How Atlas works, and why

The job

If you work in business development, you know the slide. Who else is in this indication, at what phase, with what status, enrolling how many people, finishing when. It goes into the partner memo, the licensing deck, the diligence pack. Then someone across the table points at one row and asks where it came from.

That question is the job. A landscape is only as good as its weakest row, and most of the time spent building one goes into making every row defensible: finding the trials, reconciling the registries, checking each date and status against the record, and keeping a trail you can show.

Atlas does that part. It builds the landscape from three public registries (ClinicalTrials.gov, the EU Clinical Trials Information System and the EU Clinical Trials Register) and ties every value in every cell to the record it came from.

You already have your own AI

I think most people in BD already have an AI assistant they use every day, whether it is the one their firm has approved or two or three of their own. And each of them has a way of working that suits them: their own prompts, their own templates, their own way of turning a table into a story for a particular deal.

I do not think a trial data tool should replace that with its own idea of the answer. A pre-written summary is someone else's analysis, done for a question that is not quite yours, with the evidence out of sight. You end up checking it, rewriting it, or both.

So Atlas gives you something different: the landscape as clean, structured data. One row per study. Registry identifiers first. Phase and status, sponsor, conditions, interventions, enrolment and dates in their own columns. Next to every value, a source line naming the registry, the record, the field it was read from, a link and a date. Gaps are labelled instead of left blank, and rows where the registries disagree say so.

That is the kind of input an AI model does well with. Export it as CSV or Excel and take it into whatever you already use. Ask your assistant to draft the competitive section of the memo, or compare two sponsors' timelines. Whatever it writes, the source lines are still there, so you can check any sentence against the record it rests on.

The part of the work that is yours stays yours. The thesis, the comparators you choose, your read on a competitor: that happens in your tools, under your firm's rules.

Why not just ask a general assistant

You can, and for a first look it is quick. But a general assistant asked to find every trial in an indication is working from whatever it happens to retrieve. It cannot tell you what it missed, it will not show you where the two registries disagree, and the count it gives you today may not be the count it gives you next week.

Finding every registered study, keeping each registry's statement, and labelling what is missing is retrieval and bookkeeping. That should be done the same way every time, and shown. Atlas does that step. Your assistant does the writing, on data you can trace.

Where Atlas uses AI, and where it does not

Atlas uses AI in one place. When you describe the landscape you want, Atlas chat turns your words into criteria: conditions, interventions, phases, statuses, sponsors, dates. You see every criterion, you can change any of them or set them yourself, and you can undo each change chat makes.

Chat reads your request and the release's criteria. It never reads study records, so it has nothing to summarize and no way to put a value in a cell. Building, filtering and exporting the dataset run no AI model at all. The studies in your table are the ones that match the criteria, and the values in it are what the registries say.

Why disagreements stay visible

Registries often disagree about the same trial. In 2018, Jessica Fleminger and Ben Goldacre compared trials registered in both ClinicalTrials.gov and the EU Clinical Trials Register. Of the 10,492 trials they could compare, 16.2% gave a different completion status in the two registers (PLOS ONE, 2018).

A tool that picks one record hides that. Take NCT05104866, the phase 3 study on our home page. As of 6 October 2026, ClinicalTrials.gov listed it as active, not recruiting, while its Netherlands record in the EU Clinical Trials Register said completed. If your landscape shows only one of those, neither you nor the person reading your memo knows the other exists.

Atlas shows each registry's statement as that record gives it and marks where they differ. It does not merge them or choose a winner. You decide which record your landscape relies on, and you can say why.

Why "not collected yet" and "not mapped" are labelled

A registry saying nothing does not mean nothing exists. A blank field can mean the sponsor did not report it, the registry does not ask for it, or Atlas has not collected that field yet. Those are different facts, and a blank cell makes them look the same, to you and to any model you hand the file to.

So Atlas labels them. A field the release has not collected is marked "not collected yet". A value Atlas could not map to its criteria is marked "not mapped". A value the registry does not give is marked as missing at source. A gap you can see is a gap you can footnote.

Why no rankings or probability of success

Ranking programmes against each other, or putting a probability of success on one, means making judgments: which endpoints matter, how to weigh a small study against a large one, what a delay signals. Those judgments belong to the person whose name is on the memo.

Atlas gives you the cited facts and stops there. It does not rank, score, predict or advise. You make the call, with whatever tools you trust, and you can show the records you made it from.

How it works

1. Ask. Describe the landscape in plain words, the way you would brief an analyst.
2. Check the criteria. Atlas shows the criteria it took from your request. Change any of them, or set them yourself. Atlas is set up to ask you a question only when the answer would change which studies are included.
3. Build. Atlas builds the dataset from one published reference release, as of that release's data date.
4. Open any cell's source. Every value opens the record it came from, with the field and the date it was retrieved, and links to that record on the registry's own site.
5. Reshape. Choose columns, sort, group by phase and filter by enrolment without starting again.
6. Save or branch. A saved dataset keeps its earlier revisions in its history, and undo adds a new revision rather than rewriting an old one. Save a copy to try a different cut; the copy starts its own history. You can also share a read-only link that opens without an account, and whoever opens it can make their own copy.
7. Export with citations. Download CSV or Excel, with the source of each value beside it, and take it into your own tools.
8. Refresh. When a newer reference release is published, Atlas shows which studies would join or leave a saved dataset before you update it. Your criteria carry over, and earlier revisions keep their own data in the history. An update cannot be undone, so save a copy first if you want to keep working on the earlier data.

What Atlas does not cover yet

Atlas covers trial registrations. It does not include trial results, deals or conference abstracts. If your landscape needs outcome data, licensing terms or what was presented at a congress, Atlas will not give it to you today. I would rather you read that here than find it out halfway through a project.

Where Atlas sits

General AI assistants are quick and flexible, and you should keep using yours. What they do not give you on their own is a row you can trace to a record, a count you can reproduce next week, or a list of what was left out.

Licensed intelligence platforms have depth that Atlas does not have: results, deals, analyst commentary. They are usually sold to whole teams on annual contracts, and what you read there is often their analysts' account of a trial rather than the registry record itself.

Atlas is narrower than both on purpose. It answers one question, who else is registered in this space and where they stand, from the public record, as data you can take anywhere and trace back. If something in a landscape looks wrong to you, I want to hear about it: write to us.

Tamal