BCG’s AI-at-Work Study: Productivity Gains Require Human Review and Workflow Changes
Checked out by Neil C. Hughes · 813 reads
A BCG experiment found AI helped consultants tackle unfamiliar tasks, but lasting skills and faster workflows depend on human checks and deliberate process design.
Generative AI can help employees do work beyond their current expertise, but a task completed with AI isn’t proof that the person has learned the skill. That distinction sits at the heart of a Boston Consulting Group experiment involving 480 consultants: participants using AI saw improvements of up to 49% on complex tasks outside their skillset. For employers, the finding points to an opportunity and a management problem. AI can expand what teams accomplish, while making oversight and learning harder to leave to chance.
In an interview with Techopedia, Vladimir Lukic, BCG managing director and senior partner, describes the effect as an “exoskeleton”: AI supports a worker’s performance, but the capability may disappear when the tool is taken away. The experiment was conducted with scholars from Boston University and OpenAI’s Economic Impacts research team. Its findings are also described in a research paper.
AI extends performance, not necessarily expertise
Lukic uses coding to illustrate the difference. A person with no programming background may ask a generative AI tool for a script and receive dozens of lines of Python. That can make a previously inaccessible task possible. But if the code is wrong, checking and fixing it may require knowledge the user hasn’t acquired.
“Think of GenAI as an exoskeleton. It can suddenly let a non-coder write code, but once you remove the tool, that person can’t do it alone. They haven’t learned to code.”
BCG’s experiment found that consultants without coding experience could use AI to complete advanced tasks. Their performance without the tool, however, did not retain the same improvement. The natural-language interface lowers the barrier to producing code, but it doesn’t remove the need to assess whether that code works as intended.
Nor are the benefits limited to beginners. Lukic says experienced data scientists also performed better when using AI. In his account, the tool can help less-experienced workers bridge gaps while giving specialists a faster way to work or extend their capabilities.
“People without backgrounds had enormous gains, but skilled folks also performed higher with AI augmentation.”
That combination matters for how organizations assess AI-enabled work. A strong output may reflect a useful interaction between worker and tool, rather than either person’s standalone expertise. The practical question is not just whether AI can produce an answer, but whether someone on the team can evaluate it and explain what it means.
Build review into the work
AI can also make it easier to skip the steps through which employees develop judgment. Lukic cautions against handing entire tasks to a model without considering what workers need to learn. His proposed response is to make some effort deliberate: employees should inspect intermediate work, break generated code into smaller components or get a manager’s review.
That isn’t an argument for preserving every manual task. It’s a case for deciding which steps serve a learning or quality-control purpose before removing them. If junior staff only see a finished AI-generated result, they may have fewer chances to build the intuition needed to spot errors the next time.
The same division of labor applies to analysis. Lukic recalls AI summarizing 50 interviews and producing an initial set of themes that was more extensive than any single analyst’s list. Yet people still caught subtleties the model had missed. AI can provide a broad first pass; human expertise remains relevant when interpreting nuance and making connections.
Redesign the workflow, not just the task
The biggest productivity gains may depend less on the tool than on what an organization changes around it. If AI reduces execution time but employees must still wait through the same approvals and committees, the faster task may not make the overall process faster. Lukic says BCG has seen processes go from ten days to two minutes, while warning that old sign-off structures can restore the original turnaround.
“That’s wasted potential.”
His advice is to revisit workflows, job requirements and performance measures together. If AI changes which tasks employees can handle, hiring and team design may need to account for how people work with the tools, not only the skills they already bring. Lukic frames this as “AI synergy,” with implications for recruiting and internal movement between roles.
Leaders also need direct experience using AI, he argues. Telling teams to adopt it won’t give executives a feel for what it can do reliably or where it falls short. Trying it on practical work, such as drafting a memo or analyzing data, can help managers make more informed decisions about where to place checks and how to redesign a process.
Speed brings two risks to manage
Lukic identifies a pair of concerns. First, workers may become faster while losing practice in skills that AI handles for them. Second, if an organization doesn’t offer a safe internal route for using these tools, employees may turn to public services, potentially exposing company data.
Those risks call for more than a blanket instruction to use AI or avoid it. Teams need clear ways to review outputs, protect sensitive information and preserve opportunities to learn. The BCG findings suggest that AI can raise performance for both novices and specialists, but immediate task completion and durable expertise are different outcomes.
The useful test for employers is therefore twofold: does AI improve the work, and does the surrounding process help people verify, learn from and safely use the result? The answer will depend on workflow design as much as model capability.
