AI Artifact Governance: Know What Your Agents Have Built
November 18, 2026
Ask an agent for a pipeline dashboard and you get one fast. Ask for a renewal tracker, a pricing calculator, or a page comparing three vendor contracts, and you get those too. People in finance, sales and operations do this every day now, and the output piles up. Some of it stays in chat history. Some of it gets a share link and lands in a Slack channel. Some of it reads from a live table that an agent built and nobody else knows about.
These artifacts hold customer lists, forecasts, contract terms and headcount plans. Try asking how many of them exist in your company, who can open each one, and which tables they read. In most companies there's no place to look that up. We think this is the next governance problem in enterprise AI, and it's piling up faster than the last ones did.
AI Artifact Sprawl Is Already Here
Writing for SAP News in August, Paul Taylor defined the problem on the agent side: "Agent sprawl occurs when AI agents are created, deployed, or connected across systems faster than the enterprise can inventory them, assign ownership, control permissions, monitor behavior, and optimize or retire them when they are no longer fit for purpose." The same piece cites a Gartner forecast that by 2028 the average global Fortune 500 enterprise will have more than 150,000 AI agents in use, and that only 13% of organizations believe they have the right governance in place.
Agent sprawl is about the agents. AI artifact sprawl is about what they leave behind. Every session can save a page, a document, a table or a small app, and each of those can outlive the chat that made it. When people build these with personal or team AI accounts and share them by link, you get shadow AI artifacts: real business data in places IT never sees.
The vendors have made sensible choices here. In Claude, artifacts start private and get saved to an Artifacts tab in the sidebar. On Team and Enterprise plans, artifacts stay inside the org until an Owner turns on external sharing. Once that's on, an "Anyone with the link" share opens for people outside the company. Those defaults are good. They also mean the record of what exists lives inside one vendor's product, split by user.
Enterprises Have Governed This Before: Spreadsheets and File Shares
Spreadsheets were the first wave of software built by people who weren't engineers. Banks call it end-user computing, and the London Whale is the case everyone remembers. In 2012 JPMorgan's Chief Investment Office lost about $6.2 billion. The bank's January 2013 task force report found that its value-at-risk model ran on spreadsheets fed by manual copy and paste, and that one calculation divided by the sum of two rates instead of their average. That error muted the volatility the model was supposed to measure.
The fix wasn't to ban spreadsheets. Large banks now keep end-user computing inventories of their critical spreadsheets. You can't test, review or retire a spreadsheet that nobody knows exists, so the list comes first.
File shares were the second wave. In 2016 Veritas surveyed more than 2,500 IT professionals in 22 countries for its Global Databerg Report. On average, 52% of the data companies stored was dark, meaning nobody knew what it was. Only 15% was business critical. The rest, about a third, was ROT: redundant, obsolete or trivial. The cleanup started the same way. Scan the shares, find out what's there, find an owner, delete what nobody needs.
AI artifacts look like both. Like spreadsheets, they're built by knowledge workers who now write software with AI, and they end up running real decisions. Like file shares, they pile up in storage that nobody cleans. The difference is speed. A person built a spreadsheet over days. An agent builds an app in one prompt.
What AI Artifact Governance Covers
Data governance already has the vocabulary for this. We'd borrow most of it and apply it to the things agents build.
Inventory
A list of every artifact: what kind it is (page, document, app, table), who created it, which agent made it, when, and where it's shared. Everything below depends on this list. If you can't enumerate the artifacts, you can't review their access or count their cost.
Access
Least privilege applies here the same way it does anywhere else. Artifacts should start private, then go to named people, then to the whole company, and only reach "anyone with the link" on purpose. Public links need regular access reviews, the same kind you already run for SaaS apps and shared drives.
There's a second access question that's easy to miss: access to the artifact and access to the data behind it are separate. In Claude, artifacts that use connected apps run on the viewer's connections, not the creator's, and for link viewers outside the org, "the parts that use connected apps or Claude don't work for them." Other tools handle it differently. You want to know which model each artifact follows.
Lineage
Which artifacts read which tables, and which ones write to them? When a forecast table changes, you want to know which dashboards now show new numbers. When a customer table holds personal data, you want to know which pages expose it. The London Whale model broke at a manual copy-and-paste step, which is a lineage gap. Anthropic's Claude Data goes after a related problem with certified metric definitions, so people building on a number know it's the approved one.
Usage and Retention
Who opens each artifact, and when did anyone last open it? An artifact nobody has viewed in months is the AI version of ROT, and it should be archived or deleted on a schedule. Deletion needs care. Claude's help center notes that unpublishing a legacy artifact permanently deletes its stored data, so a retention policy has to know what's stored inside the artifact before it removes anything.
Cost
Artifacts cost money to store, to run, and sometimes to use. In Claude, an artifact that calls Claude counts against each viewer's own plan limits. Other tools bill storage and compute to the account that hosts them. Finance teams already know the two moves here. Showback tells each team what its artifacts cost. Chargeback bills them for it. Both need the inventory first, and both work better with spend limits set per team or per person.
Capture Governance When the Artifact Is Saved
The spreadsheet-era registers depended on people. Someone filled in a form for each critical spreadsheet: owner, purpose, data sources, review date. That works for a model someone spends weeks building. Nobody will fill in a register for files an agent makes every few minutes, and we wouldn't ask them to.
So the record has to come from the save itself. Whatever system stores the artifact should write the log entry as a side effect: who saved it, which agent, when, what it reads, who it's shared with. Git works this way. Nobody fills in a form to say they changed a file; the commit records it.
Anchor is one example of a tool built this way. It describes itself as a file system for AI artifacts: HTML pages, Markdown documents, presentations, apps and live tables in folders, each with a link. Agents read and write to it over MCP, and people can edit in the browser. Anchor says it logs every operation and records whether it came from an agent over MCP or from a person, logs table reads and writes per app, reports usage and storage, and lets admins set org-wide and per-member spend limits. Its version history records which agent or person made each edit. For larger companies, Anchor's enterprise plan lists audit logs, SSO, and on-prem, self-hosted or managed VPC deployment.
It has limits, and you should weigh them. Anchor is a young product, and it isn't open source. It's one more system for people to adopt, and adoption is the hard part of any governance tool. It governs only what gets saved there. An artifact someone leaves in Claude's Artifacts tab or a ChatGPT canvas stays invisible to it. It also has nothing like Claude Data's certified metrics. If you want to build your own, there are open-source starts too, like artifact-server, a young project for publishing agent-made HTML to a server you run.
Why Artifact Governance Has to Be Vendor-Neutral
Most enterprises don't run on one model. Menlo Ventures' 2025 enterprise survey put Anthropic at 40% of enterprise LLM API spend, OpenAI at 27% and Google at 21%. Inside one company, the sales team uses ChatGPT, the analysts use Claude, and the engineers use Cursor, Codex or Gemini CLI.
A model vendor's admin console only sees its own product. Claude's tools govern Claude artifacts. ChatGPT Canvas work stays in ChatGPT. Gemini Canvas hands work off to Google Docs and Slides. An inventory split across four admin consoles doesn't answer the question "what have our agents built?" It answers four smaller questions, and someone still has to merge them by hand.
MCP gives you a common path. Claude, ChatGPT, Cursor, Gemini CLI and Codex all support MCP clients, so a store that speaks MCP can sit behind every agent and log saves the same way regardless of which model made them. We made a similar argument about the tools agents call in governing MCP servers, skills and plugins at enterprise scale. The artifacts agents produce need the same treatment, and they need it in one place that doesn't change when you switch models.
Don't Slow Down Artifact Creation
People use agents to build things because it's fast. If every dashboard needs a ticket and an approval, people will build them in personal accounts instead, and you'll have more shadow AI artifacts than before. The banks didn't stop analysts from using spreadsheets after the London Whale. They started keeping a list.
We'd take the same approach with AI artifacts. Let people build whatever they need. Capture the record when they save it, keep it in one place across every agent, and review it on a schedule. You don't need to approve everything an agent makes. You do need to know what exists, who can open it, and what it touches.
Disclosure: Neural Bridge and Anchor are related companies.