Your commercial brain, and what it does with what it knows.
It holds how your company wins and grows customers, keeps that current on its own, and answers from it. A commercial architect judges what the patterns are worth before your team relies on them.
See how we build a commercial brain. Takes less than 2 minutes.
What goes in, what it does, what comes out.
Ask it the question you would ask the person who was in every meeting.
Answers are grounded on two layers at once: the human-verified record, and everything you have connected, retrieved at the moment you ask.
Each question is fanned out into several alternative phrasings, reranked with a weighting toward recent evidence, and every passage carries its own date. So the answer knows which debrief is the most recent, and shows you which ones it used.
Evidence is held until it means something, however long that takes.
Most of what the brain holds is showing you nothing yet. One remark in one call is noise. The same remark from four accounts over two months is a pattern, and the only way to tell them apart is to still be holding the first one when the fourth arrives.
So evidence accumulates against the pattern it belongs to, weighted so recent evidence leads while older evidence stays in the count. This is the gap that separates AI that returns something from AI that does not: keeping evidence between sessions is what lets the engine weigh it. It is also the property that makes this a company brain rather than a search index.
It keeps evidence until a pattern crosses the bar, then puts the case for it and against it.
Each night the engine attaches what is new to the patterns it is already holding. Each is scored on how often it recurs, across how many distinct sources, and how strongly against your own baseline.
When one crosses the bar, the record is swept for corroboration and for contradiction, and the counterpoints go on the card. A pass that only ever confirms is confirmation shopping.
Every conclusion dated, attributed, and stating the conditions it applies under.
Conclusions are recorded as decision blocks rather than prose. Each one states when it was made, by whom, whether it is locked, working or gated, and what would have to change for it to stop being true.
When one of those conditions breaks, the brain raises it rather than waiting for someone to notice.
A week that starts from the same numbers, read the same way, every time.
The pipeline is snapshotted each week, so this week is read against last week rather than against a board that has already moved underneath you. One definition of a qualified lead is set once and used by every surface that counts one.
Month-to-date is compared on elapsed days, so day three is read against the first three days of the month before it.
The Monday snapshot
The pipeline is snapshotted, so this week can be read against last week rather than against a board that has already moved.
Notes land on the deal
Call notes sync onto the right deal from the meeting transcript, so the record of what was said keeps up with the week.
The scan runs
What is new is attached to the patterns already being held and scored against your own baseline. A quiet night stays quiet.
Strands are merged
A consolidation pass reads the held strands as a set and merges the ones that are a single pattern seen under two names.
It reaches the owner queue
A finding waits for a verdict with its corroboration and its counterpoints attached, alongside member proposals and generated documents.
The documents your team argues over, drawn from the context you have already agreed.
Assets are built with the artefact on the left and the chat on the right, so a draft is iterated rather than regenerated. Versions stay live, with a present mode and a PDF alongside.
At build time you pull a company, contact, deal, meeting or transcript straight in, so a one-pager starts from the real account rather than from a description of it.
It reads the systems you already run, and keeps them current every night.
Connections cover CRM, meetings and calls, documents and drives, chat, analytics, advertising, search and design. Each syncs nightly with content-hash change detection, so only what actually changed is read again.
Connecting a credentialed system is an owner action and the tokens are encrypted at rest. Every link, upload and forwarded email appears as its own row, with who added it and when.
Reach it from the assistant your team already has open.
Every workspace is an MCP server your own assistant can read from. Connect it inside the assistant your team already uses and ask your own model questions grounded in your own record, in the tool they are already working in.
Access runs on personal, workspace-scoped tokens: hash-stored, shown once, revocable. A versioned REST API and a CLI have the same standing.
What happens to everything that comes in, from the point it is received.
A CRM record, a recorded call, a support thread, an ad account, a spreadsheet, a forwarded email, a page you pointed us at. All of it takes the same path in.
It arrives, and it is read to text
A deal record, a call transcript, a chat thread, a campaign report and a web page all come in as text. Word, Excel, PDF and PowerPoint are parsed on the way in. What moves forward is the extracted text, and the text alone is what the product keeps.
It is scrubbed, if you have asked for that
With redaction on, email addresses, phone numbers and bank identifiers are replaced with typed placeholders at this point: before storage, before embedding, before any model sees it, so what lands at rest already carries the placeholder. Names are kept deliberately, because naming your champion is part of reasoning about the account.
It is split into passages and embedded
The text is divided into passages, and each one is turned into a vector so it can be found by meaning rather than by keyword. Every passage keeps its source and its date. Embedding and storage both happen in Frankfurt.
Later syncs re-read only what changed
Each sync compares a content hash against what is already held, so only what changed is parsed and embedded again. That is why a nightly rhythm across your whole stack stays cheap and quiet.
Overnight, it is weighed against everything else
The nightly pass attaches what is new to the patterns already being held, scoring each on how often it recurs, across how many distinct sources, and how strongly it reads against your own baseline. A Sunday pass then merges the strands that turn out to be one pattern seen under two names. A pattern surfaces once it crosses the bar, so a quiet night stays quiet.
It is retrieved when somebody asks
A question is fanned out into several phrasings, the passages that come back are reranked with a weighting toward recent evidence, and the answer shows which ones it used and when each was written. The verified record is in the prompt alongside them, so an answer draws on what your team agreed as well as on what the sources say.
The personal material ages out. The commercial record stays.
Transcripts, email and chat carry a retention window you set, 365 days by default. When material reaches it, the substance is first distilled into a dated, scrubbed summary that joins the permanent record, and only then is the raw item and its passages deleted. Your own documents and uploads are exempt as curated material; connected CRM and analytics regenerate on every sync. Components, decisions, findings and the weekly history stay for the life of the workspace.
Specification.
Retrieval
- Grounding
- The component record and your connected knowledge base, both always active
- Query handling
- Fan-out to several phrasings, reranked with a recency weighting
- Provenance
- Every passage carries its source and date, shown with the answer
- Live web
- Fetches a named page mid-answer, behind request-forgery checks, and keeps it as a real source
The nightly scan
- Cadence
- Nightly, with a weekly consolidation pass that merges duplicate strands
- Scoring
- Recurrence, distinct sources, and intensity against your own baseline
- Surfacing
- A finding appears once it crosses a bar. A quiet night stays quiet
- Verdicts
- Set aside, Monitor, or Lock it in
Sources
- Connections
- Reads the commercial systems you already run, and leaves them as your team left them
- Sync
- Nightly, with content-hash change detection so only changes are re-read
- By email
- Each workspace has an address; forward a thread and it lands
- Removal
- Archive and keep what was learned, or purge source and passages together
Your AI
- Routes
- MCP endpoint, versioned REST API, and a CLI. Each one reads your record
- Tokens
- Personal, workspace-scoped, hash-stored, shown once, revocable
- Your model
- Anthropic, Gemini, or any OpenAI-compatible endpoint, tested before it saves
- Managed
- EU inference on AWS Bedrock, with per-workspace daily limits
Control
- Roles
- Owners, members and platform admins. Removal revokes every session
- Versioning
- Every edit versioned, one-click restore, owner-gated
- Redaction
- Every workspace can enable PII redaction at ingestion
- Retention
- A window you set, with the record itself sitting outside it
Where it runs
- Residency
- Database, compute, inference and embeddings in Frankfurt
- Model terms
- Inputs excluded from training, processed in memory
- Audit
- Append-only, enforced by database grant, actor and before-and-after state
- Hosting
- Shared, your own dedicated database, or inside your own cloud boundary
And the limits it works within, by design.
Questions about the platform.
What data does Liffey read?
Liffey reads the core commercial systems where commercial knowledge already collects: your CRM, meeting recordings, shared drives, chat, email, analytics and advertising. You choose which sources connect and which parts of each one it reaches. Liffey reads them and leaves them exactly as your team left them.
Where does Liffey store and process our data?
Every layer runs inside the European Union. The database is managed PostgreSQL in Frankfurt, application compute runs on serverless functions pinned to Frankfurt, and AI inference is Claude via AWS Bedrock in eu-central-1 with Amazon Titan embeddings in the same region.
What happens to a file after Liffey reads it?
The text is extracted on the way in, and the extracted text is what the platform keeps. It is split into passages, each one turned into a vector so it can be found by meaning rather than keyword, and every passage keeps its source and its date. Embedding and storage both happen in Frankfurt.
Can we use our own AI model with Liffey?
Yes. Liffey's own AI can be routed through your governed endpoint: Anthropic, Google Gemini, or anything OpenAI-compatible, which covers Azure and enterprise gateways. A live call is made against your endpoint before the setting is accepted, and one click returns you to managed EU inference.
How long before Liffey is useful?
The free plan is live in minutes, seeded from your domain's public footprint. With your systems connected, the first answers come back grounded in your own numbers inside the first days. The full commercial architecture takes weeks, built alongside your team.
“We’ve grown by nearly 50% over the past year. In that time, Liffey has helped us better articulate our unique value proposition, evolve our pricing model, reset our go-to-market strategy and build scalable commercial systems. All of which has enabled us to make better decisions on a day-to-day basis and grow with more confidence.”
Nick Comer, Founder & CEO
First commercial insights within minutes.
Connect the systems you already run, and the first answers come back the same day. The full commercial architecture takes weeks, alongside your team.
See how we build a commercial brain. Takes less than 2 minutes.