AI as a Co-Pilot: Where Human Expertise and Automation Intersect

Chapters

• Why this matters now
• Copilot was a good name back in 2024
• The work humans must own
• The work AI should take first
• How to build a genuinely human-AI HR
• How to measure value without sacrificing trust
• The leadership test ahead

The next era of HR service delivery will not be defined by who automates the most. It will be defined by who designs the clearest boundary between speed and judgement, between prediction and accountability, between assistance and authority. 

Why this matters now

Previously, I talked about the future of HR service delivery becoming an experience architecture discipline, not simply a more efficient support function. I also warned that hyper-personalisation can tip into surveillance if organisations collect more data than employees can reasonably understand, challenge or control.

This picks up exactly where they leave off: not with another general argument for AI, but with a more precise question: when AI can draft, detect, route, recommend and increasingly act, what must still remain unmistakably human?

Organisational AI adoption continued to rise sharply in 2025: Stanford’s 2026 AI Index reports that 88% of surveyed organisations were using AI and 70% were using generative AI in at least one business function, yet AI agent deployment remained in the single digits across nearly all business functions. At the same time, Microsoft’s 2025 Work Trend Index found that 82% of leaders saw 2025 as a pivotal year to rethink strategy and operations, and 81% expected agents to be moderately or extensively integrated into AI strategy within 12 to 18 months. In other words, the market has moved from experimentation to expectation, but not yet to mature operating discipline. (Stanford HAI)

That gap is where HR leaders now live. The hype says “AI agents everywhere”. The reality says something more awkward: most organisations are still figuring out which work should be automated, which should be augmented, and which should remain firmly human-led. The World Economic Forum’s Future of Jobs Report 2025 also makes the broader context plain: by 2030, employers expect 170 million new jobs to be created and 92 million displaced, with nearly 40% of skills set to change, while human capabilities such as creative thinking, resilience, flexibility and collaboration remain critical alongside technical skills. HR has a work redesign problem. (World Economic Forum)

The real dividing line in HR is less about AI adoption and more about judgement design. If we do not actively decide where human expertise must sit, the technology stack will make that decision for us: usually in favour of speed, scale and lower cost, and usually long before anyone has thought properly about fairness, trust or employee dignity.

Report: Decoding the HR Service Delivery Ecosystem
HR teams are being asked to do more than ever. Everest's Tech Provider Spotlight explores how HR Service Delivery is evolving. Read Now.

Copilot was a good name back in 2024

The phrase “copilot” sounds reassuring, but it is already out of date. Early generative AI mostly helped people write, summarise and search. The new wave is more agentic: these systems can retrieve from multiple sources, plan steps, propose actions, complete workflows and hand work to other systems. That is why the copilot metaphor still matters, but only if we update it. A modern AI is not just a helpful drafting assistant. It is a system that can increasingly influence the route, the timing, the next action and, if left ungoverned, the decision itself.

The problem is that AI capability is uneven. Harvard Business School’s “jagged technological frontier” research showed that people using GPT-4 completed 12.2% more tasks, finished them 25.1% faster, and produced higher-quality work on tasks within AI’s competence frontier. However, on a complex managerial task outside that frontier, AI users were 19% less likely to produce the correct answer. Stanford’s 2025 AI Index similarly noted that complex reasoning and planning remain significant weaknesses for current systems, especially as task horizons lengthen. That is the central lesson HR leaders should internalise: AI does not fail like a junior employee. It fails like a brilliant intern who sounds certain even when wrong. (HBS AI Institute)

That means the governing question for HR is not, Can AI do this task? It is, What kind of judgement does this task require? If the work is repetitive, rules-based, time-sensitive and emotionally low-stakes, automation is your friend. If the work is politically sensitive, legally consequential, emotionally charged, ambiguous, or likely to affect a person’s pay, progression, status or sense of fairness, the human should move from reviewer to owner. The wrong way to distribute work is by technical possibility. The right way is by judgement load. (HBS AI Institute)

A model for deciding what AI should do

This is not theory for theory’s sake. It is what the evidence points to. Generative AI has shown strong performance on professional writing and routine support work: a Science study found that ChatGPT reduced writing time by 40% and improved output quality by 18% among college-educated professionals, while an NBER field study found a 14% productivity increase in customer support, with gains rising to 34% for less experienced workers. The same body of evidence also shows that performance is much less dependable when tasks demand deeper, messier human judgement. (Science)

For HR, that translates into a very practical sequence of adoption. Start with first drafts, first-pass triage, knowledge retrieval, case summaries, standard policy explanations, workflow nudges, translation, data extraction and anomaly spotting. Move later — if at all — into recommendation-heavy domains such as policy interpretation, sensitive manager guidance, compensation exceptions, complex leave cases, performance concerns and employee relations. And be extremely cautious with any system that scores, ranks, profiles, labels or infers in ways employees cannot contest. (Science)

The work humans must own

In HR, the most important human contribution is legitimacy (not warmth).

Employees do not only want a fast answer. They want to know that the answer was fair, that the process was understandable, that context mattered, and that someone accountable could still listen, intervene and reverse course when needed. NIST’s AI Risk Management Framework was built precisely around that reality, emphasising governance, mapping, measurement and management of risk as design requirements rather than clean-up work after deployment. The UK ICO makes a related point in plainer language: where automated tools support decisions about workers, human involvement must be meaningful, not tokenistic, and reviewers must have the authority and competence to disagree with the system. (NIST)

This is where a lot of “human in the loop” language falls apart. Too often it means someone clicks Approve after the model has already framed the case, narrowed the options and set the default. That is theatre. The ICO’s guidance explicitly warns against this kind of pseudo-review, saying human decision-makers should actively interpret recommendations and weigh other information rather than simply applying them. Research on automation bias has been making a similar point for years: people over-rely on automated suggestions even when they know the system is imperfect. (ICO)

For HR leaders, this has a direct operating implication. The most sensitive HR decisions should never be delegated to AI merely because the output looks neat. The EU AI Act classifies many employment-related uses of AI as high-risk — including systems used in recruitment, promotion, termination, task allocation, worker monitoring and evaluation — and imposes obligations around risk management, data governance, transparency, human oversight, accuracy and security. The Act also prohibits AI systems used to infer emotions in the workplace except in narrowly defined safety or medical situations. In parallel, the U.S. EEOC has warned that employers’ use of AI can violate the Americans with Disabilities Act where tools screen out or disadvantage people with disabilities. (EU AI Act)

The privacy line matters just as much as the fairness line. The ICO warns that excessive worker monitoring can intrude into private life and undermine privacy and mental wellbeing, especially in hybrid and homeworking contexts. In 2024, it ordered Serco Leisure and associated trusts to stop using facial recognition and fingerprint scanning for staff attendance, finding the biometric processing unlawful and not shown to be necessary or proportionate. That case should be read by every CHRO tempted by “just one more signal” in the name of predictive HR. (ICO)

Meaningful oversight

If you cannot put a named human role against all five of these layers, you do not have human oversight:

HR should stop pretending that empathy can be automated. Tone can be simulated. Fluency can be simulated. Personalisation can be simulated. Empathy, in any meaningful sense, still depends on judgement, memory, moral responsibility and the ability to absorb consequence. AI can help humans show empathy more consistently by saving them time and surfacing context. It cannot carry the ethical burden in their place.

The work AI should take first

If all of that sounds like a warning against AI, it is not. It is the opposite. The most responsible way to use AI in HR is also the most commercially sensible: give it the work it is genuinely good at, and stop forcing humans to spend their best hours on administrative drift.

The evidence here is stronger than many HR leaders realise. On routine writing tasks, AI meaningfully improves speed and quality. On large volumes of support interactions, it can standardise knowledge and lift the performance of less experienced staff. On enterprise service delivery, IBM says its internal AskHR agent now automates more than 80 HR tasks, achieved a 94% containment rate for common questions, contributed to a 40% reduction in HR operational costs over four years, and helped drive a 75% reduction in support tickets since 2016. Even allowing for the fact that those are self-reported figures, the direction of travel is unmistakable: AI is already best used as an industrial-grade remover of repetitive work. (Science)

The old model was: employee asks, HR receives, HR interprets, HR drafts, HR routes, HR closes. The better model is: AI handles the first mile, humans own the hard mile. Let AI retrieve policy, assemble history, produce the first draft, summarise the case, identify missing information, prepare the manager briefing, and suggest next steps. Then let the human decide whether the issue is routine enough to complete or important enough to slow down.

This is what “co-pilot” should mean in practice: not replacing HR judgement, but clearing space for it. (NBER)

There is also a talent logic here. Research keeps finding that AI often benefits less experienced workers disproportionately. In the NBER customer-support study, novices benefited far more than top performers. In the Science writing study, lower-skilled participants narrowed performance gaps. That matters for HR teams under pressure, because it suggests co-pilot systems can raise the floor without flattening the ceiling. Used well, AI can democratise access to better drafting, stronger knowledge retrieval and more consistent first-line service — provided experts remain available for exceptions, coaching and quality control. (NBER)

This is also why I would start with manager enablement before anything more glamorous. The highest leverage co-pilot in HR may not be an employee chatbot at all. It may be a manager-side assistant that helps leaders produce clearer performance notes, better team updates, more consistent policy explanations, sharper meeting summaries and faster escalation briefs. Senior HR leaders know this instinctively: a great deal of HR friction enters the system through hesitant, inconsistent or poorly documented management. AI can reduce that friction dramatically. The trick is to stop before it becomes the author of the judgement instead of the author of the draft.

How to build a genuinely human-AI HR

The organisations getting this right are not the ones with the most pilots. They are the ones redesigning work, governance and accountability together.

IBM’s AskHR case shows what mature tiered automation can look like in HR service delivery, with routine interactions handled digitally and human advisers reserved for more complex needs. Moderna is interesting for a different reason: it merged HR and digital leadership responsibilities under a single executive and, according to reporting, built thousands of internal GPTs to redesign which work should be done by people and which by automation. And Klarna, in customer service rather than HR, offers an equally useful warning: after heavily emphasising AI-driven efficiency, its chief executive admitted that focusing too much on cost had led to “lower quality”, and the company moved back towards more human support. These are three very different stories, but together they expose the same lesson — AI succeeds when it is embedded in service design; it fails when it is treated as a headcount strategy disguised as innovation. (IBM AskHR)

I’ve translated that into six operating moves.

1. Start with a messy, memorable journey

Do not begin with an enterprise-wide promise of transformation. Start with a journey employees remember vividly because it combines operational friction with emotional consequence: parental leave, internal moves, complex onboarding, manager help, payroll exceptions, case escalations, or return from leave. Those journeys generate the richest signal about where AI can remove effort without removing care. That logic is consistent with the wider series’ employee-first approach, which repeatedly centres moments that matter rather than internal process charts.

2. Set autonomy bands before you scale

Define what AI may answer, what it may draft, what it may complete, what it may recommend, and what it may never decide. This sounds obvious, yet most organisations still blur the line between explanatory automation and decision automation. The law is moving faster than many realise: EU employment AI is high-risk, sole automated decisions that significantly affect people are tightly restricted under EU data protection rules, and UK guidance requires meaningful human involvement if automation is used in worker-related decision-making. (EU AI Act)

3. Redesign the knowledge layer for action, not browsing

Most HR knowledge is still written as content. Co-pilot systems need it written as decision logic: ownership, jurisdiction, exceptions, triggers, escalation points and plain language. If your policy documentation cannot tell an AI when to stop and hand over, it is no more than brochure-ready.

4. Train HR and managers as reviewers, not passive recipients

The new skill is “critical review”. People need to know when the model is likely to help, when it is likely to hallucinate, how to spot overconfident nonsense, and when to override the recommendation. The WEF’s 2025 jobs report is clear that technical skills are rising, but so are human skills such as collaboration, resilience, flexibility and judgement. The winning HR teams will develop both together. (World Economic Forum)

5. Create an appeal route that is real, fast and stigma-free

Employees should be able to say, “I want a human to look at this,” without being penalised by slower service, managerial irritation or silent downgrading. If you do not offer that escape hatch, the system will look efficient right up until the moment it creates a trust problem.

6. Treat AI introduction as service redesign, not software rollout

This is the hardest mindset shift of all. Do not ask, “Where can we deploy AI?” Ask, “Where do people waste the most judgement on low-judgement work?” That is where the co-pilot belongs first.

How to measure value without sacrificing trust

One reason HR AI programmes go wrong is that they get measured like call-centre automation. Deflection. Containment. Faster closure. Lower cost per interaction. Useful metrics, yes but incomplete to the point of danger.

Stanford’s 2025 AI Index found that even where companies reported financial benefits from AI, most described them as modest and early-stage. That should not depress HR leaders; it should make them more disciplined. The first wave of value is usually real but uneven. The firms that create durable advantage are the ones that pair efficiency metrics with trust metrics from the start. (Stanford HAI (2025)

 Measuring the right things: Two-Ledger Scorecard 

Ledger one: efficiency: track draft time saved, average time to first meaningful action, percentage of routine contacts resolved without hand-off, knowledge findability, repeat-contact reduction, manager admin hours removed, and capacity released for higher-value work.

Ledger two: trust: track factual accuracy, override rate, appeal rate, fairness audit pass rate, percentage of high-stakes cases escalated to humans, employee confidence after AI-supported interactions, and quality of outcome not just speed of closure.

The rule is simple: if ledger one improves while ledger two worsens, the programme is not succeeding. This is especially important in HR because the visible cost saving often sits with the function, while the hidden cost lands with employees and line managers in the form of rework, frustration, anxiety and weaker trust.

There is also a practical ROI case that senior leaders can understand quickly. If AI reduces manager communication drafting time by 40% — the rate observed in the Science study on professional writing tasks — then an organisation producing 120,000 people-related manager messages a year that previously took 20 minutes each would recover roughly 16,000 hours annually. At a fully loaded cost of £50 an hour, that is about £800,000 of managerial time returned before counting faster case resolution, fewer escalations or better consistency. And that is a conservative use case because it sits in the safe, draft-heavy zone rather than the high-risk decision zone. (Science)

The more interesting ROI, though, is strategic rather than clerical. When AI removes first-pass admin, expert humans can spend more time on exception handling, judgement coaching, sensitive conversations, policy redesign and manager capability. In other words, a co-pilot system should not only make HR cheaper. It should make HR better at being human where it counts.

Cultivating an Employee-First Mindset
Learn how to transform HR into a people-first function that builds trust, designs better experiences, and drives real business results in this interactive, 10-minute guide. Read Now.

The leadership test ahead

There is a lazy version of this future in which HR becomes the administrator of increasingly opaque systems: faster answers, fewer tickets, more dashboards, more monitoring, more “insight”, and steadily less dignity in the experience. That future is entirely plausible. It is also, in my view, unacceptable.

There is a better version. In that version, HR uses AI to absorb repetition, shrink delay, improve first drafts, surface patterns and raise the quality floor, while keeping fairness, empathy, context and accountability in human hands. That model is much harder than simple automation because it requires real design choices: autonomy bands, governance, escalation rules, appeals, audit trails, manager training, and honest conversations about which work is fundamentally relational. But it is also the model most aligned with where the labour market is heading: widespread AI use, uneven agent maturity, rising demand for human skills, and much stricter expectations around responsible deployment. (Stanford HAI)

So here is the thought I would leave with any CHRO or Head of HR Service Delivery: stop asking where AI can replace people, and start asking where it can return people to the work only they should do. The future of HR service delivery will not be won by the organisation with the most agents. It will be won by the one that can industrialise speed without industrialising indifference.

That is what AI as a co-pilot should mean in HR. Not human replacement. Not human theatre. Human judgement, finally given the capacity to show up where it matters most.

How Applaud Helps You Make It Happen

At Applaud, we believe employees are a company’s most important customers. That’s why our technology is built entirely from the employee’s point of view—delivering more human, intuitive, and rewarding HR experiences that empower HR teams to do more for their people.

If you’re ready to turn employee-first HR from vision to reality, we’re here to help. Get in touch to see how Applaud can transform your HR Service Delivery and create a workplace where employees truly thrive.