The HR AI market is exploding. There are now 500+ vendors claiming to use AI in some form — from resume screening to workforce planning to benefits optimization. The problem? Most are repackaging basic automation or wrapping OpenAI APIs with a thin UI layer on top. The difference between a vendor that built genuine AI and one that calls ChatGPT behind the scenes matters enormously for reliability, data privacy, and long-term value.

Unlike buying an HRIS or ATS where features are visible and testable, AI capabilities are opaque by nature. You can’t look at a demo and tell whether the system is running a sophisticated custom model or a prompt to GPT-4. You need a different evaluation playbook — and that’s what this chapter provides.

“Our AI is 99% accurate.” → Accuracy on what dataset? Measured how? 99% accuracy on an easy benchmark means nothing. Ask for false negative rates by demographic group.

“It eliminates bias.” → No AI eliminates bias. Good AI can be audited for bias and mitigated. Any vendor claiming elimination is either naive or dishonest.

“It works out of the box.” → Means it hasn’t been trained on your data, your culture, or your definitions of success. Generic models produce generic results.

“We use proprietary AI.” → Could mean anything from a custom-trained model to a prompt template they wrote for ChatGPT. Ask specifically: is it a custom model, a fine-tuned model, or an API wrapper?

Every AI vendor demo uses curated datasets, best-case scenarios, and hand-picked examples. The resume the AI screens perfectly? They chose it. The chatbot answer that’s flawless? They tested that question 100 times. The sentiment analysis that nails it? Pre-selected data. This isn’t dishonest — it’s marketing. But it means what you see in a demo has almost no correlation with real-world performance.

Think of it like interviewing a candidate who only answers questions they prepared for. They sound brilliant. But can they handle unexpected situations? Edge cases? Messy real-world data? A demo tells you what the product can do at its best. You need to know what it does at its worst.

Model documentation: Does the vendor publish a model card? Can they explain what type of AI it is (custom ML, fine-tuned LLM, API wrapper)?

Bias audit results: Have they conducted a bias audit? By whom? What were the results by protected class?

Data practices: Where is your data stored? Is it used to train their models? Who has access? How is it encrypted?

Customer references: Can they provide 3+ references in your industry and company size?

Implementation timeline: Realistic timeline with milestones, not just “4-6 weeks.”

Support model: Dedicated CSM? Response time SLAs? Escalation path?

Who owns the data you put in? You should retain full ownership of all employee data, candidate data, and any content you create within the platform. This sounds obvious but must be explicit in the contract.

Can you export your data if you leave? In what format? CSV? API? Proprietary format that requires their tools to read? Data portability is your insurance policy.

Does the vendor use your data to train their models? Many AI vendors improve their models using customer data. If your employee data is training a model that serves your competitors, that’s a problem.

What happens after contract termination? Is your data deleted? Within what timeframe? Can you verify deletion? What about backups?

Per-seat: Pay per user or admin. Simple but can get expensive at scale. Watch for “user” definitions — is it admins only or all employees?

Per-transaction: Pay per resume screened, per chatbot conversation, per analysis run. Costs scale with usage, which can be unpredictable.

Per-model: Pay for specific AI capabilities. Want to add sentiment analysis? That’s another module, another fee.

Platform fee + usage: Base platform fee plus per-unit usage charges. Most common for AI-native tools. Gives predictable baseline with variable upside.

Pilot terms: Free or reduced-cost pilot period with defined success criteria. 30-90 days is standard. Push for no commitment to buy if the pilot fails.

Bias audit requirements: Require the vendor to conduct and share annual bias audits at their expense, with results broken down by protected class.

SLA commitments: Uptime guarantees, response time for support tickets, maximum latency for AI predictions. Tie SLAs to financial penalties.

Data deletion guarantees: Specific timelines and written certification of deletion upon termination.

Price protection: Cap annual price increases at 3-5%. Multi-year deals should lock rates.

Exit terms: Data export assistance, transition support period, reasonable notice requirements.

Defined success metrics: Before the pilot starts, agree on exactly what success looks like. Accuracy rate? Time saved? User satisfaction score? Adverse impact ratio? Write it down.

Time-bound: 30-90 days. Long enough to get real data, short enough to maintain urgency. Set milestone check-ins at 2 and 4 weeks.

Controlled comparison: Run the AI process alongside your current process. Compare outcomes directly. Don’t just measure the AI in isolation.

Representative data: Use real data that reflects your actual workforce and candidate pool. Don’t cherry-pick easy cases.