Three out of four Indian employees worry their company will use AI for workplace surveillance, according to Mercer’s Global Talent Trends 2026 report, a figure higher than the global average. That anxiety sits alongside a separate finding from the same survey: only 48% of Indian employees think their performance management systems are actually effective. Put those two numbers together and a harder question emerges. Companies are handing performance data to algorithms specifically to make manager decisions fairer. Whether that’s what’s actually happening is far less settled.
HR teams across India’s IT services firms, BFSI majors, and fast-growing startups have spent the past two years bringing AI into performance calibration, promotion shortlists, and attrition-risk flagging.
The pitch is consistent: humans are inconsistent, moody, and prone to playing favourites, so let the data decide. NASSCOM’s AI Adoption Index put India at 2.45 out of 4 in December 2025, with 87% of enterprises actively using AI in some business function. Performance and talent decisions are squarely inside that expansion.
The Case For AI Reducing Manager Bias
The argument for AI as a bias-correction tool starts with a real and well-documented problem: unmanaged human judgement is inconsistent in predictable, unflattering ways.
Recency bias means a manager remembers the disastrous client call from last week more vividly than a strong quarter from four months ago. Halo effects mean one impressive presentation can inflate an entire year’s rating. Similarity bias means managers tend to rate people who remind them of themselves more favourably, a pattern that has quietly shaped promotion pipelines at Indian companies for decades. None of this shows up as intentional discrimination. It shows up as noise that happens to fall along predictable lines: gender, tenure, which office location someone sits in, whether they went to the same college as their manager.
A well-designed AI layer can, in theory, correct for some of this. It doesn’t get tired at the end of a long calibration session. It applies the same rubric to every employee’s project data rather than remembering some more vividly than others. It can flag when one manager’s ratings are systematically lower than their peers’, a calibration problem HR has struggled to catch manually for years.
TCS’s internal instruction to managers to rate roughly 5% of staff in “Band D” during a recent appraisal cycle is a useful, if uncomfortable, illustration of the underlying problem AI is meant to solve. A forced distribution applied uniformly by policy removes some discretion from any single manager. An AI system doing the same job with richer, individual-level data could, in principle, do it with more nuance rather than less. That’s the optimistic case, and it isn’t a fringe one. It’s the reason HR budgets for these tools have grown every year since 2023.
How AI Actually Enters A Manager’s Decision Today
AI rarely makes the final call on a promotion or a rating. It sits upstream, shaping the inputs a manager sees before they decide.
In practice, this means an AI layer might flag which employees are “attrition risks” based on engagement survey responses, internal job applications, and manager sentiment in one-on-one notes. It might pre-populate a draft performance review using project completion data, peer feedback text, and prior rating history. It might rank candidates for an internal promotion panel using a scoring model trained on who got promoted in previous cycles.
Mercer’s India findings show why this matters at scale: 66% of HR leaders in the country are actively redesigning work to combine human and AI decision-making, and 78% of executives now see AI fluency as essential to how their organisation operates. Performance management, specifically, has become a stated 2026 priority theme in India, with HR leaders focusing on using performance data for talent and pay decisions while leaning on AI to sharpen feedback and performance conversations. The tools are only getting more central to how the decision gets made.
Why The Same Mechanism Can Work In Reverse
The counter-argument doesn’t dispute any of that. It points out that AI models learn from historical data, and historical data carries every bias the system was supposed to fix.
If an AI tool is trained on five years of past ratings, past promotions, and past manager feedback, it isn’t learning what good performance looks like. It’s learning what got rewarded in the past, including whatever biases shaped those past decisions. A model trained on a dataset where women were historically rated lower in “leadership potential” doesn’t need gender as an input field to reproduce that pattern. It only needs a correlated variable, what researchers call a proxy: which projects someone was staffed on, how many hours of visible face-time they logged, whether they took a long leave that shows up as a gap in an activity timeline.
That proxy risk isn’t hypothetical to the people advising companies on how to use these systems in the first place. At the Fortune Global Forum 2025, when asked how corporate clients were applying AI to succession planning and promotion decisions, one executive search leader put the core problem in blunt terms.
“Frankly, it saves so much. You get all your answers at your fingertips. It’s so sexy. But is it the be-all and end-all of your answers? No.” – Anne Lim O’Brien, Vice Chair, Global & CEO Practice, Heidrick & Struggles
The output looks objective because it’s a number generated by a model. Whether the number encodes last decade’s bias is a separate question entirely, and it’s the one managers are least equipped to ask.
India’s version of this proxy problem has its own texture. Tier-1 college pedigree, urban metro location, and English-fluency signals in written communication have functioned as unspoken filters in Indian corporate hiring and promotion for years. An AI model trained on outcomes shaped by those filters will treat them as predictive of performance, because in the historical data, they correlate with it. The bias doesn’t disappear. It gets a technical justification.
| Bias Source | How A Human Manager Introduces It | How An AI System Can Reproduce It |
| Gender | Assumes lower ambition post-maternity leave | Treats employment gaps as a negative signal without context |
| College Pedigree | Favours familiar institution names when assessing potential | Weights degree source as a performance predictor from historical correlation |
| Location | Assumes remote or Tier-2 city staff are less visible, less committed | Uses office attendance or activity logs as engagement proxies |
| Tenure With Manager | Rates familiar reports more generously (similarity bias) | Uses prior manager ratings as a training label, inheriting the same skew |
The uncomfortable part is that both columns produce the same outcome. What changes is how defensible it looks afterwards. A biased manager’s decision can be challenged and overturned. A biased model’s output arrives wrapped in the language of data, which makes it harder to question and easier for an organisation to treat as neutral.
Where The Evidence Actually Sits
Neither position is fully wrong, which is what makes this genuinely unresolved rather than a case of picking a side.
AI systems can outperform individual managers on consistency and can surface calibration problems that were previously invisible, like one manager’s ratings running consistently below their department average. That’s a real, measurable gain. At the same time, every documented case of algorithmic bias in employment decisions, from global recruitment tools to promotion-scoring systems, traces back to the same root cause: a model trained on data that already encoded discrimination, deployed without anyone checking whether it had.
For Indian HR teams, this isn’t a theoretical debate anymore. The DPDP Act, 2023 treats performance data as personal data subject to purpose limitation and data minimisation, and any organisation using automated scoring in employment decisions needs to be able to explain, on request, what data went into that score and why. That requirement alone forces a level of scrutiny that informal manager judgement never had to survive. It’s a backhanded benefit of the regulation: the paperwork AI now requires is pushing HR to ask questions about historical bias that nobody was asking before.
The honest position is conditional. AI reduces one category of bias (inconsistency, fatigue, memory-driven recency effects) while creating exposure to another (encoded historical discrimination, laundered through a model that looks neutral). Which effect dominates depends entirely on what data trained the model, how often it gets audited, and whether a human is actually empowered to override it in practice.
What HR Can Actually Check For
Auditing an AI performance or promotion tool doesn’t require a data science degree. It requires asking pointed questions before rollout and revisiting them on a fixed schedule that continues well past launch. Most of these questions can be asked directly of the vendor, or the internal team that built the tool, and the absence of a clear answer is itself useful information.
- Ask the vendor which historical dataset trained the model, and how far back it goes. A model trained on ten years of ratings inherits ten years of whatever bias shaped those ratings.
- Run the tool’s outputs against demographic splits, gender, tenure, work location, before trusting it in a live calibration cycle. Uneven clustering in ratings or flags is a signal worth chasing.
- Check whether managers can see why the model produced a given score or flag rather than only the score itself. A black-box number is not an audit trail.
- Confirm a human can actually override an AI-generated flag, and that doing so is tracked rather than discouraged informally.
- Revisit the audit on a fixed annual schedule. Models drift as the workforce and the business context change, and last year’s clean audit says nothing about this year’s data.
A related discipline already exists on the hiring side of HR, where algorithmic bias in recruitment tools has forced companies to build exactly this kind of scrutiny into vendor selection. Performance and promotion decisions deserve the same rigour, arguably more, since the stakes for the employee are just as high and the historical data problem is, if anything, deeper.
Organisations that have already built structured AI bias audit processes for hiring are finding it’s a short extension to apply the same framework to performance tools. The mechanics barely change. What changes is that the person affected by a biased output isn’t a rejected candidate who may never know why. It’s someone already inside the organisation, watching a rating or a promotion decision they can’t fully explain.
In The End…
Whether AI removes manager bias or quietly reinforces it doesn’t have a clear yes or no answer, and any vendor or consultant offering one should be treated with some scepticism. What’s actually true is narrower and more useful: AI can reduce specific, well-documented forms of manager bias, and it can just as easily launder older, structural forms of the same bias into something that looks like data-driven objectivity.
The difference between those two outcomes isn’t the technology. It’s whether anyone in the organisation is actually checking the training data, running the demographic splits, and keeping a human genuinely empowered to say the model got it wrong. HR teams that have already built structured, defensible performance review processes have a head start here, since the same discipline, clear criteria, documented reasoning, checked against outcomes, applies whether the reviewer is a manager or a model.
The test is simple enough to run in any organisation this week: take one AI-influenced decision from the last calibration cycle and ask to see the data behind it. An organisation that can’t produce a clear answer has found its actual finding, whatever the model’s dashboard says.
FAQs
Does AI reduce bias in performance reviews?
Partially. AI removes fatigue and recency effects that skew human ratings, but it can also inherit bias already baked into the historical data it was trained on.
Can AI performance tools be biased?
Yes. A model trained on past ratings or promotions shaped by gender, location, or college pedigree will treat those patterns as predictive rather than correct for them.
What does India’s DPDP Act require for AI-driven performance scoring?
The DPDP Act, 2023 treats performance data as personal data subject to purpose limitation. Companies must be able to explain what data trained an automated score, and why.
How can HR audit an AI performance tool for bias?
Check what historical data trained it, run outputs against demographic splits, confirm managers can see why a score was generated, and verify human overrides are actually tracked.
What is “proxy bias” in an AI performance tool?
It’s when a model uses a correlated stand-in, like project staffing or attendance logs, to indirectly reproduce a bias such as gender, without ever using that attribute directly.
