We're All Analysts Now

The workflow has three layers: human, then AI, then human. The first layer is human judgment, deciding what to ask and what sources to trust. The old programmer's warning applies: garbage in, garbage out. The AI is the filling, doing the gathering and the drafting. Then human judgment closes the sandwich, catching what was missed and testing whether the output holds against the commercial and political realities no model holds.

The trouble is that a lot of organizations are fixated on the bread. They want to know which tools to buy and how many hours they can save. Those are fair questions. The one that decides the outcome is whether the human layer is strong enough. If the person in the middle has poor judgment, AI will not rescue the work. It will help produce weak thinking faster.

The evidence on this comes from general office work, not from this field. BetterUp Labs and the Stanford Social Media Lab use the term "workslop" for AI output that looks polished but carries little substance. In their September 2025 survey of 1,150 US desk workers, 40% said they had received workslop from a colleague, and each instance cost roughly two hours to sort out. The labs' reading of that record is that AI amplifies whatever is already there. Skilled people use it to extend their expertise, while less-skilled people are more likely to generate passable-looking work that hides shallow reasoning and slows their own development.

The same pattern turns up in software, the other borrowed case, where it can be measured precisely. Salesforce's own engineers, writing about the surge in AI-generated code, found that their file-by-file review process started to break down: reviewers could no longer reconstruct the intent behind changes that touched backend logic and the interface at once. A study by CodeRabbit in 2025 across hundreds of open-source projects found that AI-assisted pull requests contained about 1.7 times as many issues as human-written code, with the sharpest rise in logic and correctness errors, even as the raw volume of work went up.

A thin market scan can now be formatted beautifully, and a briefing note can have exactly the right structure while skating past the hard question. The output looks professional without being good.

For economic development professionals, this cuts both ways. The risk is easy to describe. Teams get overconfident because the first draft already looks finished. Junior staff skip the slow work of understanding a market. Senior leaders receive slicker documents with thinner thinking inside them, and agencies mistake a productivity gain for a strategic one.

The opportunity is larger. AI can strip a lot of low-value friction out of the work, the repetitive drafting and formatting that eats time without producing insight. It can give a regional agency better first drafts.

But that value only shows up if the human spends the recovered time on better analysis, which mostly means asking better questions. The questions that matter are what is changing in this market, and what this investor needs that the region can credibly offer. Those are human questions, and they are the ones that will define the profession.

Economic development is often presented as a toolkit: export promotion, investment attraction, and a dozen other instruments. The toolkit still matters, but AI will make many parts of it easier to reach. The differentiator becomes the practitioner's ability to decide which tool fits, when to use it, and what the output means. That is analysis: connecting information to action, and knowing when the obvious answer is wrong.

A trade mission is a good test. AI can help build the country brief and set up a follow-up tracker. That is useful. The trade professional's value is in deciding whether the mission has a reason to exist at all. Often the companies are not ready, the buyers who matter are not in the room, and nobody follows up once the reception ends. AI can help run the mission. It cannot care whether the mission is worth running.

Investment attraction works the same way. AI can help build a target investor profile and draft outreach. What it cannot judge is whether a region is credible for that investor. It does not know from experience that a particular company is unlikely to move, or that the workforce story reads far better in a slide than it holds up on the ground. The practitioner still has to tell the difference between a lead and a distraction.

The diplomatic side of this work may be where the line is sharpest. AI can produce a competent summary of a bilateral relationship and draft talking points. But the work a Trade Commissioner does is not mostly about content. It runs on trust and on the ability to hear what is being signaled without being said outright. We map that craft in Commercial Diplomacy and Relationship Tradecraft.

This is why the practitioner of the next decade needs to become more analytical, not less. The job is to direct AI and correct it, and outwriting it is a losing game. The practitioner shifts from producing documents to deciding what is credible and what should happen next.

That calls for a different kind of training. Most organizations do not need another generic AI webinar. They need to build analytical muscle: people who can use AI without handing their judgment to it, and who can get from output to insight. In practice that means learning to read markets and institutions, and to tell a weak source from a strong one, fast, under time pressure.

The strongest professional knows where AI belongs in the workflow and where it does not. Some work should never go into external tools: confidential commercial information and cabinet material. And there are moments when the fastest AI-assisted answer is not the best one.

A small regional team can produce sharper market scans without adding headcount, then spend the time it saved talking to more companies.

AI is becoming a commodity. If the major models stay broadly available at today's prices, teams competing for the same investor will be working with much the same tools within a year or two, and prompting will be an ordinary skill rather than an advantage. More people will produce an acceptable first draft, and the difference will sit underneath, in who can catch the weak assumption and tell activity apart from progress.

The strongest objection is that judgment is not a fixed category. Each model generation takes in work that looked like judgment a year earlier, and ranking a shortlist of candidate markets is already partly machine work. That concession narrows the claim: the line will keep moving, and some of what the profession now calls analysis will move with it.

The test is checkable. Take a year of screening calls on which investors were serious, compare each call against which projects were actually built, and see whether a model's ranking beats the practitioner's. If it does, the argument here fails.

The measured record so far runs the other way. In the BetterUp Labs and Stanford Social Media Lab survey of September 2025, the repair work after AI output still fell to a person, about two hours at a time.