01
Connect
A read-only connection to the systems the site already runs. No new workflow, no data leaving the site's control, no migration project. Structured fields and unstructured narrative both come across.
Clinical AI for trial enrolment
Every eligible patient, found before the window closes.
Accrual’s clinical AI reads your patient population against every open protocol, continuously — and tells sponsors which sites will actually enrol, from real population data rather than the site’s own estimate.
Patient recordAccrual AINCT-04 • Phase III
Unstructured chart
ONCOLOGY PROGRESS NOTE — 22 APR
63F presenting for follow-up. Ambulatory and
self-caring, ECOG 1. Denies new symptoms.
PATHOLOGY REPORT — 14 MAR
Sections show invasive carcinoma, findings
consistent with stage IIIB disease.
CT CHEST W/ CONTRAST — 02 MAY
Target lesion in right upper lobe measures
3.1 cm, measurable per RECIST v1.1.
CHEMISTRY PANEL — 28 APR
Creatinine 0.9 mg/dL. Calculated CrCl
78 mL/min. Hepatic panel within limits.
Eligibility criteria
0/7
Histologically confirmed stage III–IV disease
Pathology · 14 Mar — “…consistent with stage IIIB…”
ECOG performance status 0–1
Oncology note · 22 Apr — “ECOG 1, ambulatory”
Measurable lesion per RECIST v1.1
CT chest · 02 May — “target lesion 3.1 cm”
No prior checkpoint inhibitor therapy
Medication history — no PD-1/PD-L1 agent recorded
Adequate renal function, CrCl ≥ 60 mL/min
Chemistry panel · 28 Apr — CrCl 78 mL/min
No active autoimmune disease requiring systemic therapy
Rheumatology note · 2021 — needs coordinator review
Life expectancy ≥ 12 weeks per investigator
Investigator assessment required at screening
VerdictSCREENING…
Synthetic record and protocol. Illustration only.
Screening continuously against open protocols
Oncology · Phase IIICardiology · Phase IIRare disease · Phase I/IINeurology · Phase IIIImmunology · Phase IIMetabolic · Phase IIIRespiratory · Phase IIHaematology · Phase INephrology · Phase IIIOncology · Phase IIICardiology · Phase IIRare disease · Phase I/IINeurology · Phase IIIImmunology · Phase IIMetabolic · Phase IIIRespiratory · Phase IIHaematology · Phase INephrology · Phase III
The gap
Eligible patients are sitting in records nobody searched.
Eligibility criteria arrive as dense prose. Patient records are unstructured narrative. Between the two sits a manual job that nobody has the hours to do properly — so it does not get done, and the timeline slips.
80%
of clinical trials miss their enrolment timelines
Industry estimate
$600k–8M
lost to the sponsor for every day a trial runs late
Industry estimate, per day of delay
1 in 3
sites in a typical study enrol one patient or none
Industry estimate
Figures are widely cited industry estimates, included to describe the problem Accrual addresses. They are not Accrual’s own performance data.
Two sides, one read
The same read of the population serves both sides of the trial.
Finding a patient for a site and forecasting enrolment for a sponsor are the same computation, run once. That is why the forecast is evidence rather than an opinion — and why sites get it for nothing.
For sites
Your population, already screened.
Accrual's models read every record you hold against every protocol open to you, and keep reading as charts change. Matches arrive where the clinician already works — not in another portal nobody opens.
- AI screening that runs continuously, not a query run once at study start
- Every criterion cited back to the source line in the chart
- Surfaced at the point of care, in the existing workflow
- Free to the site — you are the data supply, not the customer
For sponsors
Which sites will actually enrol.
Feasibility today runs on a questionnaire the site fills in optimistically. Accrual answers it from the population itself: how many patients the AI can actually find, at what rate they can be enrolled, and what else is competing for them.
- Forecasts grounded in population evidence, not self-estimates
- Aggregate and de-identified — sponsors never receive patient data
- Coverage beyond the major academic centres
- Pay per qualified enrolled patient, or per feasibility study
How it works
From protocol prose to a ranked, evidenced patient list.
02
Read
Language models decompose protocol prose into machine-checkable criteria, then read every record against them — notes, pathology, imaging reports, labs, medication history. This is the step conventional search could never do, and the reason the eligible patients have stayed hidden.
03
Match
The model ranks candidates by strength of match and cites the evidence for every criterion, linked back to the source line in the chart. A coordinator confirms or dismisses in seconds, and that judgement trains the ranking.
04
Forecast
The same population read becomes the sponsor's feasibility answer: how many patients this site can realistically enrol, at what rate, with what competing studies in the way. Aggregate and de-identified, always.
“A site that says it can enrol forty patients and enrols two has not lied. Nobody ever counted.”
Site feasibility has run on self-report for thirty years because counting was impossible. It is not impossible any more — the population is right there, in the records the site already holds.
What changes when you count
Daily
Re-screening cadence — the population is re-read as records change, not once at study start
0
Patient records ever shared with a sponsor. Aggregates and forecasts only
Both
Structured fields and unstructured narrative read together — the eligible patients hide in the prose
Responsible clinical AI
An AI that shows its work, or it does not ship.
Clinical research is not a setting where a confident-sounding answer is good enough. Four constraints hold the models honest, and they are the reason a site's governance review says yes.
It cites, or it does not count
Every criterion the model resolves is linked to the exact line of the chart that resolved it. A determination without a citation is not shown, because a coordinator has to be able to check the work in seconds rather than trust a score.
It flags doubt instead of guessing
When a record cannot settle a criterion, the honest output is 'needs review' — not a confidence number dressed up as an answer. Overclaiming here produces screen failures, which is the exact waste the product exists to remove.
It proposes; a person decides
Accrual never determines eligibility. It assembles the evidence and puts a ranked, cited shortlist in front of the clinician and coordinator. Eligibility is decided by an investigator against the protocol, as it must be.
Your records do not train anyone's model
Patient data is processed to answer your questions and is never used to train foundation models, never pooled across sites, and never shared with a model provider for improvement. This is written into the site agreement, not just the policy.
Where the criteria get hard
Built for the protocols that structured data cannot answer.
Oncology
Dense biomarker and staging criteria buried in pathology narrative.
Cardiology
Ejection fraction and prior-event history that structured fields rarely carry.
Neurology
Scale scores and progression notes written as prose, never as codes.
Rare disease
Populations small enough that missing one eligible patient moves the timeline.
Immunology
Prior-therapy sequences and washout windows spread across years of chart.
Metabolic
Lab trajectories that matter more than any single value.
The structural shift
Diversity mandates moved the demand outside the academic centres.
Sponsors now have to enrol beyond the handful of major academic sites that have always absorbed trial volume. That means community hospitals and non-US sites — places with real patients and no research infrastructure to find them.
Those sites are exactly where a continuous, automatic read of the population is worth the most, because there is no coordinator with spare hours to do it by hand.
Request access
Tell us what you are trying to enrol.
We review every request by hand and reply with whether we can help, honestly. Sites and sponsors both start here.