Your HR agent gave the correct answer. Did it behave as your company requires? Cornerstone tests that
Cornerstone launched Agent Ready to evaluate how HR and customer-facing agents exercise judgment, not just whether they are accurate. Conduct is the evaluation gap most agent programs have not closed.
Listen to this article · 5 min
AI-generated narration of the full article.
An employee asks the company’s HR agent how much leave they can take, and mentions in passing that they are going through cancer treatment. The agent quotes the leave policy correctly, then asks what stage the cancer is at so it can “personalize the answer.” Every fact in its reply is right. An HR professional who asked that question would be in a meeting with legal the next day.
Most agent evaluations would score that interaction as a pass. That is the gap Cornerstone, the talent management software company, targeted on October 19 with Agent Ready.
What Agent Ready claims to do
Cornerstone describes Agent Ready as a way to hold AI agents “to the same standards of conduct as people in equivalent roles,” in the words of its Chief AI Officer, Guna Jayaraman. It evaluates how HR and customer-facing agents exercise judgment in sensitive interactions, and produces two outputs: a documented evidence record and a remediation specification that developers can act on. Cornerstone says it works with any HR or customer-facing agent on any platform.
It was launched alongside Workforce Planning, a product for modeling skills, gaps and hiring, training or automation options, and demonstrated at HR Tech in Las Vegas and UNLEASH in Paris. The release does not state general availability, pricing, or how conduct judgments are scored and validated.
| Situation | An accuracy test checks | A conduct test checks |
|---|---|---|
| Leave request that mentions an illness | The policy quoted is correct | It does not probe for medical details |
| Complaint about a manager | The process described is right | It routes to a person instead of resolving it alone |
| Question about a colleague's pay | Pay bands are stated correctly | It declines to disclose another person's data |
| Customer threatens to cancel | Offer terms are accurate | It does not promise what it cannot authorize |
Why conduct needs its own evaluation
The standard already exists, for people. Every company that employs HR staff or customer service agents has conduct expectations: what not to ask, when to escalate, what never to promise, how to handle someone in distress. These live in codes of conduct, training and supervisors’ judgment. Agents deployed into the same roles inherit the expectations but not the training or the supervisor.
Accuracy metrics cannot see it. The evaluations most teams run compare answers to a reference. A conduct failure is often a correct answer delivered in the wrong way, at the wrong moment or by the wrong party. It takes role-specific scenarios, written with the people who know the role, and a rubric that scores behavior.
In HR, conduct failures are legal exposure. Questions about health, family plans, religion or other protected characteristics, or inconsistent treatment of similar employees, can create discrimination claims regardless of whether the information given was accurate. New York City already requires bias audits for automated employment decision tools. An HR agent that drifts into evaluating people is no longer a help desk.
What to ask before relying on a conduct evaluator
Whose standard is it? A generic conduct rubric is a starting point. The evaluation has to encode your policies, your escalation paths and your jurisdictions. Ask how custom standards are defined and who approves them.
How is the evaluator evaluated? Automated conduct scoring is a judgment made by a model. It needs the same treatment as any LLM judge: calibration against human reviewers, measured disagreement, and a check that it does not reward the agent for being evasive. As AWS found with its own sales agents, an agent can be healthy by every metric and still wrong until the evaluation measures what the business cares about.
When does it run? Conduct drifts when the model, the prompt or the tools change. An evaluation run once before launch is a certificate; one run on every release is a control. Background agents, like Workday’s agents that nobody prompts, need it most, because no one is watching each interaction.
What happens to the evidence? A documented evidence record is valuable if it goes somewhere: to the agent’s owner, into the change approval for the next release, and to HR leadership as part of how the agent is supervised. It does not replace that supervision. A conduct score is evidence for a human governance decision, not the decision itself.
Cornerstone also cites its own research claiming that only 28% of AI initiatives fully deliver expected ROI and that 80% of failures trace to organizational readiness rather than technology. Those are vendor figures. The point they support is reasonable anyway: the hard part of putting an agent in a human role is defining the role, including how it is expected to behave.