Your First Agent Shouldn't Be Your Most Impressive One

There's a reason the ambitious ideas come first. They make the strongest business case and the best demo. But a first agent has a different job from every agent after it. It's how your organisation learns to switch AI on safely: who sets up access, who checks the answers, who keeps the information behind it current, and what happens when something goes wrong.

If the first agent is also your most complex and highest-risk one, you're learning all of that at once, on the use case where mistakes matter most.

A first agent has a different job from every agent after it.

A good first agent usually has four things in common:

• It reads rather than changes data. If the answer is wrong, the agent has not directly changed the underlying HCM record, although users could still act on incorrect information.
• It answers questions people ask often. That gives you enough use to learn from, and a visible benefit quickly.
• Its answers are easy to check. The person asking would usually know if the answer looked wrong.
• A wrong answer is an inconvenience, not a harm. Nobody's pay, job or prospects depend on it.

None of this is about the AI being less capable. It's about choosing a use where you can see clearly how it behaves before you rely on it for more.

Before anyone designs a custom agent, check what Oracle already provides. Oracle's quarterly updates have introduced a growing range of preconfigured HCM agent templates since 25C, many of them exactly the kind of everyday question-answering assistants that make good first agents.

Two examples:

• The Employment Information agent (25D) gives a plain-language summary of one person's job, grade, position, department, location, contract and work relationship. For changes such as a promotion or transfer, it links to the right page with the person already loaded, but the change itself is still made in the normal page.
• Leave and Absence Analyst (My Team) (25C) answers managers' questions about their own team's leave and absence.

Delivered agents still need setup. Oracle's notes for the Employment Information agent, for example, list platform services, a profile option, permission groups and role access in AI Agent Studio. Check what your licences include and what each agent needs before you plan the work. Preconfigured does not mean universally available without enablement. Delivered agents may still require the relevant platform services, profile options, permission groups and user access before they can be used.

Some delivered agents can change records as well as answer questions. That isn't a reason to avoid them, but it changes the order you do things in.

Person Data Assistant (26C) can retrieve a range of personal information and can update supported data areas such as names, national identifiers and ethnicity. Other areas, including passports and visas, are currently available for retrieval rather than direct update. For an initial pilot, you could limit the scenarios tested and communicated to information retrieval, then introduce supported update scenarios once RBAC, effective-date behaviour, pending transactions and approvals have been validated. Whether this can be technically restricted to read-only should be confirmed for the specific template and roles being deployed.

Learning Assignment Assistant is a useful model for how an agent that acts should behave. From 26C it can update many learning records in one go, but Oracle describes it producing a summary of the proposed changes first, and learning specialists "can review and approve the proposed changes before the assistant does the updates." For a first agent that changes records, look for a meaningful control before the change takes effect, such as a presummary, user confirmation, normal transaction approval or a deep link that leaves the final action in the application. If the change is difficult to verify or reverse, it is better left until later.

Recruiting is often the first area people want to automate, because the volumes are high and the pain is real. It's also where the stakes for individuals are highest.

Oracle's Job Applicant Screening Agent Template brings together what's already on an application, and from 26A the Job Applicant Screening Advisor Enhancement: Candidate Authenticity adds candidate authenticity confidence scores. Oracle describes the score as helping recruiters identify potentially fraudulent or AI-generated candidates. My view is that it should be treated as a signal for further review, not as an unchecked reason to reject someone.

Regulation points the same way. The EU AI Act identifies AI used for recruitment or selection, including analysing and filtering applications and evaluating candidates, as an Annex III high-risk use case, subject to the Act's classification rules and limited exceptions. The assessment must consider how the agent is actually used, not simply that it sits within Recruiting. In the UK, the ICO explains that Article 22 places additional restrictions on solely automated decisions that have legal or similarly significant effects. Where it applies, organisations need safeguards including a route to human intervention and challenge. Even where a recruiter remains involved and Article 22 does not apply, the wider UK GDPR requirements around lawful processing, transparency, accuracy and fairness still remain relevant. This isn't legal advice, and your own legal and data protection teams should review any recruiting use. But it's a clear sign that screening deserves your most mature processes, not your first ones.

Once you've chosen an agent, give it to a small group first: one HR team, one group of managers, one country. Tell them they're using AI and what it's for, and give them an easy way to report a wrong or unhelpful answer.

Then measure something that matters to the business, not just the number of conversations. For example:

• questions answered without raising an HR ticket
• answers reported as wrong, and why
• how often the policies or knowledge behind it needed correcting

The last one is often the most revealing. An agent's answers depend on the underlying HCM data, configured tools, prompts and any policy or knowledge content it uses. A pilot quickly exposes weaknesses in both the technology and the ownership of those sources.

Oracle updates Fusion every quarter. Each update reaches test environments on the first Friday of the update month and production on the third Friday, which gives you roughly two weeks to check your agents before your users see the change.

Create a short regression pack containing representative questions, expected facts or outcomes, prohibited responses and role-based access checks. Run it in test after each quarterly update and record who signs off the result. From 26D, Oracle AI Agent Studio ATLAS also supports repeatable, source-controlled regression testing for workflows and agentic applications, as a more mature future option.

If most of your answers sit in the middle column, it's a strong first agent. If either "read or change" or "what happens if it's wrong" lands in the right-hand column, keep it for later, however good the business case looks.

The best first agent is rarely the one that wins the room. It's the one that answers an everyday question, only reads data, and gives you a safe place to learn how AI behaves in your own environment.

Get that right and the ambitious agents become much easier: you'll already know who owns the knowledge, how to test after each update, and how your people respond to working with AI.

Oracle-specific points in this article are based on Oracle's published readiness material for 25C to 26D. Features vary by environment, licences and configuration, so always validate in your own environments.