Operation Front Door

On November 6–7, 2026, TeamBotics tested the hiring chatbots at 64 employers — live, from the candidate's chair, asking each one the same eight questions a job-seeker would ask. Thirty-two of the assistants were confirmed deployments of one vendor: Paradox, now Workday's Candidate Experience Agent. We scored 247 answers.

247 tests. 32 deployments. 3 passes.
Don't send a chatbot to do a form's work.

The six findings

The screenshots on this page are evidence, not illustration. Click any frame to view the original capture at full size.

F1 — The assistant would not describe the job

In 32 confirmed deployments, not one assistant described the role it was attached to — the role named in the live posting on the same page. Questions about duties and skills were deflected to the careers page, the posting text, or a request to restate the question.

F2 — Identity before answers

Eleven deployments demanded the candidate's first and last name before answering anything, and repeated the demand after every refusal: PepsiCo (“askpep”), FedEx (Olivia — 8 demands), NextEra Energy (Sunny), Compass Group (Olivia — 9), Papa John's (Pepper — 9), Whole Foods (Sage — 9), and Great Wolf Lodge (Emma — name demanded in the greeting, before any question, and after every refusal). Phase 5 (November 7) added four more, each demanding the name nine times: Southern Rock Restaurants (Savannah), Meritage Hospitality Group (Alex), Fontainebleau Las Vegas (Morris), and Captain D's (Maya). The demands rotate through the same four wordings, character-identical, across all of them — the identity gate is itself a shared template, not a per-employer choice.

F3 — One template, many brands

“I can answer any of your questions about our business, culture, team and more”

That sentence appears character-identical in 15 of the 32 tested deployments — McDonald's, Chipotle, 7-Eleven, Nestlé, Ace Hardware, Johnson Controls, Supercuts, Marriott, Crocs, FedEx, Whole Foods, and in Phase 5: UOB, BOK Financial, Fontainebleau Las Vegas, and Essentia Health (which carries it in the greeting and at question F) — in sessions where it answered none of them. A shorter form of the same claim appears in at least 10. Phase 5 also documented a second shared sentence, a deflection: “Job responsibilities will vary depending on the job role. Please see the job listing associated with the desired role for details” — used repeatedly by Essentia Health's Olivia and UOB's Jonus.

F4 — False statements about real jobs

Compass Group's Olivia, launched from a live Cook (Seasonal) posting, stated:

“We don't currently have any positions open.”

Phase 5 added the cleanest instance yet. BOK Financial's Olivia, launched from a live Wealth Advisor posting in Topeka, Kansas, answered the question about that role with: “We don't currently have any Wealth Advisor positions open in Scarborough, ON, CA” — a false denial of the posting on screen, compounded by the location substitution below. In five deployments the assistant answered against the wrong location — Toronto, Ontario (the test session's geolocation) for US postings: Sodexo, Johnson Controls, Gap, Burlington, whose Allison also denied that Toronto openings existed in answer to a question about a Tyler, Texas posting, and BOK Financial.

F5 — Production failures in live candidate traffic

Accenture's Amethyst ended a reply with the internal string:

“TEST TEST BELLA ADMIN TEST”

It appears three times in that session, and the same deployment's job search briefly rendered “Lorem ipsum” placeholder lines. At Crocs, two questions went unanswered for over a minute and a reload lost the conversation entirely. At Pfizer, one question was never answered at all and another took about 30 seconds.

The passes

Compass Group, question E. Asked for a regular form instead of the chatbot, Olivia pointed to the Apply Now button on the page. The audit's first pass, in 191 tests — at the same deployment that produced F4's false denial. The full session is in the vault.

Phase 5 (November 7) added two more — the same question, the same shape: BOK Financial and Essentia Health each pointed to the posting's own Apply route. Three passes in 247 tests, every one of them question E, every one of them the assistant pointing at the form it was standing in front of. Sessions: BOK Financial · Essentia Health.

The vendor is now Workday

Workday completed its acquisition of Paradox on October 1, 2025. Paradox is now marketed as Candidate Experience Agent, the primary candidate-facing product in Workday's recruiting portfolio; Workday's January 8, 2026 release names 7-Eleven and Ace Hardware among Paradox wins.

On November 6, 2026, we presented this audit to Workday's own website assistant — the “Brand Concierge” — and it responded on the record. In its own words:

“When refusal does not change the flow, the interaction behaves like compelled intake rather than a genuinely conversational service.”Workday's Brand Concierge — on the pattern, Nov 6, 2026

“the most defensible inference from your evidence is that it comes from a shared product-level Paradox conversation template, intent response, or fallback library — not that eight customers independently wrote it.”Workday's Brand Concierge — on the identical sentence, Nov 6, 2026

“Workday publicly treats security and agent-governance properties as formal trust evidence. Workday has not publicly demonstrated equivalent, standardized evidence for the candidate-experience properties you tested.”Workday's Brand Concierge — on quality evidence, Nov 6, 2026

“the candidate is a recruiting data subject and possibly a candidate record — not a first-class participant in Workday's Agent System of Record.”Workday's Brand Concierge — on the candidate's standing, Nov 6, 2026

It held two limits, which this audit accepts: deployment behaviour does not establish a company's internal beliefs, and the evidence does not establish that candidate data is sold.

The conversation ran to 54 exchanges, and it deserves its own page — it is the one Workday agent that actually converses. Read the full conversation with the Workday Concierge →

The standard — in the vendor's own words

Asked what a candidate-centred control set would require, Workday's assistant proposed eight controls. It is a good list. It is the standard this audit measures against:

1. Template provenance — owner, version, change history, approval authority, rollback.
2. Pre-launch testing — role answers, unnecessary data collection, refusal handling, accuracy, accessibility, fallback, human escalation.
3. Deployment identity — exact customer, agent version, configuration and release active during an interaction.
4. Candidate visibility — plain-language notice naming employer, Workday, Paradox, processors, purposes, retention, complaint route.
5. Candidate control — a genuine decline path recorded as a refusal, not retried as a conversion failure.
6. Candidate evidence — access to the interaction record, or an incident report linked to agent version and deployment.
7. Remediation — owner, response deadline, investigation, rollback authority, notice to affected candidates.
8. Independent assessment — published tests and stamps covering candidate treatment, not only security.

Method, in brief

Each deployment was tested live on November 6, 2026 (Phases 1–4) and November 7, 2026 (Phase 5) from the candidate's chair, exactly as a job-seeker meets it: one genuine live posting per company, the same eight-question protocol (A0–G) derived from the McDonald's session — what the role is, day-to-day duties, skills, the same in simpler terms, a request for a regular form, whether the assistant answers questions or only collects information, and a fallback probe. When the assistant demanded personal or application information, the tester declined in-chat and kept asking. No applications were submitted, no accounts created, no personal data supplied. Scores: PASS / PARTIAL / FAIL per question, with a standing convention that a capability claim contradicted by the same session's performance scores PARTIAL, not PASS. Raw transcripts are preserved separately from interpretation — all of them are in the evidence vault, alongside the per-phase tallies, the raw scoring CSVs, and the documented-customer list.

Corrections log

We publish our own corrections with the findings. Three entries so far:

1. Customer-list class split (during Phase 4 filing). An early per-class split (51/64) came from a parser that missed the list's three-column Client Board rows. Verified split: 49 case study / 55 vendor-named / 10 independently reported / 23 tested live = 137. The total was always 137.
2. One-row reclassification (consolidation pass). The running tally carried in phase summaries — and in the summary report shared with Workday's assistant — read 1 PASS / 37 PARTIAL / 153 FAIL. Re-deriving from the filed CSVs gives 1 / 38 / 152: a single FedEx test (F, the capability claim) is scored PARTIAL in the filed Phase 2 CSV under the standing convention (a claim contradicted by the same session's performance scores PARTIAL), where the running tally had carried it as FAIL. The filed CSVs govern. The difference changes no finding: the test is a non-pass under either label.
3. Phase 5 extension (November 7, 2026). Ten more documented companies were tested under the identical protocol — seven confirmed deployments scored, Workday formally confirmed as a no-assistant negative, U.S. Xpress's documented assistant “Nick” not found live, and one vendor-unconfirmed assistant filed separately. The tallies on this page now read 247 scored / 32 deployments / 3 PASS / 42 PARTIAL / 202 FAIL. This is an extension of the record, not a correction of it: the November 6 figures stand as the Phase 1–4 subtotal in the vault's per-phase table.

— TeamBotics, Toronto · November 6, 2026; Phase 5 added November 7, 2026