How a Graduate Student Built a Live HR AI Agent for Under $35

Most guides on how to build an internal AI agent for HR skip the hard part — not the technology, but getting a real organization to approve, trust and adopt it. Laís, a Master's in Artificial Intelligence student at Atlantis University, went through the full process during her internship: executive pitch, prototype, approvals, testing, deployment. The agent has been live since March, serving 130 employees in two languages. Initial cost: a $31.50 developer account.

Start with the problem, not the technology

The company is a manufacturing business with over 130 employees, scaling quickly. Its entire human resources function consisted of two people — and one of them, the VP, was about to leave.

Meanwhile, new hires receive an employee handbook that ran to 75 pages, with variations by state because different jurisdictions require different content. Employees could theoretically find their own answers. In practice, navigating 75 pages of policy documentation across multiple menus is enough friction that most people just email HR.

So HR spent its days answering the same recurring questions — how many PTO days do I have, what's the policy on X — instead of doing strategic work.

Note what came first. She didn't start with "let's deploy AI." She started with a bottleneck that had a measurable cost and a human consequence.

Design decision one — work with the infrastructure you already have

The organization runs on Microsoft: Teams, SharePoint, the full stack.

Rather than pitching new software, she built inside that environment using Microsoft 365 Copilot, described as an entry-level agent builder. It won't create agentic systems that route requests across multiple specialized agents — but it builds a single, capable agent well.

Her reasoning is worth borrowing:

"Instead of asking and pitching to get additional software, I said — let me work with what we have. That's probably going to help me sell the return on investment later."

Using existing licenses can simplify procurement, but the project still needs the organization's security, privacy, access and budget review.

Design decision two — deliberately limit what the agent knows

Counterintuitive, and the most important choice in the build.

HR questions can have legal consequences. The team therefore restricted the agent to an approved knowledge base and defined when it should decline an answer or escalate to HR.

"If you have something that is sensitive, you want to limit what knowledge this agent has access to, so that it does not create hallucinations."

Her framing of the broader hallucination debate is the clearest articulation of it from the whole Summit:

"Everybody's saying AI is not that reliable, it hallucinates. But these are all controls that you can put into place."

Grounding answers in approved sources and constraining scope can reduce hallucinations, but cannot guarantee their elimination. Testing, monitoring and escalation remain part of the design.

Design decision three — build for the actual audience

This is where the project moves from competent to genuinely good.

A manufacturing facility employs a substantial number of workers whose primary language is Spanish. So the agent serves the company in both English and Spanish from day one.

The project also explored a SharePoint intranet interface for employees without Teams access. That route still requires the appropriate identity, permissions and licensing for the organization's deployment.

"Who is going to interact with this? Is this really going to be useful?"

Plenty of internal AI tools fail not because the technology underperforms but because nobody asked who would actually use them.

The executive pitch

She describes it as "a shark tank situation." What she brought:

• Structured implementation plan — not "let me build you a bot"
• Cost of implementation — the $31.50 developer account, plus modest usage overage described as cents on the dollar
• Identified risks — liability, hallucination, scope
• Quantified benefits — versus an HR employee spending 15 to 30 minutes per query
• Testing metrics — accuracy rates across response categories
• Scalability path — start here, expand there, at incrementally higher cost with proportionally higher output

The element that closed it: Microsoft's built-in analytics dashboard. Executives can see how many minutes the agent has worked, what it cost, and — by inputting the cost of equivalent human handling — exactly how much the company saved.

"That was a big call for the executive team."

If you're pitching an internal AI project, note the sequence. She didn't sell the technology. She sold a measurable business case with the measurement built in.

Testing — the step most internal projects skip

Testing wasn't an afterthought; it was a presented deliverable. Response accuracy was measured across multiple categories:

• Standard questions — does it answer correctly from the knowledge base?
• Sensitive questions — does it handle appropriately what carries liability?
• False negatives — questions it should answer but might decline
• Out-of-scope questions — what does it say when asked something entirely unrelated?

That last category is where poorly built agents embarrass their organizations.

Timeline: 12 weeks from start to a live agent with HR comfortable using it — roughly one semester. Approved and rolled out in March; in production and tracked ever since.

Response time sits under a minute, and each answer includes a source reference, so employees can read the underlying policy if they want to. Small design choice, significant trust effect.

Results and what came next

The HR VP did leave. A single person now manages the function, supported by the agent handling recurring questions.

And the foundation proved scalable. Additional agents now in development:

• Request routing — organizing incoming requests from other departments to the accounting team
• Sales quoting support — the company builds highly customizable engineered products, which makes learning to quote them genuinely difficult; a second agent uses that quoting knowledge base to accelerate onboarding

Next 60 to 90 days: metric reviews and capability expansion.

Frequently asked questions

What does it cost to build an internal AI agent?

In this case, a $31.50 developer account plus modest usage charges, because the organization already licensed Microsoft. Costs vary substantially with infrastructure and scope.

Do I need to be a developer?

No. Microsoft 365 Copilot's agent builder is designed for non-developers. She did use Claude to help structure project documentation and testing records.

How do you reduce an internal agent's hallucinations?

Ground answers in approved sources, limit scope, test failure cases and provide a route to HR. These controls reduce risk; they do not guarantee error-free responses.

How long does implementation take?

Twelve weeks in this case, from design through testing to a live deployment HR trusted.

What's the hardest part?

Not the build — the organizational work. Executive approval, defining liability boundaries, and getting the team to trust the system enough to let it answer on their behalf.

Can employees without Teams accounts use it?

A separate intranet interface may provide another access route. Employees still need the identity, permissions and licenses required by that deployment.

Conclusion

Figuring out how to build an internal AI agent for HR is far less about the technology than most coverage suggests. The build is accessible. What separates a deployed system from an abandoned prototype is everything around it: identifying a bottleneck with real cost, working within existing infrastructure, constraining scope for safety, designing for the actual user population, and presenting a business case with measurement built in.

A graduate student did all of that mid-program, alongside coursework — consulting her professors on tool selection and deployment approach as she went.

Which reframes the usual assumption about timing. You don't wait until after graduation for any of this to start mattering.

The Master's in Artificial Intelligence and the STEM MBA at Atlantis University are built around exactly this combination: large language models, human-computer interaction, AI ethics, and the practical judgment to deploy responsibly inside a real organization.

Become a student at Atlantis University