Run AI Agents on Your Own Data: Records, Documents and Systems, Without Copying Them Around
For operations and IT heads who want AI that knows their customers, stock and policies, without staff pasting data into a chat tool.
To run AI agents on your own data, the agent must be able to reach that data safely: the order it is checking, the policy it must follow, the system that holds the stock. A general model knows none of these.
The common workaround is copy and paste. Staff paste customer details into a chat tool, get an answer, and paste it back. Data leaks out, and nothing is recorded.
This page covers what that costs, an illustrative example, and the three ways Apragya agents reach your data.
An example: a Nagpur agri-inputs distributor
Take an illustrative distributor of seeds and fertiliser selling to dealers on credit. The credit team wanted an agent to suggest a credit limit for each new dealer order, using the dealer's history and the company's credit policy.
What the agent used for one dealer
- Dealer recordOrders, payments and overdue invoices.Checked
- Credit policySearched from the policy document.Checked
- Season calendarKharif demand, from a reference sheet.Checked
- Overdue invoicesTwo past the agreed period.Overdue
- Suggested limitLower than asked, with the reason.Ready
- Credit head reviewApproves or changes the limit.Pending
The credit head approved most suggestions as they were and changed a few. Each dealer record now shows the suggested limit, the reasons, and who approved it, and no dealer data went into a chat tool.
How AI agents on company data work
Step 1 of 5
Bound to records
An agent runs on a record event and reads that record and its links.
- Bound to records: An agent runs on a record event and reads that record and its links.
- Knowledge: Upload policies, price lists and manuals; the agent searches them.
- Connectors: Reach databases, SharePoint, S3, FTP or SAP through connectors.
- Tenant Runner: Reach systems inside your network over an outbound connection.
- Output and review: Results are saved on the record, with human review where set.
Before and after agents on your own data
Agents on your data
- The agent reads the record it is bound to
- Output is saved on the record
- Each run is logged
- Policy documents are searched as knowledge
Private data AI agents: what to check
An agent is only as safe as its access. It should read only what the task needs, run under role controls, have personal data checks on its output, and log every run. Model calls can use your own provider keys.
Chat tools against Apragya agents
| Feature | General chat tool | Apragya agents |
|---|---|---|
| Data source | Pasted by staff | Records, knowledge and connectors |
| Output | Pasted back | Saved on the record |
| Personal data | Unchecked | Guardrails and PII masking |
| Record of use | None | Each run logged |
| Model account | The tool's | Yours, where you add keys |
Checked against public information on 16 September 2026.
What sets Apragya apart when you connect AI to ERP data
Already on your records
Sales, stock, finance and HR records are in the same platform.
Built without code
Describe the agent and bind it to a record.
Guardrails
Checks on input and output, with block, warn or log.
Your keys
Model calls can use your organisation's provider keys.
Metered
Every run is counted in AI Credits.
Inspectable
Runs keep their steps and output.
See agents for credit limit recommendation and collections prioritisation, or bring your own LLM key.
Getting started
- Pick one decision that uses records you already keep.
- Collect the policy documents that decision follows.
- Build the agent and add those documents as knowledge.
- Bind it to the record event and keep review on.
- Check the first fifty suggestions against what people decided.
The policy documents usually need a tidy-up first; agents follow what is written, not what is meant.
What it costs to start
AI agents on your own data questions
Does our data train the model?
Can agents reach our on-premise systems?
Can agents change records?
Which systems have connectors?
What does it cost to start?
Let one agent read your own records
Pick one decision, add its policy, and keep review on for the first month.