AI Artifact Governance: Track What Your Agents Make

AI artifact governance is how a company keeps track of the files its AI agents make: what exists, who can open it, what data it reads, how often people use it, and what it costs. The files are the HTML pages, Markdown documents, small apps, live tables and dashboards that people ask their agents for every day. Data teams have run databases under rules like these for years. Almost no company runs its artifacts under them yet, and artifacts are piling up faster than databases ever did.

We build Anchor, a file system for AI artifacts, and we think every CIO will get asked about this within a year. Here's how we think about it, and what we built.

Why is AI artifact sprawl a problem?

An artifact takes a minute to make, and nobody plans where it goes. A sales lead asks Claude for a territory dashboard. An analyst has ChatGPT build a pricing calculator. Someone in finance keeps a table of vendor contracts that an agent updates every Monday. Each one is a file, most of them hold company data, and each one ends up somewhere different: a chat thread, a laptop's Downloads folder, a public link from the AI vendor, or a personal account the company doesn't control.

Ask a CIO how many dashboards their company's agents built last quarter, who can open them, and which ones touch customer data. Few can answer any of the three today.

Security teams call the tool side of this shadow AI, and most of them already track which AI apps employees use. Fewer track what those apps leave behind. The artifacts are what's left after the chat ends. They hold customer lists, forecasts and contract terms, and they keep running (and keep getting forwarded) long after the person who made them has changed teams.

Agents make the pile grow faster. Writing for SAP, Paul Taylor defined agent sprawl as agents being "created, deployed, or connected across systems faster than the enterprise can inventory them, assign ownership, control permissions, monitor behavior, and optimize or retire them." The article cites Gartner's estimate that the average Fortune 500 company will have more than 150,000 AI agents in use by 2028, and that only 13% of organizations think they have the governance to manage them. Every one of those agents writes files. Put "artifacts" where that definition says "agents" and you have the problem we're describing.

Have enterprises dealt with this before?

Twice, and both times the fix started with a list.

Spreadsheets came first. Banks call them end-user computing, or EUC: models and tools that business people build for themselves, outside of IT. In 2012, JPMorgan's Chief Investment Office lost more than $6 billion on the trades that became known as the London Whale. The bank's own task force found that the risk model behind them "operated through a series of Excel spreadsheets, which had to be completed manually, by a process of copying and pasting data from one spreadsheet to another." One cell divided by a sum where it should have divided by an average, and the model understated volatility by about half. Big banks now keep EUC inventories that list every critical spreadsheet, what it feeds, and who last checked it.

File shares came second. In 2016, Veritas asked 2,500 IT leaders in 22 countries about the data their companies stored, for its Databerg report. They called 52% of it dark, meaning nobody knew what was in it or what it was worth. Another 33% was redundant, obsolete or trivial, which records managers shorten to ROT. Only 15% was business critical.

AI artifacts have both problems at once. They're end-user computing, because anyone can now build an app that reads live company data. And they pile up like a file share, because one more costs nothing to make.

What does AI artifact governance include?

Data governance already has the words for this, and we borrowed them. It's a pretty short list.

Inventory. A data catalog lists every dataset a company has. An artifact inventory does the same for every page, document, app and table: what it is, where it lives, who made it and when it last changed. Every other control depends on this list, since you can't set access on a file you don't know about.

Access. Access should follow least privilege: people get the level they need, on the files they need. When someone leaves or changes teams, an access review should show what they could open and who they shared it with.

Lineage. In a data warehouse, lineage traces where data comes from and where it goes. For artifacts, it answers questions like which apps read the customer table, and how often. It also shows you the blast radius before you change anything. Retire a table, and you know which dashboards break.

Usage and retention. Which artifacts people open, how many times, and which ones nobody has touched since spring. That's how you find your ROT, and how you set a retention policy you can explain to an auditor.

Cost. Cloud teams learned to split the bill by team, which FinOps people call showback (or chargeback, once the money moves between budgets). Agents use compute and storage too, so the same questions apply: who's using how much, and on what.

How does Anchor govern AI artifacts automatically?

Most governance programs ask employees to register what they make, tag it and file it. That works for a few hundred datasets run by a data team. It won't work for agents that write a new file every few minutes. So in Anchor, the records come from where the file is saved.

Agents write artifacts into Anchor over MCP, from Claude, ChatGPT, Codex or any other agent that speaks it. People can upload and edit in the browser too. The moment a file lands, it's in a folder, in an org, and in the log. There's no separate step, so there's nothing for anyone to forget. And because Anchor works with every agent, one inventory covers the artifacts made with Claude, ChatGPT and the open model your data team runs. A model vendor's admin console only sees its own app. (We wrote more about that in our guide to LLM vendor lock-in.)

What do employees get?

Employees organize artifacts in folders, the way they'd organize files on a laptop. Sharing works per file or per folder, with three roles: Owner, Contributor, who can edit, and Consumer, who can open and use it. Share a folder with your whole org and everything inside it gets the same access, including files added next month. Someone without access can request it, and the owner decides. Anchor also marks which of your files are private, meaning nobody else can open them.

What do admins get?

Anchor records every operation: who did it, on which file and folder, what kind of operation it was, and whether it came from an agent over MCP or from a person in a browser. The admin views are built from that record.

Admins see which artifact is using what. Each time an app reads or writes a table, Anchor logs it, so you can see which apps use the customer table and how many times this month. You can also see which artifacts nobody has opened since spring, so your ROT is a list you can act on.

Admins see usage by day, week, month and billing period, with members ranked by how much compute they use. Storage breaks down by member and by file type, so you can tell whether the growth is one person's video uploads or the whole team's tables.

Admins can also cap it. Anchor has an org-wide daily and weekly limit and a per-member limit on top of it, so one runaway agent can't spend the month's budget in an afternoon!

Should companies slow down how fast people make artifacts?

We don't think so, and we don't think they could if they tried. The dashboards and apps are the good part. They're how a sales lead or an analyst gets something built this week instead of waiting a quarter for an engineer. What a company needs is to know what exists, who has it, and what it touches. Anchor records all three as people work. Employees keep their artifacts organized and share them with the right people, and admins see the inventory, the lineage and the bill. Nobody has to fill in a form.