What people actually ask.

Software that reads what arrives at your business, applies the criteria you actually work by, and hands you the decisions worth making.

A placement agent gets 10–50 inbound pitches a week and reads maybe six properly. Synthos reads all of them against his screens and returns the two worth his week — with a written reason for every decline he can send as-is. He still decides. He just stops spending Thursday on the ones that were never a fit.

Those are conversations. You ask, they answer, the session ends and nothing persists.

Synthos runs workflows against your own material on a schedule or a trigger, holds anything outward-facing for approval, and keeps a record of every decision. The models underneath are the same ones — what is different is the structure around them.

Most of those summarise or draft inside one product. Synthos works across whatever you point it at, applies rules you wrote, and stops at a person before anything leaves.

If your existing tools already do the job, use them. This is for work that spans several sources and where a wrong output reaches someone outside the business.

Nothing that reaches a person outside your business happens without your approval. Not as a setting you switch on — the system will not run a workflow where an outward step could be reached without passing a human gate.

Internal work runs on its own: reading, extracting, cross-checking, drafting. The boundary is whether anyone outside your organisation would see it.

It will sometimes. Language models misread documents, miss things, and occasionally state something confidently that is not true. That is a property of the technology, not a defect we expect to eliminate.

The platform is built on that assumption. A Verifier re-checks findings rather than trusting the agent that produced them, guardrails are deterministic rather than model-judged, and a person approves anything that leaves. Those controls reduce the risk. They do not remove it.

Yes, and doing so is how it learns. Accept something your own rules declined and the exception is recorded with your reason.

Next time a similar case comes up it is surfaced for your call rather than declined outright — never auto-approved. A system that learned “he takes pre-revenue biotech” and started acting on it would have replaced your judgement with a guess about your judgement.

You are, and the design assumes it. Every decision that reaches someone outside your business was approved by a named person at your firm, and the record shows who and when.

We are a software vendor. We are not a broker-dealer, an adviser, or a regulated intermediary, and where you use Synthos in regulated activity the supervisory obligation remains yours. That is stated plainly in the Terms.

In your own workspace. Every scoped record carries its workspace, and access is checked on the server for every request.

Our automated tests assert something specific: a record belonging to another workspace and a record that does not exist are indistinguishable to a caller who should not see it. Not filtered in the interface — refused at the data layer.

No, and neither do the model providers we route to. That is a contractual term with them, not a preference of ours.

We also do not read your workspace content for our own purposes. Support access to a specific workspace is a separate, logged action, and we ask first.

No. We do not currently hold SOC 2 or ISO 27001.

What is true today: isolation enforced at the data layer and asserted by automated tests, encryption in transit and at rest, credentials stored server-side and never returned to a browser, and an append-only record of what the system did.

Yes. You own what you put in and what comes out, and you can export at any time. On cancellation you keep export access for 30 days.

Different ones for different steps. Extraction and formatting are mechanical and go to fast, cheap models; judgement calls go to the strongest available.

That is not only about cost. A step that does not need a frontier model should not get one, and routing by what the work actually requires is easier to explain to a compliance officer than “we use AI.” On higher tiers you can bring your own provider keys and pay them directly.

The first useful run takes about ten minutes: create a workspace, write three or four screens in your own words, and send one submission through.

Getting it genuinely tuned to how your firm decides takes longer, because that comes from you overruling it a few dozen times. There is no configuration shortcut for that and we would not claim one.

It sends from your own mailbox rather than a shared address, which requires authorising it once. CRM write-back is available on higher tiers.

It is designed to feed the systems you already have rather than replace them. If you pay for a pipeline tool, keep it.

Not today. The platform is delivered as a managed service. Self-hosted deployment is something we would consider for a customer who needs it, but it is not built and we will not pretend otherwise.

Reading and screening are unlimited — you should run everything through it. What gets counted is what leaves the building on your behalf.

Starter $890/month for one workspace and one operator. Operator $2,400/month for multiple workspaces, CRM write-back and your own model keys. Firm from $6,000/month for isolated workspaces, a custom domain and compliance export.

Overage on sends is per action at a published rate, not a wall — a run never stops halfway because you crossed a line.

We bill for work the system can prove happened and that passed verification against the specification it was commissioned under. If it cannot be proven, you are not charged for it.

The demos on this site run in your browser with no sign-up — start with Deal Flow Triage and press Run. Figures in them are illustrative and labelled as such; the behaviour is not.

Beyond that, access is granted by hand. There is no self-service sign-up, and that is deliberate while the platform is early.