What's behind Pramatic.

This page is for anyone who wants to understand why Pramatic works the way it does: the problem we saw, why we solved it this way, where AI belongs and where we deliberately keep it out, and what is genuinely hard about this. It does not explain how the thing is built — that part is ours.

What we saw

Look closely at any company and the same thing is happening. The information exists, it is complete and it is correct. And still, to answer a simple question — did what we are about to pay for actually arrive in full? — somebody has to chase paper across three systems and two inboxes.

The problem is not missing data. The problem is that the data is loose. Every document stays where it was born: the invoice in an inbox, the delivery note in the warehouse, the contract in a folder, the payment in the bank. Nobody brings them together until something doesn't add up, and by then reconstructing what happened takes days.

The idea

A company runs on agreements: somebody asks for something, somebody promises to deliver, something arrives, somebody pays. Each of those moments leaves behind a different document, and none of them, read on its own, can tell you whether the agreement was kept.

Figure 1 · One agreement, and the trail it leaves

The invoice, read on its own, looks perfect.

So Pramatic does not file documents: it turns them into data that is connected to the rest. Storing a file so you can find it later was solved decades ago. What nobody solved for you is making the value in one document talk to the value in another — and that is where everything an owner needs to know actually lives.

Where AI belongs — and where it doesn't

AI is not the product. It does a job a person is doing today with a keyboard: reading each document and moving the values across. It belongs in exactly two places, and there is one point where we deliberately keep it out.

Figure 2 · AI in the system

What you already have

Invoices, delivery notes, contracts, emails, spreadsheets, machine readings — as they are.

Reads and connects

Pulls the values out of each document, keeps where each one came from, and proposes the cross-checks.

Your team

Approves

Nothing enters the base without a sign-off. The uncertain case lands here; it does not slip through.

Your database

Records connected to each other, every value carrying its origin.

Answers with the source

Ask in plain language; the answer arrives showing the documents that hold it up.

AI does the work. Your team decides. The base remembers where every value came from.

The place it does not belong is the third one. That is where the line runs between processing and being answerable for something, and a person on your team crosses that line. This is not a limitation we plan to remove in a future release: it is the part of the design that makes everything else trustworthy.

What we don't negotiate

Every value we state carries its source.

If Pramatic says you were billed for 48 units, it can show you the invoice and the page where it says so. Every time.

A value without evidence is an opinion in the shape of a fact. And in front of a tax authority, an auditor or a supplier disputing a charge, a company cannot defend itself with opinions.

Why this only became possible now

Having everything connected is not a new ambition; what did not exist was a way to get there without asking the client to get organized first.

Until recently, for a system to read your documents they had to be standardized up front: fixed forms, templates, an integration per format, long implementation projects. The cost of tidying up the mess fell on the customer, which is why this class of software only ever reached large companies with an IT team to keep it alive.

What changed is that AI now reads the mess as it is: a crooked scan, a photo of a delivery note, an email with the order in the body, a spreadsheet somebody built their own way. That is what makes it viable to offer this to a mid-sized company instead of only to a corporation. The ambition didn't change — the cost of entry did.

Why Chile is a good place to build this

Chile mandated electronic invoicing early: a pilot in 2002, voluntary operation in 2003 — the first country in the region — and universal obligation by 2018. The tax authority has been issuing a pre-filled monthly VAT return since 2017, built on tens of millions of electronic documents a month.

The practical consequence is unusual and useful: an entire economy of small and mid-sized companies whose documents are already digital and already structured at the fiscal layer — and still disconnected from everything else in the operation. That is a hard, real, unglamorous problem to build against, and the companies living it are within driving distance.

What we will not claim

Three things get asserted lightly in our industry that we are not going to assert, because we don't believe them:

That AI "understands" your business. It reads, connects and cites. That is already a lot, and you can verify it. The rest is conference vocabulary.
That a system can make a commitment on your behalf. A program can write a promise; it cannot stand behind one. It cannot be sued, fired, or lose its reputation. In payments, taxes and audit, somebody has to answer — and that somebody is a person.
That autonomy is a switch. It is a dial, set per process: the system acts on its own where a mistake is cheap and reversible; a person decides where a mistake gets paid for.

Figure 3 · How much autonomy, by the cost of being wrong

Not autonomy yes or no: how much, decided by the cost of being wrong.

This is not the caution of someone lagging behind. It is the only design that survives an audit.

What is genuinely hard

Making this work on real documents means solving things that are not obvious. We name them because knowing where the difficulty lies is part of what we offer; how we solve them is what we don't publish.

Knowing when two things are the same thingThe same supplier written four ways, the same product under three codes, the same delivery in two documents. Without this solved, there is no cross-check.
Never losing the origin of anythingKeeping the document and page behind each value is easy with ten files and hard with hundreds of thousands — especially when documents get corrected, voided and replaced.
Knowing when it doesn't knowA system always returns an answer. The valuable part is recognizing the uncertain case and routing it to a person before it enters the base, not after.
Living with real-world messOperations don't behave like diagrams: documents arrive incomplete, out of order and contradicting each other. The design has to expect the exception, not the happy path.

How this gets measured

A product that rests on evidence has to be able to prove it. The right questions are not "how good is your model?" but:

Of the values the system states, how many are actually supported by the document they cite?
For each type of value — amount, document number, date, quantity — how much does it get right, and how much does it miss?
Of the uncertain cases, how many reached human review before entering the base?

These are demanding, and they are the right ones. We are at pilot stage: building those numbers with real companies. When they exist we will publish them measured, not estimated. If someone shows you a round percentage without saying on what sample and for which type of value, they are not showing you a measurement.

Does this match what you see?

If this speaks to what happens inside your company — or if you are considering investing in or backing one that thinks this way — write to us. We work with a small group of pilot companies.

Let's talk

A note in the margin

Pramatic came out of looking at real operations, not out of a library. Later, reading, we found that others had arrived on their own at similar intuitions: a Chilean engineer, Fernando Flores, wrote in the 1980s that a company holds together through the agreements people make and keep, and that any claim should be able to show what it rests on. And long before that, the Sanskrit word pramāṇa — where our name comes from — already named the same demand: that knowledge is worth what its grounding is worth.

Running into those echoes was a pleasure and gave us confidence in the direction. But the product didn't come from there: it came from invoices, delivery notes and emails of Chilean companies, and from the stubbornness that no value should be left without saying where it came from.