A BCG study on India’s GenAI adoption found that 92% of employees use GenAI regularly, compared with a global average of 72%, and only about one-third report adequate training.
ADP’s People at Work 2026 survey confirms the pattern from a different angle: 80% of Indian employees use AI multiple times a week, and 41% use it daily, the highest frequency recorded across any surveyed market.
The gap between usage and training is where every real risk lives. The Digital Personal Data Protection Act, 2023 holds employers (as “data fiduciaries”) liable for breaches of up to ₹250 crore per contravention. An employee pasting a client’s personal data into a public chatbot is now a compliance exposure, not a bad habit. The same pattern explains why Indian employees still don’t trust AI tools even as their own usage climbs.
Working “around” AI, then, really means letting liability compound in silence. The policy says one thing. The desktop activity says another. HR holds the gap.
What “With” Actually Looks Like, Done Properly
The companies putting serious money against the “with” option tend to be the ones already dependent on knowledge work. Wipro committed $1 billion over three years to train all 250,000 of its employees under its ai360 programme. TCS has made similar large-scale commitments through its “Ideate and Build with AI” initiative, and Infosys has rolled out tiered “AI-Aware, AI-Builder, AI-Master” certifications across its workforce.
Scale alone doesn’t settle the question of whether the training works. The same BCG study found that employees who get five or more hours of hands-on, in-person coaching are significantly more likely to become regular, confident users. Mandatory e-learning videos do not clear that bar, which is why enterprises are now rethinking AI-based reskilling and upskilling spend rather than just adding more modules.
From Tool Fluency to Judgment Training
Most AI training programmes still focus on the wrong thing. They teach prompt syntax and tool interfaces as though the problem were technical fluency. It isn’t. The problem is judgment: knowing what to delegate, what to verify, and what to never send to a public chatbot in the first place.
Judgment training looks different from a software demo. It uses real work artefacts (a sanitised offer letter, a redlined contract, a candidate screening transcript) and asks employees to decide where AI helps, where it introduces risk, and where the output quietly degrades the task. This is closer to creating differentiated AI literacy paths than to running yet another compliance deck.
Training the Override, Not Just the Prompt
The second shift is the opposite of trust. Employees need to recognise when a model is confidently wrong, when the output needs a second human reader, and when the right answer is to close the tab and finish the work by hand.
That capability doesn’t develop on its own, and no vendor dashboard will install it. It comes from supervised practice, feedback from a senior reviewer, and a culture that treats “I didn’t use the tool here because X” as a correct answer, not a lazy one.
Where Human-Only Lanes Still Earn Their Keep
There are workflows AI shouldn’t be inside, and naming them is part of a “with” strategy, not a contradiction of it. The companies that get this right design explicit no-AI zones and defend them as rigorously as they defend the rest of the AI estate.
POSH investigations are the clearest example. The Internal Committee is a legal body under the POSH Act, 2013, with specific procedural obligations around confidentiality, documentation and timelines.
Running a complaint through a generative tool for “efficiency” compromises confidentiality, creates discoverable artefacts, and may expose the complainant to re-identification risk through training data, which is why workplace AI tools and POSH risks have become a serious concern for HR leaders.
Grievance conversations, performance improvement plans, exit interviews, mental health disclosures and sensitive compensation discussions sit in the same category. The quality of these conversations is the organisation’s duty of care. A chatbot drafting the manager’s script is a bad trade, however much time it saves.
Human-only lanes aren’t nostalgic resistance. They’re a design choice, and they’re defensible only if the rest of the AI estate is governed well.
The Shift HR Teams Have to Lead
The practical move isn’t buying another LMS seat or bolting one more compliance module onto onboarding. It’s treating AI competence as a core job expectation, with layered training that matches what the role actually does and distinguishes one-off training from durable learning.
A workable training stack looks like this:
| Layer | What It Covers | Who It’s For |
| Baseline AI Literacy | What LLMs are, what they can’t do, DPDP-sensitive data types | Everyone |
| Function-Specific Workflows | Vetted use cases per role, approved tools, prohibited inputs | By team |
| Verification Skills | Fact-checking, source checking, hallucination spotting | Everyone touching AI output |
| Judgment and Override | When not to use AI, when to escalate, when to redo by hand | Managers and above |
| Policy and Compliance | DPDP consent trails, client data rules, incident reporting | Everyone |
A legal team’s AI risks are not a marketing team’s. A customer service team’s risks are not a product manager’s. Generic literacy training satisfies the compliance tick-box and almost nothing else.
In the End…
“With” versus “around” sounds like a philosophical debate about human-AI collaboration. It isn’t. It’s an operational question about whether a company puts rules, training and tools around behaviour that’s already happening, or keeps pretending it isn’t.
Moves that actually matter this quarter:
- Publish an AI use policy that names approved tools, forbidden data types (client PII, salary data, legal drafts, M&A material) and consequences for violations. Vague “be responsible” guidance is cover for liability.
- Run function-specific training sessions, not generic literacy modules. Give each team two or three real artefacts and let them map out safe and unsafe uses before anyone touches a live prompt.
- Designate human-only workflows in writing. Put POSH, grievance handling, mental health conversations and sensitive compensation decisions off the AI table formally, not as unspoken norms.
- Measure training by whether behaviour changes, not by completion rates. Audit a sample of prompts, outputs and incident reports every quarter, and tie findings back into the next cycle.
Teaching people to work with AI isn’t an endorsement of every tool or every use case. It’s an admission that the shadow economy is already inside the organisation, and that pretending otherwise is the most expensive position a company can take.
FAQs
How widespread is AI use in Indian workplaces right now?
A BCG study found 92% of Indian employees use GenAI regularly, against a global average of 72%, while ADP’s People at Work 2026 survey put daily AI use at 41%, the highest frequency across any surveyed market.
What’s the compliance risk when employees use AI tools informally?
The Digital Personal Data Protection Act, 2023 holds employers liable as “data fiduciaries” for breaches of up to ₹250 crore per contravention. An employee pasting client personal data into a public chatbot now reads as compliance exposure, not a bad habit.
How much are Indian companies investing in AI training?
Wipro has committed $1 billion over three years to train all 250,000 employees under its ai360 programme. TCS runs its “Ideate and Build with AI” initiative, and Infosys has rolled out tiered “AI-Aware, AI-Builder, AI-Master” certifications.
Does mandatory AI e-learning actually work?
Not reliably. The same BCG study found employees need five or more hours of hands-on, in-person coaching to become regular, confident users. Compliance videos don’t clear that bar.
What should AI training actually teach beyond prompts?
Judgment, not tool fluency. Employees need to know what to delegate, what to verify, and what to never send to a public chatbot. The second shift is training the override: recognising when a model is confidently wrong and when the right answer is to close the tab.
Which workflows should stay AI-free?
POSH investigations, grievance conversations, performance improvement plans, exit interviews, mental health disclosures, and sensitive compensation discussions. Running these through a generative tool compromises confidentiality and creates discoverable artefacts.
Why are POSH investigations specifically off-limits for AI?
The Internal Committee is a legal body under the POSH Act, 2013, with specific obligations around confidentiality, documentation, and timelines. Using AI risks re-identification of the complainant through training data and creates procedural exposure.
What should a workable AI training stack look like?
Five layers: baseline AI literacy for everyone, function-specific workflows by team, verification skills for anyone touching AI output, judgment and override training for managers and above, and policy and compliance covering DPDP consent trails and incident reporting.
How should HR measure whether AI training is working?
By behaviour change, not completion rates. Audit a sample of prompts, outputs, and incident reports every quarter, and feed the findings into the next training cycle.
