We run dozens of distributor research and go-to-market consulting projects every year. Ask us for a proposal.

Article

AI in distributor search and selection: what the machine is now better at, and where it stops

A long-list of two hundred candidate distributors can now be assembled in an afternoon, screened against sanctions lists more accurately than a person would manage, and translated out of four languages. None of that was ever the hard part. The hard part is that the three facts which decide a distributor appointment are not published anywhere, by anyone, and the party that holds them has reasons not to share.

Published 5 November 202614 min read

There is a version of this article that argues machines cannot do judgement, relationships or nuance. It is worthless, because it cannot be checked and every competitor publishes it. So this is the other version: what the measured evidence says about distributor search and selection, including where automation now beats experienced people by a wide margin.

The conclusion is not that AI is oversold. On several tasks that used to absorb weeks of analyst time it is better than the people it replaced. The conclusion is narrower: AI has changed what it costs to find a candidate, and changed almost nothing about what it costs to know whether that candidate is real.

1. What AI is now measurably better at than people

Start with the case against the rest of this article.

On sanctions and name screening, the Federal Reserve Board's research arm tested four model families against the fuzzy-matching algorithms the compliance industry has used for twenty years. Against the best of those baselines, the models cut false positives by roughly 92 per cent while raising true matches detected by about 11 per cent: the difference between an analyst reviewing a thousand alerts and reviewing eighty, and finding more of what matters in the eighty.

On entity matching, the dull core of building any distributor list, a study of 755,540 labelled company and person pairs from 293 datasets across 31 countries put the best model at an F1 score of 98.95 against 91.33 for the rule-based system it replaced. One caveat matters here: the authors report model failures concentrating on cross-script transliteration between Arabic, Cyrillic and Latin, which is precisely the emerging-market case.

On facts that are genuinely buried, the result is more striking. A benchmark published in April 2025 built 1,266 questions each verified to have a findable answer that five ordinary searches would not surface. Human researchers, given around two hours per question, solved 29.2 per cent and gave up on 70.8 per cent; a purpose-built research agent solved 51.5 per cent. Where the answer exists somewhere and nobody can find it, the machine is now better than a determined person with an afternoon.

And the structural point underneath it: work on cross-border marketplace trade between 62 countries found the distance penalty roughly 65 per cent smaller online than offline, and the reduction largest for exporters from more corrupt and less well-known origins. Platforms substitute for the trust infrastructure missing where institutions are weak. That data runs to 2007, so treat the magnitude as indicative.

Where the argument actually sits

Anyone claiming AI cannot screen, match, translate or retrieve at scale is arguing against measurement and will lose. The interesting question is what happens to the 200-name list afterwards, when somebody has to decide which eight companies a client should meet and stake a recommendation on it.

2. The error that retrieval leaves behind, and why it does not look like a gap

The standard answer to model error is retrieval: stop asking the model to remember and make it read a source. It works, it is not enough, and the size of what is left over is now well measured.

The cleanest test comes from legal research, because the conditions are as favourable as they ever get: three commercial systems retrieving over closed, licensed, professionally curated databases. Researchers ran 202 preregistered queries through them for the Journal of Empirical Legal Studies in 2025. The best tool was accurate on 65 per cent of queries, the other two on 41 and 19 per cent, and hallucination ran between 17 and 33 per cent of responses. A financial-reporting benchmark isolates the residual more precisely: handed the exact correct filing, the strongest model of the day still got 15 per cent of answers wrong. Those tests ran in November 2023 and the levels are stale; the shape is not. Adding the right document narrows the error and does not remove it.

ConditionWhat was measuredWhat was left over
Retrieval over a curated, licensed, authoritative corpusLegal research tools, 202 queries, 2025: best tool accurate on 65% of queries17% to 33% of responses hallucinated
Handed the single correct source documentFinancial questions, November 2023: 85% correct15% wrong, with the right document open
General web search augmentationAccuracy 28.6% rising to 75.6%About 24% still wrong
Search augmentation on slow-changing facts12.0% rising to 59.2%Four answers in ten still wrong

That last row is the one that matters here. Slow-changing is precisely the category "who distributes this product in that country, and do they still hold the licence" belongs to.

But the residual error rate is not the real problem. The real problem is that these systems do not get quieter when they do not know. The benchmark that measured buried-fact retrieval also measured calibration, and the system that solved 51.5 per cent of those questions had a calibration error of 91 per cent: it was confident almost exactly when it should not have been. Researchers at Columbia's Tow Center ran 1,600 queries through eight AI search tools in early 2025, handing each a verbatim excerpt and asking it to name the publisher, headline and URL — the easiest possible version of attribution. The tools were collectively wrong on more than 60 per cent. One misidentified 134 of 200 articles, signalled low confidence fifteen times out of two hundred, and never once declined to answer.

What that does to competent professionals has also been measured. In a study of 758 consultants at a large strategy firm, published in Organization Science in March 2026, AI access on tasks inside the model's competence produced 12.2 per cent more work completed, 25.1 per cent faster, with quality up about a third. On one task deliberately built to sit outside it, the unaided group was correct 84.5 per cent of the time and the AI-assisted groups 70.6 and 60 per cent. The participants could not tell which side of the line they were on.

The specific problem with a generated list

A distributor search produces a table, and a table has cells. When the underlying research cannot establish whether a company holds the licence it needs, the honest output is a blank and a note. The output of a generative system is a sentence, in the same register as the cells that are correct. The absence of information and the presence of plausible information are rendered identically, and the reader cannot tell them apart. A better model does not fix this, because the thing being misreported is the system's reach, not its accuracy.

3. The information nobody publishes, in any language

A good deal of what a distributor search needs does not exist in machine-readable form, and in many markets not in any public form. The constraint is statutory and structural, and it does not move when the model improves.

What is missingMeasured positionSource and date
The rules themselvesOf 54 economies assessed, 25 give free electronic access to all legislation, 19 publish only the main laws and not the ministerial decrees where distribution and agency conditions live, and 10 provide no access at all. Nothing is published by 20% of the Middle East and North Africa economies in the sample, or 33% of the Sub-Saharan African onesWorld Bank Business Ready, December 2024; described as preliminary
Who owns the companyOf 38 Latin American and Caribbean jurisdictions, 28 have a beneficial ownership registry. One is public, five are partly accessible, 22 are closed to anyone who is not a competent authority. Globally, only 9% of assessed countries met the effectiveness requirements on beneficial ownership while registries existed in essentially all of themGlobal Financial Integrity, July 2025; Financial Action Task Force, April 2022
And not only in emerging marketsOn access to official company data the EU averaged 40 out of 100 and the United States 31. Austria and Spain scored zero. Estonia offered full access without a share-alike licence for 286,000 eurosOpenCorporates, 2020; may have moved since
Who is actually importingCustoms manifest data answers this well for the United States, India, Vietnam, Indonesia, Turkey, Mexico and most of Latin America. One major provider's published coverage lists exactly one African country, Cameroon. China is not covered, and the entire Gulf is absent, as are Nigeria, Kenya, South Africa and EgyptProvider's own coverage statement, read November 2026
The economy in no databaseInformal employment is 57.8% of total employment worldwide, from 12.9% in high-income countries to 81.5% in lower-middle-income and 88.3% in low-income onesInternational Labour Organization, May 2025

One note on reading that table: the figure you would most want, the share of firms in a given market with any web presence, is not reliably published anywhere. The United Nations Conference on Trade and Development said as much in June 2024 and convened a task group because the statistics are inadequate.

4. What a distributor will not tell a form

Set aside everything merely unpublished. There is a second category: information that exists, that someone knows, and that the person who knows it has a direct interest in your not knowing.

The best-documented study of international distribution remains a two-year field investigation of eight corporations which between them entered almost 250 new country-markets, published in Harvard Business Review in 2000. Its most useful case concerns a global hydraulics components leader which did not know that its Hong Kong distributor was selling nearly half its output into mainland China at roughly half the Hong Kong price. When the manufacturer took the market over directly, it inherited price-reduction demands, parallel imports and service complaints from Chinese customers it had never known it had. No register, filing, website or customs feed would have shown that, because the distributor's whole interest lay in it not showing.

The same study identifies a selection error no dataset can surface. Manufacturers choose for market fit, meaning a distributor who already sells to the right customers, and those distributors tend to represent the market's status quo, frequently carry competing lines and resist repositioning. Because many of those entries were reactions to distributors who approached the manufacturer, the paper's conclusion is blunt: the most eager distributors are often the wrong partners. Enthusiasm is the one attribute an inbound process selects for and the one that predicts nothing.

Where conduct risk is sharpest, the mechanism is margin. In May 2023 the Securities and Exchange Commission announced that Philips had agreed to pay more than 62 million dollars over sales of medical equipment in China. In the Commission's own description, special price discounts to distributors created a risk that excessive distributor margins could fund improper payments to government employees, and sub-dealers prepared sham bids using other manufacturers' products to create the appearance of competition. The Department of Justice and the SEC list "unreasonably large discounts to third-party distributors" among the red flags in their FCPA resource guide. A distributor's margin is visible to you; what happens to it two layers down is not, and no public-record screening reaches the sub-dealer.

Then the most mundane obstacle of all. A field experiment published in the American Journal of Political Science in 2025 sent 15,407 email solicitations to 5,140 banks worldwide. Among institutions in developing countries, 79.65 per cent never replied at all. An unsolicited remote approach in these markets has a low ceiling however well written, and that ceiling belongs to the recipient, not the sender. The fieldwork ran from 2019 to 2020, so levels may have shifted.

5. Who does the verifying decides what gets found

This is the strongest argument here and the least obvious, and it needs no claims about relationships or human touch. Verification is corruptible, the pattern is measurable, and an automated screen sits at the worst end of it.

The evidence comes from supply-chain auditing, where the same inspection is performed under different arrangements often enough to compare. A study of 16,795 code-of-conduct audits covering 5,819 factories in 66 countries found that audits paid for by the supplier turned up about 8 per cent fewer violations than audits paid for by the buying brand, and that audits including an auditor who had previously audited the same supplier found about 4 per cent fewer. Who writes the cheque changes the finding.

A 2026 working paper took this further using more than 10,000 audits of 2,313 factories across 34 countries for a single global brand. Auditors employed by the buyer recorded 9.5 per cent more violations than third-party auditors; forced labour was cited in 3.7 per cent of their audits against 1.5 per cent, nearly two and a half times the rate. And the finding that matters most: each unit increase in regional reported bribery was associated with 0.16 fewer violations cited by third-party auditors, while buyer-employed auditors showed no such effect. Outsourced verification degrades precisely where the institutional environment is weakest, which is precisely where a distributor search most needs it to hold. The same paper found unannounced audits citing 0.89 more violations than pre-announced ones. It is a working paper, not yet peer reviewed, and it uses violations cited as its measure of audit quality.

That this is causal rather than correlational was established by a randomised field experiment in Gujarat, published in the Quarterly Journal of Economics in 2013. Under the status quo the third-party audit system was, in the authors' description, largely corrupted: auditors systematically reported plant emissions just below the regulatory threshold although true emissions were typically higher. Changing their incentives made them report more truthfully, and the plants then cut actual pollution.

What this evidence says about automation

Read the list of conditions that degrade a verification: the subject pays, the inspection is announced, the inspector has seen this subject before, and the institutional environment is weak. An automated screen is the limiting case of every one of those at once. It reads what the subject chose to publish, at a moment the subject controls, through sources the subject can shape, and it degrades hardest in exactly the markets where the published record is thinnest. This is not a claim that software is dishonest. It is a claim about what a verification sees when nobody is in the room, and the effect sizes have been measured four separate ways.

And one corollary worth holding on to: resistance to inspection is itself a signal, and it is a signal that only exists once somebody asks to inspect.

6. What meeting in person actually buys, and what it does not

Here the conventional wisdom in this industry is simply wrong, and saying so is more useful than repeating it. The claim that in-person negotiation produces better commercial terms is not supported by the experimental evidence, and anyone building a case on it is exposed.

The foundational experiment, published in 2000, had 134 business students run a simulated wage negotiation either able to see each other or not. The final wage was 10.56 against 10.53, p = 0.47. Two more recent preregistered studies reach the same place with better methods: one comparing video against synchronous text found no effect on integrative value, with a Bayes factor of 11.31 in favour of the null; another, with 187 participants in 2024, found no significant difference in individual profit at all.

What richer contact reliably changesWhat it does not change
Mutual cooperation: 64% against 20%Final negotiated price or wage: 10.56 against 10.53, p = 0.47
Time to agreement: 8.21 days against 12.90Integrative value created: null, Bayes factor 11.31 for no effect
Explicit cooperative signalling: up to 90% against 13.8%Individual profit: no significant difference, p = 0.698
Information volume: 870 words against 280

The authors of the video study conclude that the benefits may be primarily perceptual, and that concession belongs in any honest account. But look at what does move. Cooperation, speed, signalling and disclosure, and disclosure is the commodity a distributor search is actually buying. The three facts that decide an appointment, which are willingness to take the line on, the price level the partner will really work at, and appetite for the investment the deal needs, are disclosed rather than discovered. They are not attributes of the company waiting to be looked up. They are positions the company adopts in a conversation, and they change depending on who is asking and what else is on the table that quarter.

One further finding sets the honest limit on all of this. Research on 399 buyer and supplier exchanges in China found that relational reliability did not substitute for legal enforceability against partner-conduct risk. Good relationships do not solve everything, and a piece arguing for human access should say where human access runs out.

7. The division of labour the evidence supports

The most instructive measurement of human and machine working together comes from medical literature screening, and it reads like this trade's problem in miniature. Researchers ran 23,521 articles through a model. Alone, it achieved recall of 0.92 and precision of 0.45; with a human reviewing its output, recall was the same 0.92 and precision was 1.00.

Read that carefully: the machine found nothing extra at the recall end. What the human added was precision, eliminating the 55 per cent of flagged items that were wrong, while the model cut 1,800 candidates to 274 and so reduced manual work by about 80 per cent. The model makes the work affordable and the human makes it correct, and those are different jobs rather than degrees of the same job.

Mapped onto a distributor search, the division falls out naturally.

StageWhat it involvesWho does it
Long-listCandidates from registers, trade data, directories and the open web; entity matching across scripts; sanctions and ownership screening; translationMachine, and paying people to do this by hand is indefensible
Medium-listCapability, licences actually held, channel reach, and current portfolio conflictsThe machine narrows it and a person starts asking, because the competing line a distributor carries quietly is the one that matters and the one it does not advertise
ShortlistWillingness, indicative price levels, appetite for the deal, and what the company will actually commit toHuman only. On our own engagements the team speaks directly to companies from the medium-list stage onwards, and each shortlisted company is approached discreetly and anonymously before anything reaches a client

Which is why a shortlist, done properly, is a set of pre-qualified counterparties prepared to engage rather than a list of names to investigate.

If that sounds like special pleading, consider who else prices it the same way. The United States Commercial Service sells two tiers of background report on a named foreign company. The cheaper tier, at 1,050 dollars per market, is explicitly "a general background report on a specific foreign company based on publicly available information". The dearer tier, at 2,000 dollars, adds three things: references contacted by the agency, a site visit, and interviews with the principals. Transport beyond 80 kilometres from one of its offices costs extra again. A government body has priced the gap between what is discoverable and what is not at roughly 90 per cent plus mileage, and the three things on the far side of that gap are the three a model cannot do. Those fees were published as of November 2026 and the agency notes its schedule is being updated. The same agency will separately introduce you to five interested distributors, and claims no vetting at all in doing so. Cheap introductions are widely available; qualified counterparties are a different product.

8. Six questions to ask of any AI-assisted distributor search

Whether you are buying this work or building it, these separate a search from a generated list. They are deliberately answerable.

1. Show me a cell where the answer was "we could not establish this"

If every field in the deliverable is populated, either the market was unusually transparent or something filled the gaps. A provider who cannot produce one gap is telling you their process has no way of reporting its own reach.

2. Which of these companies did you speak to, and who did you speak to there?

Not "contacted". Spoke to. The difference between a company that exists and a company that will take your line on is established in a conversation and nowhere else.

3. What is the primary source for each regulatory claim, and on what date was it read?

Regulators get merged, renamed and stripped of functions, and the web lags by years. A claim about a licensing regime with no dated primary source is a claim about what was true at some point.

4. What competing lines does each candidate carry?

This is among the most common reasons an appointment underperforms and among the least likely facts to be published. If the answer came from the candidate's own website, it is the answer they want you to have.

5. Which of these would you drop, and why?

A list is not a recommendation. Someone who has done the work holds a view about which candidates are weakest and can say what would change their mind. A generated list has no view, and it will not tell you that it has no view.

Questions we get asked

Does using AI make a distributor search cheaper?

It makes the list-building stage considerably cheaper and faster, and that stage is a real share of the work. It does not reduce the cost of the shortlist stage, because that cost is people talking to companies and the number of conversations does not fall. A quote that is cheap because the whole process is automated is quoting for a different deliverable.

Can AI tell me whether a distributor is financially sound?

In some markets yes, where filings are published and current. In many it cannot, because the accounts are not public, not filed, or not meaningful. The useful question is not whether a provider can produce a view, but whether they will tell you which of those three situations you are in.

How should AI be described in a research methodology, honestly?

As an efficiency and coverage advantage at the stages where it has been measured to deliver one, with the stages it does not touch named explicitly. Our own position is that the system cuts time and cost on search routes and pattern-matching, that all work is human-led, and that every candidate is contacted directly before anything reaches a client. Any description that cannot say where the automation stops is not a methodology.

Need a shortlist you can act on rather than a list of names?

We run distributor search and selection across 40+ countries and 100+ product markets, for companies from Fortune 500 groups to businesses exporting for the first time. Tell us the market and the sector and we will tell you what can be established from the record and what will take conversations.