The C-suite is feeding AI gamed data, and making workforce decisions on it

AI has crossed a threshold in the boardroom. It has moved from the margins to the centre of executive decision-making, and active use in strategic decisions is set to more than double within three years, according to Capgemini's survey of 500 CXOs.¹ The Conference Board's 2026 outlook puts the same shift in plainer terms: AI now shapes workforce planning directly.² This is not hypothetical. Some 31% of companies have already cut frontline headcount because of AI, the third most common reason for reductions in 2025.³

What makes this worth a warning is that executives have already named the weak link, and not connected it to their own people. In the Capgemini work, 60% of CXOs cite lack of quality enterprise data as a top concern in AI-supported decisions, and 53% rank poor, biased or hallucinated output as the leading risk.¹ Poor data quality is not a soft governance matter. Over a quarter of organisations lose more than five million dollars a year to it, and the cost scales with AI spend, because a model amplifies whatever it is fed.⁴

Here is the thesis, stated plainly because this audience skims: trust-based workforce analytics is not an employee-experience nicety but a hard data-quality requirement for any AI consuming workforce signals.

Garbage in, but gamed, not random

The most corrupted dataset most organisations own is the one describing how their people work. Surveillance-style monitoring measures presence, keystrokes and screen time performed under observation, which is compliance theatre rather than authentic activity. The OECD's review of algorithmic management notes that 55% of US firms now use monitoring tools, and that evaluation tools are the category most likely to produce biased treatment of workers.⁵ The signal is degraded at the point of collection.

This is where the obvious counter-argument enters: modern models are designed to tolerate noisy, incomplete, outlier-laden inputs, so the priority should be model capability, not chasing clean data. Peer-reviewed work does make that case for ordinary noise.⁶ It does not rescue surveillance data, because surveillance data is not statistical noise. The distortion runs one way, since everyone is optimising to look busy. Tolerance for random scatter is not tolerance for systematic distortion, and no architecture recovers a pattern that was never honestly recorded. The same research still concludes that data quality must be a top priority.⁶

When AI eats biased people-data, it launders the bias

The failure compounds when the data is historical. Feeding biased or unrepresentative workforce data into AI reproduces and amplifies the bias at scale, and the data-driven framing lends those outputs what MIT Sloan calls an 'aura of neutrality' that makes them harder to challenge than a human judgement.⁷ The peer-reviewed review of AI-enabled recruitment is blunt about the mechanism: bias stems chiefly from data selection, and if the input is skewed the output will be skewed.⁸ AI does not fix bad people-data. It industrialises it, and then dresses the result as objectivity.

That is the real stakes. A discriminatory human call can be questioned. A discriminatory model output, wrapped in the language of data, is far more defensible to the person acting on it and far more damaging at volume.

Fix the collection layer, not just the decision gate

The reflex defence is human oversight: keep a person in the loop on every high-impact people decision and the input matters less. Harvard research is right that today's AI cannot substitute for human judgement, so oversight is necessary.⁹ It is not sufficient. A reviewer can only judge what the data shows. If it encodes gamed behaviour or historical bias under an aura of neutrality, the human rubber-stamps a flawed recommendation they have no way to see through. Oversight guards against bad data only when leaders can also tell that the data is bad, which means repairing collection rather than policing the decision.

A trust-based layer is what produces an honest signal. When employees are the primary users and beneficiaries of a workplace-analytics tool and control consent, they stop gaming it, and the patterns reflect how people actually work. Capgemini's own guidance points the same way: decisions with significant people impact always need human involvement and review, with AI recommending rather than deciding.¹ Before scaling any AI on top of workforce data, three questions sort honest signal from gamed garbage. Is it gamed, collected under observation and optimised for appearance? Is it consented, so people have no reason to distort it? Is it representative, or does it encode a bias the model will launder?

Whoever answers those questions first gets a structural advantage, because they will out-decide rivals running sophisticated models on corrupted inputs. The uncomfortable version of the question for any executive scaling AI this year: how confident are you that the workforce data feeding your models describes how people work, rather than how they perform being watched?

Footnotes

1. Capgemini Research Institute. (2026). Inside the C-suite: How AI is quietly reshaping executive decisions. Capgemini. https://www.capgemini.com/wp-content/uploads/2026/01/Final-Web-Version-Research-Brief-Gen-AI-in-Decision-Making.pdf ↩ ↩² ↩³

2. The Conference Board. (2026, January 15). AI and the C-suite: Implications for CEO strategy in 2026. https://www.conference-board.org/research/ced-policy-backgrounders/ai-and-the-c-suite-implications-for-ceo-strategy-in-2026 ↩

3. Mourgelas, I. (2026). C-suite survey finds AI already cutting jobs at one-third of companies, even as hiring rebounds. Chief Executive. https://chiefexecutive.net/c-suite-survey-finds-ai-already-cutting-jobs-at-one-third-of-companies-even-as-hiring-rebounds ↩

4. Krantz, T., & Jonker, A. (2026, January 23). A compounding threat: The true cost of poor data quality. IBM Think Insights. https://www.ibm.com/think/insights/cost-of-poor-data-quality ↩

5. OECD. (2025). Algorithmic management in the workplace. OECD Publishing. https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/02/algorithmic-management-in-the-workplace_3c84ed6d/287c13c4-en.pdf ↩

6. Mehrotra, R., et al. (2023). Re-thinking data strategy and integration for artificial intelligence. Applied Sciences, 13(12), 7082. https://www.mdpi.com/2076-3417/13/12/7082 ↩ ↩²

7. Carey, N. (2025). AI is reinventing hiring, with the same old biases. Here's how to avoid that trap. MIT Sloan Ideas Made to Matter. https://mitsloan.mit.edu/ideas-made-to-matter/ai-reinventing-hiring-same-old-biases-heres-how-to-avoid-trap ↩

8. Köchling, A., & Wehner, M. C. (2023). Ethics and discrimination in artificial intelligence-enabled recruitment practices. Humanities and Social Sciences Communications, 10, Article 567. https://www.nature.com/articles/s41599-023-02079-x ↩

9. Harvard Business School Institute for Business in Global Society. (2025). AI won't make the call: Why human judgment still drives innovation. HBS BiGS. https://www.hbs.edu/bigs/artificial-intelligence-human-jugment-drives-innovation ↩