What the Research Actually Says About AI and Productivity

Everyone has an opinion about AI. Your uncle thinks it's a fad. LinkedIn influencers think it'll replace every job by next Tuesday. The truth — as usual — is somewhere in between.

But we don't have to guess. Researchers at Stanford, Harvard, MIT, and other institutions have been studying AI's real impact on work. Not in theory. In actual workplaces, with actual workers, doing actual tasks.

Here's what they found.

The Stanford Study: Customer Support

In 2023, researchers at Stanford published a study tracking over 5,000 customer support agents at a Fortune 500 company who were given access to an AI assistant (Brynjolfsson, Li, & Raymond, 2023).

The results:

14% increase in productivity on average across all workers
34% improvement for the least experienced workers — the AI essentially helped new hires perform like veterans
Experienced workers saw smaller gains, because they'd already developed the skills the AI was providing

The takeaway isn't "AI makes everyone faster." It's that AI compresses the learning curve. A new employee with AI support can reach competence in weeks instead of months.

For Malaysian SMEs that struggle with training and high turnover — especially in customer service roles — this is a significant finding. The AI doesn't replace staff. It makes new staff useful faster.

The Harvard/BCG Study: Knowledge Work

A 2023 study from Harvard Business School, in collaboration with Boston Consulting Group, tested what happened when 758 management consultants used GPT-4 for typical consulting tasks (Dell'Acqua et al., 2023).

Key findings:

Consultants using AI completed 12.2% more tasks, 25.1% faster, with 40% higher quality compared to those without AI
On tasks within the AI's capability boundary, every consultant improved — regardless of skill level
But on tasks outside the AI's strengths, consultants who relied on AI actually performed 19% worse than those working alone

The researchers called this the "jagged technological frontier" — AI is brilliant at some things and unreliable at others, and the boundary isn't always obvious.

What this means in practice: AI is genuinely powerful for structured, information-heavy tasks like drafting, summarising, analysing data, and answering standard questions. But you still need human judgement for novel situations, strategic decisions, and anything requiring deep context that the AI doesn't have.

The MIT Study: Writing Tasks

MIT researchers (Noy & Zhang, 2023) ran a randomised experiment with 444 professionals — marketers, HR specialists, managers — and asked them to complete realistic writing tasks.

Results:

AI users completed tasks 37% faster
Quality improved, especially for workers who started with lower writing ability
The gap between "good" and "average" writers narrowed significantly when both had AI access

This aligns with the Stanford finding: AI acts as an equaliser. It doesn't make great writers better so much as it brings average writers closer to great.

For businesses, this means your team's output quality becomes more consistent. The variability between your best and worst communicator shrinks.

The Microsoft/GitHub Study: Software Development

GitHub's research on Copilot (Peng et al., 2023) tracked developers completing coding tasks:

Developers using Copilot finished tasks 55.8% faster
Completion rates were higher across all experience levels
Developers reported higher satisfaction and less frustration

This is specific to coding, but the principle applies broadly: when AI handles the repetitive, mechanical parts of a task, humans can focus on the creative and strategic parts.

What Does "14% More Productive" Actually Mean?

Numbers like "14% improvement" or "37% faster" sound abstract. Let's make it concrete.

If your customer service team handles 100 enquiries per day and AI makes them 14% more productive, that's 14 additional enquiries handled — without hiring anyone new.

If your marketing person spends 3 hours a day on writing and AI cuts that by 37%, that's roughly an hour saved daily. Five hours a week. Over a year, that's more than six full work weeks of time back.

These aren't transformative overnight. They're incremental. But they compound. And they compound every single day.

The Limits (Because There Are Limits)

The research is clear about where AI struggles:

Novel reasoning. AI models can synthesise existing knowledge well, but they don't truly "think" through unprecedented problems the way humans do.

Context-heavy decisions. Your AI doesn't know that Puan Aminah always orders extra sambal, or that your Penang branch manager prefers to be called on his mobile. Humans carry contextual knowledge that AI needs to be explicitly taught.

Emotional intelligence. Handling a genuinely upset customer, navigating office politics, reading the room in a meeting — AI can assist, but it can't replace the human element.

Accuracy on the edges. The Harvard study's "jagged frontier" finding is important. AI doesn't know what it doesn't know. It'll confidently produce wrong answers for questions just outside its capability boundary.

This is why the most effective AI implementations pair AI with humans, not replace humans with AI. The research consistently shows that human + AI outperforms either alone.

What This Means for Malaysian Businesses

Malaysia has some specific characteristics that make these findings particularly relevant:

Multilingual environment. Modern AI models like Claude by Anthropic handle BM, English, Mandarin, and Tamil naturally. The Stanford productivity gains likely underestimate the benefit in multilingual environments where the AI can switch languages seamlessly.

High mobile usage. With over 70% of internet traffic on mobile, Malaysian customers expect fast responses via WhatsApp and Telegram. AI assistants that handle messaging channels directly address this.

SME-heavy economy. Malaysia's economy is 97% SMEs. The research shows AI benefits smaller teams disproportionately — because the productivity gains per person matter more when you have fewer people.

Talent competition. Every Malaysian business owner knows how hard it is to find and keep good staff. If AI can compress training time (Stanford finding) and equalise output quality (MIT finding), that directly addresses one of the biggest pain points.

The Practical Question

The research paints a clear picture: AI produces real, measurable productivity gains — especially for routine tasks, customer-facing work, and content creation. The gains are largest for less experienced workers and most consistent for structured tasks.

The question isn't whether AI can help your business. The research has answered that. The question is which tasks to apply it to first, and how to set it up so your team actually uses it.

If you're thinking about where AI fits in your business, we're happy to talk it through. No pressure, no jargon — just an honest conversation about what might work for your specific situation.

Book a free discovery call →

References:

Brynjolfsson, E., Li, D., & Raymond, L. (2023). "Generative AI at Work." NBER Working Paper.
Dell'Acqua, F., et al. (2023). "Navigating the Jagged Technological Frontier." Harvard Business School Working Paper.
Noy, S. & Zhang, W. (2023). "Experimental Evidence on the Productivity Effects of Generative AI." Science.
Peng, S., et al. (2023). "The Impact of AI on Developer Productivity." arXiv.

Related reads: