, every PR merged to date stays in your repo, and you owe nothing for the final week. A consulting acceptance clause cannot do this; it requires written notice and a negotiation.

Three questions, no jargon. (1) Will you put the engineers' GitHub handles in the SOW, with a specific date by which they will be added to CODEOWNERS in our repo? (2) Will you accept a numeric eval threshold in the contract, wired to a CI gate, where missing the threshold pauses your invoice automatically? (3) When the engagement ends, what files are in our repo and what files are in yours? A real FDE answer is yes, yes, and 'every artifact is in your repo, none are in ours.' A consultancy answer is usually about access policies, acceptance reviews, and the firm's accelerator library being licensed on a separate Order Form.

Three concrete differences. First, scoping: every PIAS engagement starts from a free 60-minute scoping call that produces a written one-pager naming the agent, the rubric thresholds, the feature flag path, and the named senior engineer. Second, contract shape: a six-week engagement, week-1 first-PR clause, week-2 numeric refund clause, week-6 production gate, and a no-handoff-no-invoice tail clause. Third, leave-behind: zero PIAS-hosted runtime, zero platform license, zero seat fee. The Big Four offerings we have read tend to bundle a managed runtime, a per-seat accelerator library, and a 12 to 16-week phased plan. Both can be the right call. The honest difference is what the contract says about week 7 onward.

Because the bottleneck is usually senior MLE capacity on a specific production agent, not advisory hours. If your internal team has time and skill to write rubric.yaml, both gates, the runbook, and the cases corpus, and to keep the eval green for the trailing seven nightly runs, you do not need an FDE. If the constraint is that two senior engineers would need to be pulled off another deadline for six weeks to do that work, an FDE is the cheapest way to ship without breaking the other deadline. The win is that the leave-behind is identical to what your team would have built, so internal ownership is intact when the engagement ends.

Cleanly, in the contract. Every PR merged to your main is your IP, work-for-hire, and there is no carve-out for the firm's accelerator library because there is no accelerator library. The PIAS engineering practices (the file shapes, the rubric format, the gate workflows) are open-source patterns we publish on the public guides at fde10x.com/t. They are not licensed back to you, you owe no fees for using them, and you are free to fork and rename. The only thing the firm keeps is the ability to publish a redacted case study, with your written consent, after the engagement ends.

Six weeks for a single agent shipped to production with a runbook, an eval harness, and a feature-flag flip. Two weeks for a contained eval-harness retrofit on an existing agent. Twelve weeks for a multi-agent orchestration with cross-agent eval, when the rubric carries multiple agents. Anything longer than 12 weeks usually means the engagement is shaped like a consulting retainer, the senior engineer is partly on advisory work, and the leave-behind is harder to point at. We say no to those.

Two specific people, not a bench. Their full names and GitHub handles are in the SOW. They are added to CODEOWNERS in your repo on day 1 of week 1. They show up in your standup or async standup channel. They open PRs from their own GitHub account. They are the same two people on day 1 and day 42; if either rolls off mid-engagement, the engagement is invoiced at zero for any partial week without that engineer present, and replacement is your call (continue with one engineer, accept a substitute, or pause the engagement). The opposite pattern, a rotating bench staffed from a regional pool, is what most repackaged FDE offerings actually deliver.

When the deliverable you actually need is a recommendation, a vendor-selection memo, an audit report, or a regulatory filing. When you have to convince a board that a particular AI strategy is sound and you need an external authority on the deck. When you need a firm-name signature on a process audit. When the work is genuinely advisory, a senior consultant with no commit rights is the right hire. The FDE pattern fits when the bottleneck is a missing senior engineer on a specific shipping deadline, and what you need is code on main, not a recommendation.