How we're different
Every HEOR, market access and regulatory consultant has had the same week. A dossier section due Friday. A model that could draft it in twenty minutes. And a sponsor agreement that says the data in front of you doesn't go anywhere near it.
So you do what everyone does. You work around it.
Find-and-replace the drug name. Swap the sponsor for “Company A.” Round the hazard ratio. Delete the site names. Then paste the rest and hope the model can still do something with it.
What it costs you: Forty minutes of redaction for twenty minutes of drafting. And the model, staring at “Company A” and “[REDACTED],” writes you something generic enough to be useless — or worse, confidently fills the gap with something that isn't true. You then spend another hour putting the real values back and checking that it didn't quietly reason about the wrong ones.
The part nobody says out loud: one missed value in one prompt, one time, and the redaction wasn't a control. It was a habit.
The licence says the vendor doesn't train on your data. Procurement signed something. Your firm considers the matter closed.
What it costs you: Nothing — until a client asks. Because that agreement is between your firm and the AI vendor. Your obligation is between your firm and your client. The second is not discharged by the first. Your sponsor did not sign that DPA. They were not asked. And a confidentiality clause that says “shall not be disclosed to any third party” does not contain an exception for third parties who have promised to be careful.
The data still goes to the model. Every value. In full. The protection is a promise about what happens after it arrives.
Some sponsors have their own walled AI. Eventually you'll be provisioned. Eventually there'll be training. Eventually.
What it costs you: The engagement is ten weeks long and the access request is in week six. Meanwhile you have three other clients, and none of their environments talk to each other. You are not going to run your practice inside five different sponsors' sandboxes.
Literature summaries. Generic method text. The introduction nobody reads. Anything that touches the actual engagement stays off it entirely.
What it costs you: The AI helps with the part that was already easy. The confidential 80% — the analysis, the argument, the numbers you were hired for — is exactly the part it can't touch. You have automated your least valuable hour.
Not out of recklessness. Out of arithmetic. The deadline did not move when the policy arrived.
What it costs you: It moves off the firm's laptop and onto your phone, where there is no log, no policy, no oversight and no one to ask. The exposure didn't go away. It just went where nobody can see it.
A ban doesn't stop AI use. It stops supervised AI use.
Four of the five ask you to give up capability to stay compliant. The fifth asks you to give up compliance to stay capable.
Nobody has ever offered you the option where you don't have to choose — because until now the only way to protect a confidential value was to remove it, and a model cannot reason about what isn't there. Blank it out and the model stalls. Swap in a meaningless code and it treats the code as a real thing and writes fluent nonsense around it. Both roads end in output you have to throw away.
That is why “just anonymise it first” has never actually worked for consulting deliverables. It isn't a discipline problem. It's a physics problem: protection and usefulness were the same dial, and it only turned one way.
Untraceable reads your document on your machine and finds every confidential value in it — the sponsor, the molecule, the indication, the site, the endpoint, the number that hasn't been published yet. Then it replaces each one with a cloak: a stand-in that carries everything the model needs to reason correctly, and nothing that could ever identify anything.
Not a blank. Not a hash. A meaning.
The model reads a study of a drug, in an indication, at an institution, with a hazard ratio and a confidence interval. It understands the structure, the relationships, the clinical logic, the statistics. It drafts as well as it drafts on anything.
It has never seen a single one of your values. Neither have we. When the answer comes back, it restores on your screen — and only on your screen.
The study evaluated Pembrolizumab in patients with non-small cell lung cancer at Memorial Sloan Kettering. The primary endpoint showed a hazard ratio of 0.68 (95% CI: 0.51–0.89).
The study evaluated NSSA_Drug_1 in patients with NSSA_Indication_1 at NSSA_InstitutionalInvestigator. The primary endpoint showed a hazard ratio of SSA_OSHR_POB_SMCID_KVM_CLIN_054 (95% CI: SLC_004431_012–SLC_004532_054).
Roughly 95% of the model's natural writing and reasoning quality survives the cloak. That isn't a claim we'd like you to take on faith — it's a preprint, with the method, the task set and the numbers.
| Approach | Real values reach the model | Output quality you can actually use | Satisfies your client's NDA | Audit trail if you're asked |
|---|---|---|---|---|
| Redact by hand | No | Poor — the model reasons around holes | Only if you never miss one | None |
| Enterprise AI seat | Yes, all of them | Excellent | No — that's your vendor's promise, not your client's consent | Vendor logs, not yours |
| The client's own environment | Yes | Good | Yes — for that one client | Theirs |
| AI on the safe 20% only | No | Good, on the work that didn't need it | Yes | N/A |
| Ban it (and use it anyway) | Yes, invisibly | Excellent | No | None, by design |
| Untraceable | Never | ~95% of unrestricted | Yes — the value was never disclosed | Every value, every cloak, every submission |
Look down that last column. In four of the five workarounds, the honest answer to “can you show me what your AI saw?” is no. That question is going to start being asked.
We're not a chatbot with a compliance badge. If what you need is a general assistant to answer general questions, you already have one, and it's cheaper than us.
We're not a consulting firm. We never see your data, so we can't do your engagement. We build the tools; you do the work. Where you'd rather not run it alone, we partner with consultancies in your field.
We won't help you with public data. If nothing in the document is confidential, cloak nothing and use whatever you like. We're for the other 80%.
And we can't sign your NDA for you. What we can do is make sure there was nothing to disclose in the first place — and give you the audit trail that proves it.
Untraceable was built by an HEOR consultant with twenty years in the field — someone who has written the dossier, sat in the payer meeting, and had the conversation with a sponsor about what may and may not leave the room.
That's why this isn't an API you'd have to build a workflow around, or a security tool your CISO buys to stop you working. It's the tool the work actually needed: drafting that knows what a value dossier is, statistical support that knows what an SAP is, models that know the difference between what CDA-AMC wants and what G-BA wants — all of it running on data the AI is never allowed to see.