What AI in HR actually looks like - modeled end-to-end on a scaling tech company.

They shouldn't trust it as a final product. They should treat it as a first draft built from evidence they would otherwise have to reassemble themselves - 1:1 notes, goal attainment, peer feedback, prior ratings. The manager is the final author. What AI removes is the blank-page problem and the reconciliation tax.

In the demo, the bias scan surfaced two weaknesses in its own draft - recency bias and a vague growth area. That's the model's job: make itself challengeable.

Three mechanisms designed into the architecture, in order:

1. Evidence-anchored drafts. The AI is only allowed to cite from real evidence the manager already has: goals, 1:1 notes, ratings, peer entries. No vibes, no inferred personality. If there's no evidence, it says so. AICAP HR specifies what the AI is allowed to read and how it must cite; the engineering layer enforces those rules.

2. Automated bias check. Every output runs through a second pass that flags recency bias, gendered language, vague growth areas, over-weighted single events, and any pattern where ratings systematically favour one group over another. AICAP HR designs what to check for; the build team or vendor implements the checks.

3. Calibration as cross-check. The calibration pre-read (Chapter 4) surfaces rating drift across managers - so individual bias can't hide in aggregate. AICAP HR designs the calibration methodology and the committee playbook; managers run the conversation.

None of this makes bias impossible. It makes it visible, which is the condition for fixing it.

The architecture AICAP HR designs keeps employee data inside the company. The AI runs in an environment the company controls (a private setup, a vendor's enterprise-tier offering, or a company-managed instance), HR data stays in the HRIS, and no employee records get sent to public AI services. AICAP HR writes the data-handling policy and the architecture spec; the client's engineering team or a partner systems integrator builds to it.

Phase 2 (once governance is defined - month 3 onward) allows for opt-in metadata-only signals from calendar and collaboration tools, with employee-level opt-out always available. Full audit log of every AI decision from day one - AICAP HR specifies the log schema and review cadence; the engineering layer implements it.

A skills ontology is a structured list of the skills the company needs to run - something like "P&L ownership at $50M scale," "incident-command decision-making," "B2B contract negotiation." A good ontology lets you ask real questions: who in the company has the skills the open VP role requires, including people outside Operations?

Most companies in this size range don't have one yet. The current demo uses performance + role-family proximity as a proxy - which is why the slate only surfaces Ops candidates, not cross-functional talent. Building the ontology is Phase 2 (month 6). Options: Lightcast (off-the-shelf), SFIA (global standard), homegrown, or AI-inferred from work artifacts. AICAP HR's default recommendation is AI-inferred seeded by self-declaration, audited quarterly.

Month 0-3: Governance design, audit-log spec, HRIS rationalisation (off spreadsheets into a real platform if not already there). Pilot the review drafter with one business unit (Ops, 40 people).

Month 3-6: Expand review drafter to all 263 active employees. Launch calibration pre-read for FY26 cycle. Begin skills ontology selection.

Month 6-9: Skills-per-employee seeded. True skills-based successor slating goes live. Begin flight-risk early-warning system.

Month 9-12: Full portfolio (13 processes from the automation matrix) in production. People-data warehouse joining HR + business outcomes.

AICAP HR scopes the phasing; the engineering layer builds to it through one of the three delivery paths (client engineering, partner SI, or vendor stack).

Vendor-agnostic. AICAP HR helps clients evaluate and select the right tools based on scale, data, regulatory posture, existing stack, and team capacity. No defaults, no kickbacks.

The categories typically in play: an HRIS that supports clean integrations and role-based permissions; a comp platform with live market data; a performance and engagement tool; an AI layer (Claude, ChatGPT, Gemini, or a company-managed alternative) deployed in a way that keeps employee data inside the company; and a connective layer that sits across all of them and keeps an audit log of every AI action.

The vendor mix varies by client. AICAP HR's job is to help you select what fits, design how the pieces talk to each other, and oversee the build - not to push you toward any particular vendor.

~60% reduction in manager + HR hours per review cycle. Every high flight-risk employee (48 today) gets a named owner and a playbook - historically zero of them did. A single retention save at the director level pays for the full tool stack for a year. And the compounding: HR's time redirects from drafting to strategy, which is what the company needs as it scales from 263 to 800 people without 3×-ing HR headcount.

Three failure modes I watch for:

1. Rubber-stamping. Managers sign the AI draft without editing. Counter: require manager edits as a non-skippable field; audit-log diff of what changed.

2. Data staleness. The model reasons over a snapshot that's two weeks old and misses a triggering event. Counter: near-real-time sync for the highest-signal sources (engagement, 1:1s, goal updates).

3. Trust loss on a high-profile error. One visibly biased or wrong AI output becomes the story. Counter: start with internal-only pilots, heavy manager training, and make the bias scan default-visible - not a hidden feature.

Yes. Existing tools add AI features inside their own product: AI-drafted review inside one tool, AI summary inside another. That is useful but narrow.

What AICAP HR designs is the connective tissue across all your HR tools. So a performance rating in one system can trigger a calibration review in another, which flags a retention risk in succession, which opens a comp review. No single vendor does that end-to-end. AICAP HR designs how the tools talk to each other, sets the access rules, and oversees the build by the client's team or a partner. That is the gap a scaling company can own by engaging AICAP HR before they're locked into a single vendor's assumptions.

HR owns the policy, the principles, and the audit. Engineering (the client's team, a partner integrator, or a vendor that provides this out of the box) builds the connective tissue across the tools. CEO and board own the governance question: what is AI allowed to recommend versus decide? That line gets set once, in writing, and every tool lives inside it. AICAP HR is the architect across all three: writes the spec, runs design reviews, oversees the build, audits the output.

I've sketched an org + ownership model in the blueprint deck. Happy to walk through it.

This is the single biggest implementation risk and it is almost always under-designed. Right answer: build permission rules in from day one, so the AI can only ever see the data the user in front of it is allowed to see. Not bolted on later.

The access rules AICAP HR designs, by role:

HR sees everything. Managers see their direct reports plus the people on their succession slate (with rationale, not raw peer comparisons). Employees see their own data, their team's org structure, and their own skills - never peer ratings, peer flight-risk, or peer comp. Executives see aggregated views plus their own personal successor slate. Boards see only aggregated trends with names removed.

AICAP HR designs the access rules and writes the technical pattern that enforces them. The build is delivered by the client's engineering team, a partner systems integrator, or a vendor that already supports this. The choice is made with the client based on what fits. AICAP HR runs design reviews, signs off on the implementation, and audits it against the spec. This is where most consultants gloss; AICAP HR designs it on day one because once you lose employee trust, you can't get it back.

All four are either live or imminent and all four apply to HR.

NYC Local Law 144 (live since July 2023): mandatory annual bias audits on any Automated Employment Decision Tool used for NYC-based hiring, with public disclosure. Penalties from $500 per violation up to $1,500 per day for ongoing violations.

Colorado SB24-205 (live Feb 1, 2026): consumer protection law treating any AI that's a substantial factor in employment decisions as "high risk." Deployers carry duties around impact assessment, risk management, and disclosure.

EU AI Act (high-risk enforcement begins August 2, 2026): recruitment, performance evaluation, promotion, and termination AI all classified as high-risk. Compliance requires risk management systems, data governance, transparency, human oversight, and accuracy / robustness controls.

Canada AIDA (Artificial Intelligence and Data Act, in legislative review): broadly similar high-risk framework expected once it lands. Pre-IPO Canadian companies should design to it now.

AICAP HR designs the architecture to include the audit log, bias scan, human-decision retention, and disclosure language needed for all four. The compliance work isn't a Phase 2 retrofit; it's specified into the architecture from Month 1 and built by the client's engineering team or the chosen vendor stack under AICAP HR oversight.

The data backs this up: 83% of GenAI pilots in 2026 fail to reach production, and the failure is change management, not the model. Only 1 in 4 HR professionals played a leading role in their own company's AI rollout. That's the gap AICAP HR closes. Running global People functions at scale taught me where adoption actually breaks, and how to design around it from day one.

Three operating principles AICAP HR brings to every engagement:

1. HR leads, doesn't follow. The CHRO is the executive sponsor from week 1. Standard AICAP HR practice is not to take engagements where HR is brought in after the tech is bought - the design and adoption work has to happen up front.

2. Manager pilots before scale rollout. Eight to twelve volunteer managers run the new workflow for one cycle. Their feedback rewrites the prompts and the policies. Only then does the system go org-wide. AICAP HR designs the pilot, recruits the cohort, runs the feedback loops; the engineering layer ships the iterations.

3. Visible wins inside 60 days. One or two outputs (the review drafter, the calibration pre-read) ship early and visibly. Managers see hours back on their calendar. That earns the right to do the harder, slower work (skills inference, succession).

Adoption is engineered into the engagement plan from week 1, not bolted on after the tech is live.

Yes, they are. 70% of knowledge workers are using GenAI outside their company's official policy in 2026. Managers are pasting employee names, ratings, and feedback into ChatGPT and Claude right now to draft reviews, prep for tough conversations, or summarise 1:1 notes. The data has already left.

This is the shadow AI problem. The CHRO's job isn't to stop it - that boat has sailed. It's to bring the usage under governance before an incident forces the conversation.

AICAP HR's standard 30-day intake includes a shadow AI inventory: an anonymous survey of what employees are actually using, for what tasks, with what data. AICAP HR designs the survey, analyses the results, and turns the inventory into the policy baseline. From there, the company replaces unofficial AI use with an approved set of tools that the company controls and can audit. AICAP HR designs that set; the engineering team or the chosen vendors build it. People still get the productivity gain; the data stays where it belongs.