Korea's investment agency stopped guessing which firms would expand, and let a model do the shortlisting.

Every investment officer runs a model in their head. They read the filings, the hiring, the customs flow, and they place bets on which companies are about to add a line or a second site. The trouble is that the model lives in a few senior heads, it does not scale past their calendars, and it walks out the door when they retire. What KOTRA, Korea's trade and investment agency, has done is write that instinct down as machine learning. The result is instructive less for the technology than for the three choices around it, each of which an agency with a fraction of KOTRA's budget can copy.

The first choice was the target. KOTRA did not point the model at the glamorous problem of finding brand-new foreign investors who have never heard of Korea. It pointed it at reinvestment, scoring companies already in the country, or already known to it, on how likely they were to put in more. That is the cheapest FDI there is. The relationships exist, the data is real rather than purchased, and an expansion closes faster than a greenfield courtship. Between 2023 and 2025, 38.5% of the companies the model flagged went on to invest, against a 17.7% baseline for the agency's normal targeting. That is better than double, and it landed on the slice of the pipeline most agencies neglect because aftercare is quiet work and greenfield wins get the press release.

The second choice was ownership. The model was not bought from a vendor and dropped on the IT department. KOTRA stood up an in-house task force that put its digital-transformation engineers in the same room as working FDI specialists, and fed the thing on inter-agency data it could actually get: customs records, financial statements, employment figures. The officers then stayed in the loop after go-live. They review every lead the model surfaces, they adjust its weightings when market intelligence tells them the data is missing something, and they validate each proposal before anyone contacts an investor. A targeting model the front line neither trusts nor tunes is shelfware with a dashboard. This one earns its place partly because it also handed those officers back more than a thousand hours a year they had been losing to manual screening.

The reason this matters well beyond Seoul is that the capability is concentrating. UNCTAD's latest agency survey finds that 82% of the IPAs reporting AI use sit in high or upper-middle-income economies. Across 76 agencies in the least-developed economies and small island states, just 13% have deployed any visible AI tool, and most of those are simple chatbots. Investor targeting is quietly becoming another axis of the digital divide, with the best-resourced agencies getting sharper at exactly the reinvestment game the rest cannot yet play. Left alone, the gap compounds: better targeting wins more expansions, expansions generate more data, and more data sharpens the next model.

The encouraging half of the same report is how low the first rung sits. The functions ladder from operational automation, a chatbot handling routine investor questions, through information synthesis and predictive targeting up to generative drafting, and no agency has to start at the top. Entry-level tools run between $300 and $6,000 a year, roughly the price of a CRM add-on. When the Democratic Republic of the Congo's agency put a chatbot on its investor enquiries, it reported 30% more qualified contacts, responses running 50% faster and a 20% lift in conversion, none of which needed a data-science unit. The precondition is boring and non-negotiable: clean, structured data. A predictive model trained on bad records will point officers confidently at the wrong companies and burn investor trust faster than having no model at all.

So the move to copy is not KOTRA's algorithm, which is calibrated to Korea's data and mandate. It is the sequence. Start with the segment where you already own the data and the relationship, your existing investors, and ask the narrow question a model can actually answer, which of them is about to grow. Keep the tool in the hands of the officers who will act on it. Spend the first money fixing the data rather than buying cleverness. And before any of it, write down the one number that tells you whether it worked: the share of the companies you target today that actually invest. KOTRA knew its was 17.7%. Most agencies have never worked theirs out, which is the real reason they cannot yet say whether a machine would beat them.