Alembic sees the page. Then it checks its own work.
Most tools hand you the characters and lose the layout — the column header, the row a figure sat in, the note in the margin. Alembic sends the page itself to a model that looks at it, then puts every value it produces through checks it has to pass before you see it. Four layers, and not one of them takes the layer before it on trust.
Three approaches. Three dead ends.
Every team extracting data from documents lands on one of three paths. Each one works just well enough to seem promising — and just poorly enough to create real problems downstream.
Reading text, missing meaning
OCR tools convert pages to raw character strings — then throw away everything that made the document make sense. That table with merged cells, that handwritten annotation, that logo distinguishing an amendment from the original? Gone. You get text. You lose the document.
Brilliant, but unsupervised
Large language models are genuinely impressive at reading documents. The problem isn’t capability — it’s reliability. One prompt, one model, no validation, no memory. It works great on the demo. It hallucinates on page 47 of a real contract, and nobody catches it until the data’s already in your system.
Powerful, if you have six months
Enterprise platforms can handle complexity — after weeks of template configuration, months of training data, and a team dedicated to maintaining the rules. They’re built for organizations with dedicated ops staff and seven-figure volumes. For everyone else, the implementation cost outweighs the extraction value.
The dead end, dramatised
An assistant that answers from memory sounds like the fix. Here is what it costs when the answer was never in the document it claimed to have read.
You’re Absolutely Right! — 48 seconds.
Four layers. One system that actually learns.
Alembic combines visual AI, orchestrated agents, and a learning engine into a single system. Each layer solves a specific failure mode — and they compound.
The AI sees the page like you do
Most tools convert your PDF to plain text before the AI ever touches it. Alembic skips that lossy step entirely. Your documents go directly to the AI as visual input — the same way you’d hand a page to a smart colleague and say “pull out the key terms.”
- Tables stay tables. Merged cells, nested headers, multi-page layouts — processed visually, not reconstructed from text fragments.
- Handwriting and annotations included. Margin notes, stamps, signatures, and corrections are part of the input, not discarded noise.
- Format-agnostic from day one. PDFs, scanned images, photos of paper — no special handling, no separate code paths.
Multiple AI agents, each with a job to do
A single AI model running a single prompt is a demo. Production extraction requires coordination — one agent to classify, another to extract, another to validate, and an orchestrator to manage the whole flow. Alembic assigns the right model to each task automatically. Fast models handle simple lookups. Powerful models handle the hard stuff.
- Per-space orchestration. Each space gets its own extraction team, tuned to that document type. Your invoices don’t share a brain with your contracts.
- Judgment over scores. Clean work flows through. Anything uncertain gets flagged with the AI’s actual reasoning — not a binary pass/fail, an explanation.
- Proactive monitoring. Each space’s agent watches the work, resolves routine issues on its own, and files the rest as a short morning check-in you can decide from directly.
Every correction makes the system permanently smarter
When you fix an extracted value, Alembic doesn’t just update the record. It records the correction as a pattern, and once a pattern has held up across enough documents to be reliable, it is promoted into your space’s standing instructions — so the extraction team reads it before it starts work. Patterns accumulate into a knowledge base specific to your documents, your terminology, your edge cases. Corrections compound: the same mistake shows up less the longer you use it, not more.
- Zero training data required. Start extracting immediately. The system learns from production corrections, not pre-labeled datasets.
- Built by conversation. Tell the AI what you need in plain language. It designs the fields, assembles the extraction, and tests it on your samples.
- Institutional knowledge captured. When your best analyst retires, their corrections and judgment calls live on as patterns in the system.
Answers you can audit
Language models are persuasive readers and unreliable accountants. So when you ask a question about a document, Alembic never asks you to take its word. The coordinator that runs your question holds no reading or arithmetic tools of its own — it can only delegate to scoped specialists, and every number comes back computed rather than recalled. Available today on Enterprise as an API capability: one endpoint that streams progress and returns the cited, verified answer.
- Numbers are computed, not remembered. Arithmetic runs in a code sandbox, so a total is the result of an actual calculation — not a plausible-looking figure a model produced from memory.
- Tables are read as tables. Column structure is detected geometrically from the page, not guessed from flattened text, so a figure never drifts into the wrong row.
- Every claim is checked against the page. Before an answer reaches you, its claims are verified against the source page they came from — and each one arrives with a citation you can open.
Extraction that makes decisions, not just data dumps
The point was never the extraction itself. It was always about what comes next — the approval, the payment, the flag, the decision. Alembic closes the loop between “data extracted” and “action taken.”
See it work on your documents
Upload a sample document and watch Alembic extract, validate, and structure your data in real time. No credit card. No sales call. No six-month implementation plan.
Start free