Stop chasing everyone: the quiet lesson in how the best agencies are actually using AI

For a few years the AI conversation inside investment promotion agencies has been mostly theatre: a chatbot bolted onto the website, a pilot announced at a conference, a working group that meets and adjourns. UNCTAD's latest IPA Observer, published this month, marks the point where that changes. Its survey finds AI moving from early testing into routine operations, and it sorts the use into four functions that are worth memorising because they double as a value ladder. At the bottom sits operational automation — the chatbots and workflow tools that field routine investor questions and licensing steps. Above it, information synthesis, to digest the flood of market and company data an agency drowns in. Above that, predictive decision-making, to estimate which investors are actually worth pursuing. And at the top, generative AI, drafting proposals and pitch materials. Most agencies start at the bottom rung. The returns are at the top.

The clearest evidence of what the top of the ladder buys comes from the Republic of Korea. Invest KOREA, the FDI arm of KOTRA, built a model that scores existing and prospective investors on their likelihood of putting in more money — and then concentrates the agency's outreach on the ones the model flags. Of the companies it identified, 38.5% went on to invest between 2023 and 2025, against a 17.7% baseline for the field at large: more than double the hit rate. The same system freed up over a thousand staff hours a year. This is not a chatbot saving a receptionist some typing. It is a targeting instrument that changes who an investment officer spends a Tuesday calling — and the difference between a 17.7% and a 38.5% conversion rate is the difference between a team that chases everyone and one that chases the right names.

What is easy to miss is that the Korean result is not really a software story. The model was built in-house, by a task force that paired KOTRA's digital team with FDI specialists who knew the work — so the engineering stayed tethered to how promotion actually happens, not how a vendor imagines it. And it runs on data the agency had earned access to: a sharing arrangement with the Korea Customs Service that lets the model see real trade and investment behaviour rather than guess from a stale CRM. Estonia's comparable effort works for the same reason — its AI sits inside a national e-government backbone that already holds clean, connected records. The intelligence is downstream of the plumbing. Where the plumbing is good, the model has something to learn from; where it is not, the smartest algorithm produces confident nonsense.

Which is why the global picture is less a technology gap than a capacity one — and it is widening. Among agencies reporting AI use, 82% sit in high- or upper-middle-income economies; only 16% are in least developed countries or small island states. Across 76 agencies UNCTAD reviewed in those poorer economies, just 13% had any visible AI tool at all. The reflex reading is that rich agencies can afford the software. The truer reading is that they can afford the years of unglamorous data work the software needs — investor registries that are current, CRM records that are filled in, customs and tax data they are allowed to see and able to join together. An agency without that foundation cannot leapfrog to a predictive model by buying one. It can buy a chatbot, which is exactly why the chatbot is where most stop.

For practitioners the danger is mistaking the visible rung for the valuable one. A chatbot is demonstrable; you can show it to a minister. A targeting model is invisible in the lobby but it is the thing that moves the conversion rate. The agencies pulling ahead, on UNCTAD's reading, share three habits and none of them is about the model: a clear strategy for what the tool is meant to change, careful sequencing so each step builds on a working one beneath it, and steady investment in the data and the people. They also keep a human in the loop on the decisions that matter — the model ranks the call list; the officer still makes the call. AI that fabricates a confident but wrong read of an investor does not just waste a week, it can cost the trust a small agency runs on.

So the craft move this year is to resist the demo and do the boring thing first. Pick one decision your agency makes over and over, that is rich in data you already hold, and where being right more often has a clear payoff — and aftercare is the obvious candidate. Existing investors are your best-documented relationships and, as Costa Rica's reinvestment record and now Korea's targeting both show, your highest-converting prospects. Score them, rank them, and point your limited outreach hours at the top of the list before you spend a cent on anything generative or public-facing. That is how the gap between a 17.7% and a 38.5% hit rate gets closed: not by buying intelligence, but by building the data underneath it and aiming the tool at the job that pays.