The second reader
AI tools can now draft the codes that decide what a hospital is paid, and the payers who rely on those codes need a check that sits outside the tool. Beamtree already sells that check, and its small balance sheet is the main obstacle to the long contracts a checker needs.
A hospital coder reads each discharge summary and turns it into the diagnosis and procedure codes that decide how much the hospital is paid for that patient. The work is slow and skilled, and AI tools can now draft the same codes in seconds. Whichever tool drafts the codes, the health department or insurer paying on them needs a way to know the codes are right, and the oldest answer to that problem is a second reader who works independently of the first.
What Beamtree does
Beamtree, an ASX-listed health software company, already sells that second reader. Its PICQ software checks coded hospital data against the rules of the Australian classification and flags the episodes most likely to be wrong, so a coder looks again at the doubtful cases and leaves the rest.
Beamtree runs three businesses, and the largest, Health Roundtable, is a benchmarking service with 189 member hospitals across Australia and New Zealand that brings in about 42% of revenue. The coding products (PICQ, RISQ and PICQ Audit) bring in about 30%, and RippleDown, a rules engine that helps pathology laboratories interpret test results, brings in about 28%. Beamtree reported revenue of A$29.2m for the 2026 financial year, and A$27.5m of it was recurring.
Why a second reader matters
Every AI coding tool sends some of its work back to a person, and that review bill decides whether the tool saves anything. An August 2026 paper from Scale AI tested 16 AI systems on 750 clinical audit cases drawn from real patient records, and two of the systems scored 72.8% and 72.5% on accuracy, a gap no buyer would notice. To reach the same reliability target of 76%, one of the two needed a human to review 39.2% of its answers and the other needed 29.6%, so across a thousand cases the first system sends about 96 more back to a clinician. The paper tested clinical audit judgements, and hospital coding raises the same question for any tool whose output a person has to check.
The saving from automated coding survives only while people review a small share of what the tool produces, and a tool with badly calibrated uncertainty pushes coders back into checking everything until the saving disappears. The number a hospital should ask any vendor for is the review rate at a stated reliability target, and a rules engine that clears the consistent episodes and routes the doubtful ones to a person lowers that number directly.
The checker also has to sit outside the tool, for the same reason that double-entry bookkeeping and the flight recorder on an airliner keep the party that performs an action away from the official record of it. An AI coding tool that assigns codes and then explains its own choices puts both jobs inside one system, and researchers testing AI agents found that agents in seven of the eight harnesses they tested would delete their own traces when asked, which shows how weak a self-written record can be.
A useful audit compares the tool's output with an independent reference, which for hospital coding means the clinical record and the coding standards. A rules engine applies the same written rules every time and records which rule fired on which episode, so payers and auditors can rerun the same rules on the same episodes and get the same answer.
What has to change
The checker becomes standard when a payer or auditor asks for it, and no regulator we found requires an independent check on AI-drafted codes yet. Regulators ask for transparency and human oversight and leave the method to the hospital, and a human coder signing off on each code may satisfy current rules, so until a payer or auditor asks for more the checker has to win on the review bill alone.
Beamtree is building both the drafting tool and the checker, and its recurring revenue from coding grew 21% in the first half of the 2026 financial year. In an Australian trial of its own AI-assisted coding, Beamtree reports that the AI agreed with a human coder on more than 95% of the episodes it coded at one site. In May 2026 Beamtree signed a A$2m, twelve-month coding and coding assurance contract with Fakeeh Care Group in Saudi Arabia, a country that codes hospital activity with the Australian classification, and a Canadian public hospital has also tendered for AI-assisted coding.
We give Beamtree a reception score of 14 out of 30 and give the independent checking layer itself 21 out of 30.
What could slow it
- Large coding vendors such as Solventum, Optum and Epic can build their own checking layer and bundle it with their coding tools.
- Large language models make rules cheaper to write, which lowers the value of a curated rule library.
- PICQ is built around the Australian classification, so its natural market is the health systems that use it, including Saudi Arabia and Singapore, where Beamtree has signed coding contracts.
- A rules engine misses the cases nobody wrote a rule for, and PICQ needs steady maintenance as coding standards change.
- Beamtree selling its own AI coder alongside PICQ puts the drafting tool and the checker inside one company, and PICQ makes the stronger independence case as the check on other vendors' tools.
- Beamtree is a small company, with A$2.5m in cash at 30 June 2026, a further A$3.0m of undrawn debt facility and an operating profit of A$0.4m for the year, and a health department choosing a checker it will rely on for years weighs the supplier's size before it signs.
What to watch
- We predict Beamtree will announce a new paid contract for AI-assisted coding by February 2027, and the Canadian tender is one candidate.
- A partnership that puts PICQ behind a larger vendor's coding tool would put the checker in front of far more hospitals.
- Coding recurring revenue needs to keep growing at the pace of the first half of the 2026 financial year.
- A payer or auditor in Australia or Saudi Arabia requiring an independent check on AI-drafted codes would be the strongest single signal for the checking layer.
What would change our view
We would raise the score if a payer or auditor started requiring an independent check on AI-drafted codes, or if Beamtree signed a partnership that puts PICQ behind another vendor's coding tool. We would lower the score if a large vendor bundled an equivalent checker into its own product.