An Approval Queue for AI Outreach Can Only Check What the Reviewer Can See
Per-message approval of AI recruiting outreach catches context errors only when reviewers see the record behind a draft. It cannot catch an opt-out missing from the data, and some risks are better designed out than reviewed.

Workday says outreach from its recruiting agent still needs recruiter approval, while describing agent-sent messages as a path it intends to offer. Three written accounts suggest the setting matters less than what a reviewer can actually check: a reviewer shown the candidate's record can catch context errors, no reviewer can catch an opt-out missing from the data, and some risks are better designed out than reviewed.
Watch the analysis
On Oct. 13, 2026, Workday used its Rising conference to announce autopilot capabilities for its Talent Acquisition Agent. The press release describes an agent that finds candidates as soon as a job is posted and, when a pipeline runs thin, notifies recruiters “with timely sourcing actions and draft outreach for review.” Then it draws a line: “Sourcing runs in the background, but outreach to candidates still requires recruiter approval before anything is sent.” Early adopter customers can use the autopilot capabilities now, the release says, with general availability planned for early 2027.
An interview video published Oct. 28, produced in partnership with Workday and recorded on the conference's first day, shows how the company thinks about that line. In it, Workday's director of AI product for its HCM business described a recruiter's morning shifting from sourcing and drafting to an inbox holding “a couple dozen outreach messages to approve.” Customers would differ, the director said: “some customers are going to want every outreach message to be approved by a recruiter.” Others would say, “We're okay with the agent actually sending out the messages.” The director added: “I think both are okay.”
The two statements leave open whether customers can switch on automatic sending of candidate outreach today, and neither offers customer outcome data. They do make clear who decides. Customers, the release says, “decide which actions an agent can take on its own, which need review, and when a person makes the final call.”
That framing treats approval as a single dial: how much of it to keep. For the talent-acquisition and HR compliance teams who will set it, a more useful question comes first. When an AI tool can contact candidates or leads for you, what does approving each message actually catch, and what has to be enforced in the data or in the design instead? Three written responses to PSEEDR suggest the answer depends on where the fact that makes a message wrong actually lives. A reviewer can catch it only if it is in front of them. If it never reached the data, closer reading will not find it.
Errors a reviewer can see, if the reviewer has the record
Pratik Singh Raguwanshi, manager of digital experience at CXrove, describes what happened when approval was loosened. “Three years ago, we relaxed human approval to meet aggressive volume targets for a specialized sourcing operation, only to see engagement rates crater and negative feedback rise,” Raguwanshi wrote. In this account, the AI “scraped data accurately but failed the logic test.” It “frequently congratulated candidates on ‘promotions’ that were actually lateral moves during corporate layoffs or used a tone far too casual for senior executive targets.”
Both errors were in the text of the message, but neither was visible from the text alone. A congratulation on a new title reads well in isolation. It is wrong only against the facts behind it: the move was lateral, and it happened during layoffs. A breezy tone suits some recipients and offends others, and a reviewer can judge it only by knowing who the recipient is. A reviewer checking grammar and phrasing would pass both messages.
The team's fix fits that reading. It moved to what Raguwanshi calls a “logic audit” model. “We stopped checking for minor syntax issues and focused strictly on contextual relevance, compliance markers, and privacy norms,” Raguwanshi wrote, describing the reviewer as “a final sanity check for empathy rather than a traditional editor.” Response rates recovered, according to the account, which gives no figures and dates from roughly 2023.
The case supports something narrower than “always keep a human in the loop.” Relaxing approval to hit a volume target removed the human check on a type of error the model kept making. The check became useful again when it got a specific job. PSEEDR's reading is that the job has a condition attached. “Does this message fit this person's situation?” can be answered only if the review screen shows the record the draft draws on, such as the career history it congratulates or the seniority of the person it addresses. Raguwanshi does not say what the reviewers were shown, so that condition is inferred from the type of error, not reported. Nor does the account show that this kind of review holds up at any volume.
A fact the reviewer cannot see
Lilach Bullock, an AI implementation consultant and fractional CMO (lilachbullock.com), describes a failure that closer reading could not have caught. “My follow-up agent sent one lead a follow-up message four days after she had asked to be removed,” Bullock wrote. The case involved “my own sales leads rather than candidates,” so it is not a recruiting incident. It belongs here because the mechanism carries over to any agent that writes to people from a contact record.
The lead's opt-out had been logged somewhere the agent could not see; Bullock gives a phone call recorded as a note as an example of such a place. “Approving each message wouldn't have caught it, because a reviewer can't spot a fact that's missing from the data.”
Nothing in a follow-up's wording signals that its recipient has asked to stop. The fact that made this message wrong sat outside the data the agent drew on. If a reviewer sees the draft and the record the agent used, neither shows the request, and reading them more carefully cannot supply it.
Bullock moved the check to the front of the process. “Before I let the agent run unsupervised, I made do-not-contact a hard field that the agent checks first, and that fixed that failure.” Bullock is careful about the evidence: “The four days is measured, and it covers one case rather than a rate.” For recruiting teams, Bullock suggests giving HR compliance staff the fairness, consent and employment-law checks, “with the same hard opt-out field sitting above everything else.” Bullock has not run such candidate checks and claims none.
The order matters, and PSEEDR draws one further inference from it. A suppression field checked before drafting means a contact who asked to stop never produces a draft, so there is nothing for a reviewer to approve or overlook. But the field knows only what someone writes into it. An opt-out given on a call, in an email reply or to a colleague still has to reach that field. The control does not remove the work. It moves the work from inspecting messages to making sure every channel through which someone can ask to stop updates the one record the agent must consult.
Opt-out rules give this some context, though they do not fit recruiting neatly. For commercial email in the United States, the Federal Trade Commission's CAN-SPAM compliance guide says opt-out requests must be honored within 10 business days. The guide says the law makes no exception for business-to-business email and that a company cannot “contract away” its legal responsibility by hiring another company to handle its email marketing. In the EU, Article 21(3) of the GDPR says that once a person objects to direct marketing, their personal data “shall no longer be processed for such purposes.” Bullock gives no jurisdiction for the case, and this article does not assess whether any of these rules applied to it.
For candidate outreach, the fit is less certain. The FTC lists messages that provide “information about an employment relationship or employee benefits” among its narrowly read transactional categories. The sources reviewed here do not establish whether a cold recruiting email counts as commercial email. GDPR Article 21(1) gives a more general right to object to processing based on legitimate interests; how it applies to sourcing is a question for counsel. The practical point does not depend on the legal one: a stated wish not to be contacted should stop drafting, not wait for a reviewer to notice it.
Will reviewers keep reading as the queue grows?
For errors that a reviewer could catch, a second question follows: will reviewers keep catching them as volume rises? None of the respondents measured this. Bullock says there is no measured figure for how closely drafts are read as volume grows.
Research in other fields describes the mechanism, though not its size in recruiting. A 2012 systematic review in the Journal of the American Medical Informatics Association by Kate Goddard, Abdul Roudsari and Jeremy Wyatt examined automation bias, “the tendency to over-rely on automation.” Of 13,821 papers retrieved, 74 met its criteria, drawn from several research fields with a focus on healthcare. The review identified workload, task complexity and time constraints as environmental factors that strained users' cognitive resources and shaped reliance on the system, and it counted trust and confidence among the attitudes that did the same. Mitigations included training, emphasizing user accountability and design choices, such as where advice appears on screen and whether the system offers information rather than a recommendation.
A 2022 paper by Ben Green reviewed 41 policies requiring human oversight of government algorithms. According to its abstract, Green argues that evidence suggests people cannot perform the oversight such policies intend, and that the policies can lend legitimacy to flawed systems. The EU AI Act names the same risk. Article 14(4)(b) requires that the people assigned to oversee a high-risk system be enabled “to remain aware of the possible tendency of automatically relying or over-relying on the output produced by a high-risk AI system (automation bias).”
None of this research measures recruiters reviewing outreach drafts. It identifies conditions: workload, time pressure and the user's trust in the system. Workday's release opens with the example of a recruiter managing 25 open jobs, and the interview pictures a couple of dozen drafts waiting each morning. Neither is a measured queue size. A queue that grows with every open requisition is where the conditions in the research become plausible.
The interview's advice for new deployments adds a possible tension. The director recommended starting “least autonomous” and building up, with agent memory and audit trails helping customers grow more comfortable over time. That approach asks reviewers to ease off as the agent proves itself, and the review literature counts trust among the factors that drive over-reliance. A run of good drafts could teach reviewers to approve faster just as the system meets an unusual case. PSEEDR offers that as a hypothesis to watch for, not a finding. Workday's release says every agent action is logged, including “who approved it.” Such a log can show that an approval happened. It cannot show whether the approver compared the draft with the candidate's record.
Removing the risk instead of reviewing it
Shawn Mintz, CEO of mentoring-software company MentorCity, approaches the question from outside recruiting. “I am not a recruiter, so take this from the other side of the table,” Mintz wrote. During 2026, MentorCity added AI to a product where the software could have written and sent messages to participants in a mentor's or an administrator's name. “We decided it would not.” The company's AI mentor, Lexa, speaks as itself and asks questions instead of handing out advice. Matching stays with participants and administrators, and administrators can turn the AI on or off for each program.
Mintz's reasoning addresses a third type of error: a promise made in someone's name. “Anything AI produces, sent outside your organisation under your name, can promise something you do not actually do, and you own that promise,” Mintz wrote. The company chose not to rely on a reviewer to catch such promises because, in Mintz's judgment, “real reading drops off the moment volume climbs.” The conclusion Mintz draws: “if software can act in your name, the approval step is the product, not the overhead.”
Mintz reports no measured numbers. This is a design choice in Mintz's own company's product, so the case shows that the choice is possible, not that it reduced errors. Recruiting also differs from mentoring. A first approach to a candidate often comes from a named recruiter, so an AI speaking as itself will not fit every message. The design question still applies: which messages actually need a person's name on them? Where one does, it needs real review. Where one does not, the AI can still say something wrong, but it no longer commits a named person to it. That narrower risk is designed out rather than left to an approval queue.
What the rules do and don't decide
Teams hoping that AI-hiring law will settle these choices will find less than they might expect. New York City's Local Law 144 requires bias audits and notices for automated employment decision tools. The city's enforcement FAQ, dated June 29, 2023, says those requirements do not apply when a tool is used to “scan a resume bank, conduct outreach to potential candidates, or invite applications.” They apply to tools used to assess people who have applied for a specific position, or employees being considered for promotion. The FAQ also notes that federal, state and city anti-discrimination laws still apply.
In the EU, Annex III of the AI Act lists as high-risk AI systems intended for recruitment or selection, “in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates.” Whether a particular outreach-drafting agent falls into that category is a classification question that this article does not try to answer. An unofficial explorer of the Act lists Dec. 2, 2027 as the date when the Article 14 oversight duties begin to apply to Annex III systems.
These rules may shape parts of a recruiting workflow. But at least one prominent local AI-hiring law leaves outreach out of scope. Under that law, how an employer approves outreach and suppresses unwanted contact is a design choice, often made through a vendor's configuration settings, within whatever privacy, marketing and anti-discrimination rules apply. This article does not provide legal advice for any jurisdiction.
The decision: require, relax, or move the check
Taken together, the three accounts split the approval question into separate parts.
| Type of error | Evidence | Where the deciding fact lives | Can a per-message reviewer catch it? | Where the control belongs |
|---|---|---|---|---|
| A wrong reading of the person's situation | Raguwanshi: “promotions” that were lateral moves during layoffs; a tone too casual for senior executives | In the candidate's record | Yes, if the reviewer is shown that record and asked a context question | Keep approval and define the reviewer's task; do not relax it to hit volume |
| A fact missing from the data the agent reads | Bullock: an opt-out logged where the agent could not see it; a follow-up four days later (sales leads, not candidates) | Outside the fields the agent reads | No | A hard do-not-contact field checked before drafting, fed by every channel where someone can opt out |
| A promise made in someone's name | Mintz: AI messages sent under a person's name can commit to what the organization does not do | In what the sender actually intends | Only with close reading, which Mintz judges declines with volume | Design: decide which messages need a person's name at all |
| Declining attention | Automation-bias research in clinical and government settings; no recruiting measurement | In the reviewer's workload and trust | This is the weakness of review itself | Keep the queue reviewable, give each review a narrow job and watch for faster approvals as trust grows |
The answer to the opening question follows from that split. Approving each message catches what a reviewer can see, and only if the review screen shows the record the draft relies on. It cannot catch what the data never held. A decision to relax approval for volume should therefore start by asking which type of error the remaining reviews were catching. Raguwanshi's account suggests that context errors return when review is removed. Bullock's case shows why a do-not-contact rule belongs ahead of drafting, where no reviewer is needed to enforce it, as long as every opt-out reaches the field it checks. Mintz's design suggests that some risks need not be reviewed at all if the AI does not speak in a person's name. Workday may offer both paths, but the more useful configuration choice is which check sits at which point in the process.
That leaves one question open. If a team keeps per-message approval, what record would show whether reviewers are still reading? One candidate is the share of drafts edited or rejected, tracked as the queue grows. None of the accounts here reports such data. Teams that keep those logs could show whether approval at volume still works as a check or has become a click.
Reporting note: Pratik Singh Raguwanshi, Lilach Bullock and Shawn Mintz supplied written responses to PSEEDR. Their accounts were not independently verified. Bullock's case concerns sales leads, not candidates; Bullock discloses earning from sponsored content and paid placements on the lilachbullock.com website. Mintz describes MentorCity's own product. The Workday interview is a partnered video, and Workday's statements are vendor claims. The automation-bias research cited here comes from fields other than recruiting. Nothing in this article is legal advice.
Sources
- https://investor.workday.com/news-and-events/press-releases/news-details/2026/Workday-Launches-Autopilot-Agents-That-Get-Real-Work-Done-Across-HR-Finance-and-IT/default.aspx
- https://www.youtube.com/watch?v=szLOSHwVWtY
- https://www.ftc.gov/business-guidance/resources/can-spam-act-compliance-guide-business
- https://gdpr-info.eu/art-21-gdpr/
- https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/
- https://arxiv.org/abs/2109.05067
- https://artificialintelligenceact.eu/article/14/
- https://www.nyc.gov/assets/dca/downloads/pdf/about/DCWP-AEDT-FAQ.pdf
- https://artificialintelligenceact.eu/annex/3/