Domain expertise is the variable most people leave out when they try to explain why two colleagues get wildly different results from the same AI tool. The assumption running underneath most conversations about these tools is that they flatten the field, that a well-constructed prompt hands anyone an expert-level output and seniority stops counting for much. The evidence points somewhere more awkward than that.
be10X co-founder Aditya Goenka put the counter-case plainly in Forbes India, arguing that AI does not deliver its largest gains to people starting from zero but to people who already know their field, so that a strong marketer or analyst working with AI becomes a considerably stronger one rather than a replaceable one. On that reading, AI amplifies expertise instead of substituting for it.
The study that appears to say the opposite
The same article cites an NBER study of customer support work, which found that AI assistance lifted productivity by roughly 14 percent, with the largest gains concentrated among the least experienced staff. Read quickly, that looks like a flat contradiction. Novices gained most, so domain expertise must matter less than claimed.
Both findings are correct. The reconciliation is the part worth having.
Floors and ceilings are not the same economics
Customer support at scale is work with largely known-correct answers. A novice fails at it by not knowing the answer, and a system with access to the right material supplies that answer. What improved in that setting was the floor. The gap between the weakest agent and the accepted standard closed, and domain expertise was not the binding constraint there.
Judgment-heavy work behaves differently. No knowledge base holds the correct answer for which market to enter, which anomaly in a dataset is real rather than an artefact, or which campaign result is signal rather than noise. AI can generate a hundred plausible options for any of those. Choosing correctly among them is a ceiling problem, not a floor problem, and ceilings are raised by depth.
Domain expertise is what makes evaluation possible
This is the practical mechanism, and it gets skipped constantly. Someone who knows a field can tell when an output is wrong, often without checking it against anything, because the answer contradicts something learned from having done the work. A fluent, confident and entirely incorrect output is close to invisible to a reader without that grounding.
be10X works the same problem from the other end in its piece on the one question to ask AI before trusting any data insight, which makes the point that confidence and accuracy are not the same signal, and that a clean, specific-sounding conclusion can be entirely unsupported by the data underneath it. Domain expertise is a large share of what catches that.
Which explains why returns look so uneven inside companies
Two teams handed identical tools produce different results, and the diagnosis usually lands on training or adoption. Often the real variable is the domain expertise the team already held before AI arrived. A group of specialists adds AI and compounds what it was already good at. A group without depth adds AI and produces more output that nobody in the room can vet, which registers as activity rather than progress.
The honest part: depth can also work against a professional
Domain expertise is not automatically an advantage. Long experience builds strong priors, and strong priors make it easy to reject a correct output because it looks unfamiliar, or to defend a manual process precisely because the expert is unusually good at it and has the most to lose by handing it over. The professionals who lose ground are frequently not the least skilled ones.
They are the ones whose skill was attached to a specific method rather than to the judgment underneath it. be10X names the mechanism in its post on how AI is reshaping mid-career professionals, observing that a fresher arrives with no old habits to unlearn and no established workflows to rebuild, while an experienced professional carries both, which makes integration the harder task rather than learning.
What this changes about where to spend learning time
The instinct when facing AI is to start over somewhere else: a new field, a new toolset, an identity switch. For most people the stronger move is narrower. Take the function already understood in depth, isolate the part of it that is repetitive and rule-bound, and hand that portion across. The domain expertise stays as the asset. The tool absorbs the volume.
That combination is also why generic tool tutorials tend to disappoint. They teach a tool to an audience with no shared context, and the value of the tool depends almost entirely on context. The people advancing fastest are not chasing every new release. They are picking two or three tools that touch work they already know well, and going deep on those.
Anyone weighing up the best AI course for converting existing domain expertise into an AI-assisted advantage should be looking for workplace workflows rather than theory, which is what be10X’s AI Career Accelerator Program is built around.
