Company brains /
Agent-ready data architecture is the next enterprise problem
Picture a construction firm's shared drive. Every safety report, every subcontractor certificate, every site photo from the last five years, all sitting in folders marked "everyone in the company can view." A person browsing that drive knows, without anyone telling them, that the certificate from March has probably expired and the site photos from the wet season job are the ones to trust for drainage questions. An agent given the same access knows none of that. It sees a folder, not a decade of judgement about which parts of the folder are still true.
That is the gap. Most companies do not have a lack-of-data problem. They have a permissions-and-shape problem that nobody notices because humans have been quietly working around it for years.
The data was never the hard part
Companies have spent two decades organising the data that fits neatly into rows and columns. Orders, invoices, inventory, customer records. That work is mostly done, and it was never really what an agent needs most.
What agents need next is the material that has always resisted structure:
- Site photos and safety walkthroughs.
- Voice notes from a technician between jobs.
- PDFs of certificates, tenders and compliance archives.
- SOPs that exist as a laminated sheet in a site office, not a file anywhere.
- Old inspection reports, some scanned, some handwritten.
None of this is new information. It has been there the whole time, doing its job for the humans who already knew where to look and which version to trust. It was never built to be found by something that cannot ask the person next to it.
Permissions built for a person do not survive contact with an agent
A human's access to a shared drive comes with judgement attached for free. They know the March certificate lapsed because they were in the room when it did. An agent reading the same folder has the file, not the room.
This is not a bigger version of cleaning your data. Cleaning gets the facts right. Agent-readiness is a separate question: who is allowed to know this, what are they allowed to do with it, and can someone check afterwards why the agent acted the way it did.
What an agent-ready layer actually needs
Getting the permission model right is a specific, buildable thing. On the Shopify agent I run, this is not an abstract requirement, it shaped how the system was staged into production.
- Scoped access. One region of that system is deliberately read-only. Not because the data was unavailable, but because someone decided an unreviewed write in that market was not worth the risk yet. A separate test environment carries full read and write, precisely so mistakes land somewhere safe before a real store does.
- Routing, not blanket access. The system asks which store a request is about rather than assuming one. Product IDs, currencies and catalogues differ by region, and treating them as one pool of data would have been the fast way to get a confident wrong answer.
- A record of what happened. Every action the agent takes comes back wrapped as a clear success or failure, and it will not report something as done unless that wrapper says so. That is the citation and observability layer in miniature: not just "the agent did something," but a specific, checkable claim about what.
Where this matters most
This is not only a large-enterprise problem. It shows up hardest in businesses whose real knowledge was never digital to begin with: a field-service team's fault history, a construction site's evolving safety record, a freight desk's memory of which lane goes wrong in which season.
That knowledge sits in photos, voice notes and paper, permissioned by nothing more formal than "ask Dave." Making it agent-ready is not a smaller version of a tech company's data pipeline. It is closer to writing down, for the first time, who is allowed to know what and act on it, in a business that never had to answer that question before because a person was always the one holding the file.
The rule
Do not start the next AI project by asking what data you have. Ask what an agent could safely know, cite and act on today, and what is stopping it.
Usually the answer is not missing data. It is a permission model built for a person standing in the room, applied without change to something that never stood anywhere.