Ana-Maria Badulescu has over 20+ years of software industry and data management domain experience.
She currently leads Precisely AI Labs, a team of data scientists focused on innovation using artificial intelligence to solve emerging customer challenges with data integrity. Prior to her current role, Ana-Maria has held various technical and leadership roles developing the big data and data integration family of products, as well as conducting research on emerging technologies and incubation of new product offerings. Ana-Maria received a Master’s degree in Computer Science from University of California, Irvine, and a Bachelor’s degree in Computer Science from Politehnica University, Timisoara, Romania.
Approximately 70-80 percent of organisations use AI in at least one business function.
Tools like generative AI, chatbots and AI assistants enhance efficiency, facilitate real-time decision making and optimise workflows. However, AI carries risk. Many users take outputs at face value, and the integrity of the data fuelling those systems is often ignored.
Nevertheless, AI’s outputs are a result of the information inputted into the system, whether through training data or through the interactions and feedback they receive from users. If that data or user base is unrepresentative, then it will create biased results. As AI is given greater automation and influence, it could unintentionally reinforce discrimination. In fact, the International Labour Organisation has even issued a warning that generative AI isn’t gender neutral and often reflects and reproduces the gender biases embedded in society.
The real-world impact of gender bias in AI
The rise of autonomous AI agents makes the challenge of gender biased data even more critical. While standard LLMs generate text or image-based outputs, agents can act autonomously to carry out real-world actions. Without human oversight, biased decisions could have severe social, ethical and business implications across industries.
Additionally, biased decisions don’t only impact women, they can impact entire organisations. AI models operating on unrepresentative data can lead to flawed market insights, inaccurate results and financial losses.
For example, research shows that AI tools may favour male university applicants over female ones as they label women as having a higher dropout risk. When educational institutes use biased systems, they risk litigation and compliance penalties under anti-discrimination laws like the UK’s Equality Act 2010.
Beyond these regulatory risks, gender bias presents a serious concern in healthcare. Research from MIT revealed that when AI was presented with identical clinical symptoms, female patients were significantly less likely than male patients to receive a formal clinical evaluation.
Such disparities occur because AI can produce biased outputs and create unreliable assumptions when running on inaccurate, incomplete or unreliable data. If biased data is entered into a system, it can accidentally result in automated actions that systematically disadvantage certain groups.
As LLMs and agentic systems become embedded across public and private organisations, the biased outputs produced by these systems require urgent correction, or they risk amplifying societal gender inequalities further.
Creating trustworthy and transparent AI
Establishing a foundation of data integrity is essential for addressing bias and creating AI-ready data. This requires moving beyond policy statements to concrete practice: breaking down silos, enriching training data with curated, AI-ready attributes and spatial insights, and enforcing rigorous governance.
Start with data completeness. AI trained on siloed or fragmented data can only reflect a partial view of reality. For example, when HR, clinical, or admissions data lives in separate systems, AI is limited to only a fraction of the available information, and lacks the context to surface patterns across groups. This results in gaps in women’s representation going undetected. Integrating data across cloud and hybrid environments ensures completeness, significantly reducing the potential for biased outputs.
Enrichment matters equally. If training data draws heavily on historical records – for example, hiring decisions from the 1990s, or clinical studies with male-majority cohorts – the model will encode those skews as norms and simply replicate and automate with outdated information rather than reflecting current reality. Combining first party-data with curated third-party sources, such as labour market demographics or population health data, gives AI a more accurate baseline to work from.
Governance is also critical. 71 percent of organisations that report high data trust share a common trait: fairness and transparency are built into the data pipeline itself, not reviewed after the fact. In practice this means defining bias thresholds before deployment, requiring demographic balance checks as part of data ingestion, and assigning clear ownership for monitoring gender representation over time. Observability tools can flag when distributions shift – for example, if training data for a recruitment tool begins skewing male following an acquisition – before those shifts affect outputs.
In practice, three questions should be non-negotiable before any AI system goes live: Does our training data reflect the demographic breakdown of the population it will affect? Can we flag when that balance drifts post-deployment? And who is accountable when it does? If any answer is “we’re not sure,” the system isn’t ready.
Even when organisations improve training data, bias can re-emerge through uneven adoption and participation. AI systems are shaped not only by historical data, but also by the people who actively engage with them.
The impact of who uses AI
Data quality is only half the problem. The other half is who shapes the system once it’s running.
This matters because modern AI systems don’t just reflect their training data; they are increasingly shaped by user interactions, feedback loops, and workflow patterns after deployment. Every prompt, correction and interaction is an input. If women are underrepresented among the people testing, challenging and applying these tools, their needs and edge cases become less visible in the workflows that emerge. The model doesn’t just fail to represent women – it gradually optimises away from them.
The gap is already visible. McKinsey found that only 21 percent of entry-level women report being encouraged by their managers to use AI, compared to 33 percent of men at the same level. Women are also more likely to perceive AI use as a “shortcut” – a framing that discourages experimentation and signals that AI proficiency is less expected of them.
From data to action: closing the gender gap in AI
As AI adoption accelerates, exposure to AI-related bias inevitably rises. Mitigating gender bias requires a proactive strategy that addresses both the data feeding these systems and the people shaping them.
On the data side, investing in balanced, representative datasets, breaking down silos, and embedding governance into the pipeline – not bolted on afterwards – are the foundations of AI that works fairly for everyone. With 66 percent of people relying on AI outputs without evaluating their accuracy, the cost of getting this wrong is not abstract.
On the human side, the barrier to more inclusive AI use isn’t aptitude – it’s access, encouragement, and organisational culture. Google’s AI Works pilot found that just a few hours of targeted training can double daily AI usage, with that uplift sustained months later. Critically, many employees simply needed explicit permission to use AI before they would engage with it at all. Role-specific training, visible leadership sponsorship, and tracking who actually uses these tools – and who doesn’t – are levers organisations already have. They just need to use them.
Both gaps are closeable. Organisations that treat gender bias as a data integrity challenge – not merely a compliance exercise – will build AI systems that are more accurate, trusted, and commercially effective. In an era where AI increasingly shapes decisions, trust is not a byproduct of innovation – it is a prerequisite for it.
