Most of the money chasing “AI in healthcare” is looking in the wrong room. It’s focused on the software that reads the scans and the gadgets that give you the results, while the machine learning that really changes the economics is working in the lab downstairs, deciding what tests can be built. The gap between where AI is and where cash is being directed is exactly the kind of market that is mispricing. And in one of the unpopular areas of medical diagnostics, mispricing is compounded by a second disruption unrelated to AI. It’s a quiet schism in the way tests are provided to the people who use these tests.
Saliva Diagnostics expanded two sales channels from the same saliva sample. These are home tests that are sold in pharmacies and supermarkets, or that anyone can buy off the shelf. The other is through a bulk purchasing system that supplies boxes by the box to hospitals, ministries of health, and poorer countries. These two routes set price risk in opposite directions, and small groups of publicly traded companies are located along these routes, rarely in the same location or to the same degree. The market tends to treat lots as one bet. A side-by-side reading of the published documents clearly shows that this is not the case.
Where is AI actually located?
Almost everyone starts with the part that goes back. Lateral flow tests, the humble piece of paper behind home pregnancy and HIV test results, are not the place for smart machine learning. A difficult problem in calculating saliva is noise. The samples are messy and full of biological static, and it takes real horsepower to extract a clean signal from them. December 2025 Review Published in the Journal Clinica Chimica Acta,title “Revolution in salivary biomarkers through machine learning and artificial intelligence” Saliva data is high-dimensional and complex, so machine learning is required to derive usable patterns from it. That’s the upstream job. It determines which tests can exist. The final strip you spit out remains a simple strip.
Dr. Omar Deutsch, co-founder and CEO of the Israeli company Salignostics, drew the same line unprompted in an interview that preceded the Gates Foundation’s announcement. When asked where AI is integrated into the company’s workflow, he was candid. “Yes, it’s not a basic lateral flow assay, but primarily the biomarker discovery and data analysis aspects.” Once the noise reduction layer is applied, “we combine proteomic and genomic characterization with AI-based analysis tools to identify biomarker clusters associated with specific physiological or pathological conditions,” he said. Two of the current studies are in pancreatic cancer and Alzheimer’s disease, neither of which are part of the grant-funded research. These are research, not products, and work on their own clocks.
For those evaluating these companies, it’s important to keep the two tiers separate. Abingdon Health, a York-based contract manufacturer in London’s AIM market, has its own AI tool, a smartphone reader called AppDx, that interprets lines of completed tests. It’s a downstream AI that reads the results after the fact, and it’s a separate animal from the upstream animal that looks for new biomarkers. Same buzzword, both ends of the pipe. Treating them as one would misjudge where the defensible edge actually lies.
Two roads from the same sample
The second confusion concerns plumbing rather than technology. Consider what moves on each route. The path to consumerism is well-trodden. The saliva pregnancy test, sold as Salistick, is currently distributed in the UK through U-TEST, with a set price per unit sold to anyone who walks into the store. That’s something road analysts who follow this group already understand.
The organizational path is the path they often miss, and it’s no different than retail. In poorer countries, buyers often buy through collective procurement, a government- or donor-funded program and a system in which many countries band together to negotiate one large price. of Global Fund Purchasing Platform And agencies like Unitaid and PEPFAR are advancing diagnoses by the million based on that. The grant that Deutsch discussed, the $1.87 million Gates Award the company plans to announce for its HIV and Pregnancy Complications Panel, points to this second path rather than the first. He candidly said that route hasn’t been determined yet, saying, “There are several commercial routes in mind…and we’re not currently focused on any particular structure or channel,” adding that adoption will likely be led by KOLs, key opinion leaders who will get the word out to the clinical field ahead of a broader rollout.
Why is it useful to differentiate? Shelf revenue rises and falls depending on shoppers. Procurement revenues rise and fall depending on donors’ budgets, and those budgets have been hit hard. of Latest information on UNAIDS 2025 It estimated that there would be 40.8 million people living with HIV and 1.3 million new infections in 2024, and warned that the number of people infected could increase by 6 million by 2029 due to sudden funding cuts. Orasure Technologies, the only maker of commercially available oral HIV self-tests in the United States, is already in trouble. The company’s sales of diagnostic drugs in the third quarter of 2025 were down 34% compared to the same period last year, and the decline was certain. Sales of HIV tests are sluggish amidst unstable government funding. Companies that lean toward the institutional path are exposed to political weather that supermarkets never have.
Why does the gap remain open?
Knowing the breakup is one thing. The more difficult question is why it persists. Because the edge is usually arbitraged away the moment someone notices it. The answer is structural. Yakov Amihud’s highly cited 2002 paper About illiquidity and stock returns We explain why lightly traded stocks tend to carry an extra return premium and why information permeates stock prices quickly. Smaller companies that are not followed by most analysts continue to welcome mispricing. Because there won’t be an army of researchers to discover it, and there won’t be a ton of deals to fix it. AIM’s microcaps meet all of these criteria: no broker coverage, low trading volume, and none of the appeal of an AI fund.
The name confirms it. Abingdon’s FY25 sales rose 40% to £8.6m, with a much stronger second half, and the company £3.2 million in October standings Smaller companies such as Australia’s Lumos Diagnostics and US-listed Biomerica occupy the neighborhood point-of-care niche, while larger players such as Abbott, Hologic and QuidelOrtho trade with deep liquidity and analyst coverage that the smaller players would envy. The cause of the discrepancy is not that the numbers are difficult to read. The Abingdon results will be made public, the UNAIDS figures will be made public, and the procurement amounts will be made public. Few people will have a hard time reading these side by side.
History describes in two rhymes what happens when number crunching finally cracks the market that smart money was crushing. Weather derivatives have long been dismissed as unmodelable, based on the logic that it is not possible to put a price on next winter’s temperatures until statistical forecasts are well established and a real market is built. CME. Subprime auto lending followed the same course, being deemed too opaque to be cleanly underwritten until machine learning credit models learned to distinguish between good and bad borrowers. In both cases, the edges were modest in size and turned out to be oddly long-lived. The main reason is that it remained out of fashion long after the area ceased to be inefficient.
Things that can still go wrong
There are a number of things that could potentially overcome this situation, and they are worth naming. The grant-funded rollout moves at a bureaucratic pace, potentially delaying the rollout for years between announcement and the time someone changes hands on a single test. Procurement could bypass listed agents entirely, such as by purchasing directly from manufacturers or private partners, and Deutsch himself emphasized that the channel is undetermined and that this is a two-way disconnect. The same openness that creates mispricing also means that listing agents may not make any profit at all. Microcaps can become cockroach motels, easy to check in but difficult to leave once capacity is exhausted. And the Gates-backed panel may never reach commercial scale. This is because many subsidy-funded diagnostic drugs fail to reach the market even after passing testing. The uncertainties Deutsch has stated on the record do not undermine the debate. That is the discussion.
conclusion
This is not the time to buy specific small-cap stocks. A helpful habit when sizing diagnostic minnows touted as AI plays is to ask two simple questions. It’s about whether that revenue goes to shoppers or donors, and whether that AI actually lives in the lab or at the label. The market continues to lump these companies together and give them the wrong answer, and when you read the filings side-by-side, it’s clear that the two companies are separate companies. The edge remains untouched not because the data is tough, but because the space is so dull that few investors are giving it a second look.
