Make your AI work only with what your organization has approved.
The Artificer turns your standards, definitions and methods into the rules your AI follows. Your data never passes through The Artificer.
And when someone asks what criteria it worked with, you'll have the answer: which rules it used, which version, and who approved them.
Works with any MCP-compatible AI assistant.
The problem
Your AI answers with what it finds, not with what counts.
The model isn't the problem. What you give it to read is.
It reads everything
Drafts, old versions, duplicate copies. Nobody has said which one is right. The more documents you give it without criteria, the faster it gets things wrong.
One thing, three definitions
“Customer” means one thing in sales, another in finance and another in yesterday's report. Your AI doesn't know which one is your organization's.
Nobody knows where an answer came from
When the AI gets it wrong, there's no way to know which document it used or who signed off on it. People argue over who's right instead of fixing the source.
What it is
An archive of what's approved, and a gate to get into it.
The Artificer organizes what governs your organization —standards, definitions, methods— in two places. An intake desk, where everything new arrives and waits. And the approved archive, which is the only thing your AI consults. We call that archive the Governed Corpus.
Between the two there is a gate. To pass through it, every document is reviewed and a person approves it. Nothing is deleted: when a document changes, the new version goes in and the previous one is kept.
What matters most in data is quality. And quality isn't checked at the end: it's forged at the source.
What isn't defined can't be governed. What can't be measured can't be improved. And what isn't improved ends up degrading. That's why quality isn't declared: it's designed, carried out and proven.
Everyone knows that without quality, data isn't an asset but a dangerous liability. Even so, it usually gets checked too late: when the report is already wrong or the AI has already answered. In The Artificer, quality is checked at the gate, against what your organization has defined. If it doesn't comply, it doesn't get in. This is how it decides what gets in:
Gate rules
Every document has to say what it is, where it comes from, which version it is and what it covers. If that information isn't filled in and can't be completed, it doesn't get in.
It still matches the concept
Your organization writes down what each thing means. For example: a customer is a person with an active contract. If a customer document arrives that includes canceled contracts, it's out of line, because to you a canceled contract isn't a customer. So it doesn't get in.
A person decides
Your AI assistant does the review, but the AI can make mistakes. That's why what gets in is always approved by a person, and it's recorded who.
It works today, with Artificer Core.
Products
Three pieces. The AI always does the work.
One piece says how the work is done, another is the place where it's done, and the third handles what always has to come out the same.
Artificer Core
The knowledge engine. It gives your AI assistant the method: the gate rules, the steps for each job and the templates, so it can review your documents against your own definitions.
- It connects to your AI assistant.
- It never receives your documents: it only gives instructions.
- Your documents stay in your folders.
Artificer Forge
The workplace. Forge doesn't do the work: the AI always does, guided by Forge. What it adds is that working with Core becomes simpler and can be done at scale.
- All jobs on one screen; it pauses when a person has to decide.
- Document map: a graph of your concepts and the documents linked to each one.
- Document marketplace: you see the catalog of what's approved and request access to what you need.
- Cost tracking: how much the AI uses on each job.
Artificer Anvil
The deterministic part. If you want a job to run the same way every time, an AI isn't the right tool: it's expensive and can give a different result each time. Anvil adds engines that do exactly what was approved, the same way every time.
- First engine: loading data tables.
- The AI starts it; the engine does the exact part.
- Cheaper and more repeatable than asking an AI.
In short: Core says how. The AI does the work. Forge is where you organize it. Anvil does what always has to come out the same. And a person decides.
An example
What a customer is, from the definition to the count.
Your organization defines once what a customer is. Today, with Artificer Core, that definition stops what doesn't fit. Tomorrow, with Anvil, it will also do the math.
Today, with Artificer Core
You write down what a customer is
For example: a person with an active contract. Your organization decides, not the tool.
A person approves it
The definition goes into the approved archive, with its version and who approved it.
A report arrives that doesn't fit
The sales team submits a report that counts as customers people who have already canceled their contract. Your assistant reviews it with Core's instructions, compares it with the approved definition and doesn't let it through.
A person decides
Your assistant shows you what doesn't fit and why. The AI can make mistakes, so a person decides: fix it, split it or discard it.
Tomorrow, with Artificer Anvil In development
The count comes from the definition
How many customers there are each month is calculated by an Anvil engine from the approved definition, on your data and in your systems. Always the same: no AI interpreting it each time.
The past can still be explained
Every count leaves a trail of which definition it followed, which version and who approved it. If the definition changes tomorrow, the new one is approved, and the previous one stays there to explain past counts.
What you gain
What changes for you.
No more mixed-up versions
Your AI only uses what's approved. What's been replaced stops being used, but it doesn't disappear: you can always see what it said.
Each team, its own
What's shared is defined once, and each team sees only its own: it only accesses the documentation that applies to it.
What your AI knew on day X
Every job done with Core leaves a trail of which rules it followed, which version and who approved them. You look it up; you don't reconstruct it.
Your data is yours
The Artificer governs the method, not your data. It doesn't hold it: your documents stay where they are.
Ready for when you're asked
The EU AI Act requires traceability for high-risk uses. With The Artificer, the criteria your AI followed are already written down.
Essays
Why your AI fails, and how to fix it.
Two reads, each under ten minutes.
Your AI isn't hallucinating. You're feeding it garbage.
Two invisible laws explain why your assistant fails even as models keep getting better: volume without criteria, and the same concept living in three places.
Read the full essay
The fix for your AI already existed. Nobody had brought it to documents.
Kitchens solved it with three habits, and the data world arrived at the same ones thirty years ago: a contract at the door, a memory that never gets overwritten, and a shared recipe book with a local menu.
Read the full essay