News
An AI agent on a live ERP: the control is older than the technology
November 12, 2026 · PenDraco
IBM's 2026 breach study: 92 percent of organizations with an AI-related breach lacked proper AI access controls. One captured session on a live ERPNext shows the three controls an agent needs before it writes to the books, and where each one comes from.
An AI agent that only answers questions can be wrong on the screen. An agent that can write to the system of record can be wrong in the ledger. The second kind is the one companies now want, and the evidence says most are not ready for it.
IBM's 2026 Cost of a Data Breach study covers 602 organizations that had a breach. Among those that reported an AI-related breach, 92 percent lacked proper AI access controls. Across the study, only 40 percent of organizations reported using access controls on their AI models and data.
So the useful question about an agent is not how capable it is. It is three smaller questions. What can the agent change? Who approved the change? What record remains? This article reads one public session against them.
What the session shows
The session is on jakada.ai. Jakada.ai is PenDraco's AI platform for ERPNext. The page shows it working a live ERPNext system that holds demonstration data. PenDraco captured the session from a live run. Every tool call in it is real.
The session has six requests in plain language. The agent lists the document types in the system and reports 30. It reads the Sales Order schema and reports 170 fields. It searches the customer records. It creates a to-do to follow up on an overdue invoice. It lists the to-dos. It runs the General Ledger report.
Five of the six requests are reads. One is a write, and the write does not happen when the agent decides it should. The agent states its intent in one sentence, and the session stops at a confirmation gate. A person approves. Only then does the record exist. Every step leaves a lineage entry: the time, the tool, the result.
Three controls, each older than the agent
None of this is new control design. Each part answers a requirement that predates language models.
A short list of things it can do. The OWASP Top 10 for LLM Applications lists excessive agency as risk LLM06. It names three root causes. The agent can do too much. It holds too many permissions. It acts with too much autonomy. The first remedies concern the tool list. Give the agent only the tools the work needs. Keep each tool narrow. Avoid open-ended tools, such as one that runs any shell command. The agent in the session works from a catalog of 16 named tools. Each of the five it uses does one thing: list the document types, read a schema, search documents, create a document, run a report.
NIST SP 800-53, the federal catalog of security and privacy controls, already names the case. Its least-privilege control, AC-6, covers users "or processes acting on behalf of users." An agent is such a process.
A person who authorizes. The Government Accountability Office publishes the federal standard for internal control, the Green Book. One of its control activities is segregation of duties. Management separates four responsibilities: authorizing transactions, processing and recording them, reviewing them, handling the related assets. The stated purpose is "so that no one individual controls all key aspects of a transaction or event."
The authors wrote that sentence for people. It applies to software without a change. An agent that can decide a write and also commit it holds two duties that the standard keeps apart. The confirmation gate separates them again. The agent processes and records. A person authorizes.
OWASP reaches the same place from the security side. It says a human must approve high-impact actions before the system takes them. It also says to enforce authorization in the downstream system, not in the model's judgment.
A record that answers an auditor. The same NIST catalog says what an audit record must establish. Control AU-3 lists six items: the type of event, the time, the place, the source, the outcome, the identities involved.
A chat transcript records what the model said. The control asks for a record of what the system did. Each lineage entry in the session carries the time and the tool, with the result. The entry for the write also carries the approval. Jakada.ai's own description: a committed change records the documents the agent read and the fields it wrote. The product page calls this "lineage your auditor can query, not a chat transcript."
Where it runs
The last question is location. In a Jakada.ai deployment, the model runs on infrastructure the customer controls. So do the tools and the data. Prompts and documents stay inside that boundary. The product page states one exception and names it: a license heartbeat, the only outbound connection, with its contents documented.
What this means for a leader
Ask four questions of any agent that will touch a system of record. Ask them of us too.
What can it do? Ask for the tool list. A short list of narrow tools is a control.
What can it change without a person? For a system of record, the right answer is nothing.
What record does it leave? Your controller should be able to query it without reading a conversation.
Where do the model and your data run?
The control is older than the technology. A new accounts-payable clerk does not approve the payments that the clerk enters. An agent should not approve its own writes.
Jakada.ai is the platform PenDraco's engineers deploy inside a customer's ERPNext and leave running there. Watch the session. If you run ERPNext and want to see the same controls on your own system, get in touch.
Sources
- IBM and Ponemon Institute, Cost of a Data Breach Report 2026, July 2026: 602 organizations with breaches from March 2025 through February 2026; 92 percent of organizations that reported an AI-related breach lacked proper AI access controls; 40 percent reported using access controls on AI models and data.
- OWASP Gen AI Security Project, LLM06:2025 Excessive Agency, OWASP Top 10 for LLM Applications 2025: the three root causes and the prevention strategies (minimize extensions, require user approval, complete mediation).
- U.S. Government Accountability Office, Standards for Internal Control in the Federal Government, GAO-25-107721, May 2025, Principle 10, Design Control Activities: segregation of duties.
- National Institute of Standards and Technology, Security and Privacy Controls for Information Systems and Organizations, SP 800-53 Revision 5: AC-6, Least Privilege, and AU-3, Content of Audit Records.
- Jakada.ai, the captured session and the product description, read November 12, 2026. Jakada.ai is a PenDraco product.