AI-native vs AI-painted: the difference that matters in HR Tech

I bought 14 HR Tech products in the last 18 months to evaluate the competition. All of them said “AI-powered” on their site. Half were trivial wrappers around GPT-4: one call to OpenAI with a fixed prompt, no structure, no auditability, no real differentiation beyond the logo.

I call that AI-painted: applying a layer of AI on top of a product that wasn’t internally reconfigured to leverage it. The “AI” label works for the sales deck and not much more.

AI-native is the opposite: a product thought out, modeled, and built with AI as a structural piece, not decoration. This post explains the difference with concrete criteria, and shows what it looks like in Talen.to.

---

Four criteria to tell AI-native from AI-painted

These are the criteria you’d apply if you were CTO or Head of People buying HR Tech in 2026. None of them is answered by “yes, we have AI”:

Criterion 1: What does your AI specifically do?

AI-painted: “intelligent candidate summaries”, “automatic suggestions”, “AI-powered matching”. Vague, not falsifiable.

AI-native: a concrete list of tasks the AI performs, with defined inputs and outputs. Examples:

• Generate 3 role description suggestions from a 2-line brief.
• Infer 12 technical competencies with required levels from a job posting.
• Suggest a value set of 5-7 prioritized values based on industry + size + declared culture.
• Draft a 200-word narrative per candidate post-assessment, with configurable tone.
• Produce a 400-word team summary aggregating individual profiles.

If your vendor can’t list 5+ tasks at that level of specificity, it’s AI-painted.

Criterion 2: Are the prompts inspectable and editable?

AI-painted: the prompts live in the vendor’s code. They are their “secret sauce.” You don’t see them and you can’t change them.

AI-native: prompts are client configuration, not vendor configuration. Each company can inspect, edit, and version the prompts that generate output about their candidates.

Why does it matter? Three reasons:

1. Compliance. If your Head of Diversity needs to ensure the AI doesn’t use biased language when describing candidates, the prompt has to be inspectable.
2. Tone and brand voice. The vendor’s default doesn’t necessarily match how your company speaks.
3. Fast iteration. If a prompt produces mediocre output, you can adjust it without waiting 6 months for the vendor’s next release.

Criterion 3: Is there versioning and audit?

AI-painted: the model changes when the vendor updates. You don’t find out. Today’s outputs aren’t comparable to those from 6 months ago.

AI-native: prompt and model changes are versioned, with logs of which version generated which output. This is mandatory to comply with emerging regulation (EU AI Act, Art. 12 — record-keeping in high-risk systems).

I go deeper on regulation in Algorithmic transparency in HR Tech.

Criterion 4: Is the AI integrated into the scoring engine or just the cosmetic output?

AI-painted: the AI only writes the final report. Scoring is still deterministic and the AI just “narrates it more nicely.”

AI-native: the AI participates in specific parts of the reasoning (not in all — hard scoring should remain deterministic to be auditable). Legitimate examples of AI use inside the engine:

• Inferring soft competencies from OCEAN + Values patterns (not from the candidate’s direct self-report, which is biased by social desirability).
• Generating follow-up interview questions based on specific gaps.
• Synthesizing multi-evaluator assessments (with disagreement detection).

---

What this looks like in Talen.to

I’ll be concrete, not aspirational. This is what’s in production today:

AI suggestions to create roles

In /roles/ai-suggestion: describe the role in natural language (“Senior Backend Engineer for fintech, team of 8, Go + PostgreSQL stack, high-autonomy culture”), and the platform proposes:

• Technical competencies with required levels.
• Ideal OCEAN profile (ranges per dimension).
• Preferred archetypes.
• Priority values.

You edit before publishing the role. The AI doesn’t decide — it suggests a reasonable starting point.

AI suggestions for value sets

In /custom-value-sets/ai-suggestions: if your company doesn’t have formalized values, the AI proposes a set of 5-7 values with operational definitions based on your industry + size + declared culture. You edit before applying. Values are versioned (see Why generic assessments fail at culture fit).

AI report generation per candidate and per team

Each assessment generates a ~200-word narrative that synthesizes the candidate’s profile against the role. Re-generable. The prompt lives in /companies/[id]/settings/psychometric/prompts — editable and versioned per company.

At the team level, the platform generates team narratives (~400 words) that synthesize team composition: dominant archetypes, collective gaps, distinctive strengths. Useful for management conversations and hiring planning.

Inferred soft competencies (not self-reported)

This is the piece I like the most and the one that makes the biggest difference. Soft competencies (communication, adaptability, leadership) have a historical problem: if you measure them with a direct test, people lie. Not out of bad faith — out of social desirability bias: we all think we’re good communicators.

The Talen.to solution: soft competencies are inferred from the OCEAN + Values profile using a validated competency_ocean_mappings table. The candidate doesn’t answer “how adaptable are you?” — adaptability is inferred from their pattern in Openness + Emotional Stability + SR + certain values.

This isn’t “AI” in the marketing sense (it’s not an LLM). It’s applied psychometric modeling. But it’s exactly the kind of intelligence integrated into the product that distinguishes AI-native from a wrapper.

Editable and versioned prompts

Every prompt that generates output about your candidates lives in your company’s settings, not in our code. If you want the narrative to never mention age, gender, origin, or marital status: you add it to the prompt and it’s versioned. If you want to change the tone to more formal or more casual: you configure it.

What we do NOT use AI to do

Equally important. There are decisions where generative AI should not intervene:

• Computing the overallFit (it’s deterministic, formula visible, auditable).
• Assigning archetypes (it’s Euclidean distance against empirical centroids).
• Per-dimension weights (you configure them).

If your vendor uses generative AI to compute the candidate’s final score, run. That’s not auditable, doesn’t meet regulation, and reproduces training corpus biases invisibly.

---

The “AI-first” trap

A final note on the discourse. Many vendors call themselves “AI-first” to suggest AI is at the core of the product. That sounds virtuous, but often it’s the opposite: it means the product was built around what an LLM can do, not around the customer’s problem.

A solid hiring system is science-first and AI-augmented, not AI-first. Science (psychometrics, empirical validation, calibration with real data) provides the deterministic, auditable core. AI adds layers where it provides real value (synthesis, inference, text generation, suggestions) without replacing the engine.

That’s the difference between AI-native (AI integrated with judgment) and AI-painted (AI glued on with marketing).

---

Practical checklist for evaluating AI in HR Tech

If you’re buying, copy and send:

• Can you list 5+ concrete tasks your AI performs, with defined inputs and outputs?
• Are the prompts inspectable and editable by the client?
• Is there prompt and model versioning with output logging?
• Does the AI participate in fit scoring or only in text generation?
• How do you handle bias in generative output (language, tone, omissions)?
• Can you deliver the technical documentation required by EU AI Act Art. 11?

Three or more “no” → AI-painted. Worth exploring another option.

---

Closing

“We have AI” means nothing in 2026. Everyone has AI. The question is what it does, where it’s audited, whether it can be modified, and whether it’s integrated into the product or stuck on top.

If you want to see AI-native in practice, create an account at app.talen.to/sign-up and open settings/psychometric/prompts. It’s all there, with nothing hidden.

If you want the full panorama of capabilities: What Talen.to actually does. If you want to see your specific case: book a demo.

— Fran