Ideas
From ERP to ARP: what agents need to run the business
What is the shared record that makes agentic resource planning work?
One system of record
Companies did not adopt ERP for its screens or its reports. They adopted it because it put finance, sales, purchasing and operations on one shared record of the business. Before ERP, every department kept its own version of the truth, and the numbers rarely agreed. After ERP, a sales order, an invoice and a shipment referred to the same thing, for everyone.
The features changed every few years. The shared record is what lasted.
A new term is now circulating for the next step: Agentic Resource Planning, or ARP. The idea is that agents, not only people, will run large parts of the work. It is worth asking the same question companies asked about ERP thirty years ago: what is the shared record that makes this work?
What ARP means
The term is still young, and nobody owns its definition yet. One clear version describes ARP as ERP’s data backbone with a network of autonomous agents added on top (Lucidworks). In that picture, an inventory agent watches stock and reorders on its own, and a finance agent manages cash flow instead of only recording transactions. The same piece draws the analogy directly: ERP replaced MRP, and ARP will replace ERP.
The vendors are already moving. ERP platforms now ship agents that act inside the ERP itself, and expose the ERP to outside agents through standards such as MCP (Logistics IT, May 2026).
Whether the name sticks matters less than the direction. Agents are moving from answering questions to doing work, and doing work requires knowing what the company knows.
What ERP never held
ERP is a record of transactions: orders, invoices, stock, payroll. It was never where a company keeps what it knows.
That knowledge lives in files. Contracts and the memos that explain them. Designs and the reasons a design was rejected. Audit workpapers, test reports, campaign briefs, board decks. It sits on file shares, object storage and document platforms, built up over decades, with permissions set folder by folder.
This is the material an agent needs to do real work. An inventory agent can reorder from ERP data alone. An agent drafting a contract, answering a regulator or preparing a client proposal cannot. It needs the files, the right ones, and it needs to know which of them it is allowed to read.
So ARP inherits ERP’s backbone for transactions, but it has no backbone for knowledge.
The loop every agent runs
Strip any knowledge task down and the same three steps remain.
- Search. Find the knowledge the task needs, across everything the company holds.
- Accurate. Work from the right and current version of it, limited to what this person is allowed to see.
- Place. Put the result back where the next person or agent will look for it.
The third step is the one that matters most, and the one most often skipped. What gets placed today is what gets searched tomorrow. When the loop closes, every task leaves the company a little smarter than it found it. When it doesn’t, every task starts from zero.
People have always run this loop in the file system. They open the shared drive, find the last version, edit it, and save it where colleagues will find it. The file system is the one place where all three steps happen on the same material.
Why today’s tools don’t close the loop
Each of today’s common approaches handles part of the loop well. None of them was designed to run all three steps on the files people already work in.
Data catalogs answer where data lives, who owns it and how it is classified. That is valuable for governance. A catalog holds metadata, not content, so the agent still has to go and fetch from each system on its own, and nothing it produces comes back through the catalog.
Retrieval (RAG) answers which passages look most like the question. Mature retrieval systems can enforce document permissions at query time and return whole documents rather than fragments. What they do not do is take work back. Retrieval reads from its own index; it is a lookup, not a place to work.
Connectors and MCP servers open a door into each system, one system at a time. They solve access. They do not give agents one view across systems, and they say nothing about where an agent’s output should go. In practice it lands in a chat thread, a message or a new folder nobody will search.
All three are built for the first step, and sometimes the second. The third step, placing the work back into shared knowledge, has no owner.
How Keep gives ARP its knowledge backbone
Keep is an AI file system. It sits between the agents a company runs and the storage it already has, and gives agents the same file system people have always worked in. Each step of the loop maps to something Keep does.
One estate, no migration. Keep presents file shares, object storage and document platforms to agents as one file system. The files stay where they are, and no second copy is made.
Search. An agent asks Keep for what a task needs and gets a working set of the right files, instead of crawling the estate on every task.
Accurate. Each working set is built at the moment the agent asks, from the permissions already on the files. An agent sees what its person is allowed to see, and nothing more.
Place. What agents produce is written back beside the files it came from, with the same permissions, where people and the next agent will find it.
Storage that holds up. Keep fetches what a task needs rather than everything, so agent traffic does not compete with the people working on the same storage.
The result is the shared record ARP is missing: one place where people and agents search, work and place results on the same knowledge. A company can start with one team’s files and extend across the estate, without moving anything.
Questions to ask before the first ARP project
ARP will be sold as agents. The agents are the easy part. Before the first project, a CIO can test whether the company is ready with four questions.
- When an agent needs a contract, a design or a past report, where does it look, and does it see the same files my people see?
- Who decides what an agent may read, and is that decision made from the permissions already on the files?
- When an agent finishes a task, where does the result go, and will the next agent find it?
- Are we about to copy our knowledge into one more system to make this work?
ERP succeeded because everyone worked on the same record. ARP will succeed on the same condition, with one difference: this time the record has to hold what the company knows, and it has to be shared by people and agents alike.
That record already exists in the company’s files. Keep makes it one agents can work in.
The series
Get the next post.
One email when the next piece in the series goes up. No product news, no sales follow-up.
Unsubscribe in one click. We do not share the list.