AI Guardrails for Enterprise: Check What Goes Into and Comes Out of Every Agent
For compliance officers, IT heads and business owners who want AI agents at work, with clear rules on what they may read, say and spend.
AI guardrails for enterprise use are the rules an agent must pass before it acts. They catch an Aadhaar number in a prompt, a blocked term in a reply, or a run that is about to cost far more than it should.
Without them, the only safeguard is a person reading everything, which defeats the point of automation. With badly designed ones, agents are blocked so often that people stop using them.
This page explains how guardrails work in Apragya, with an illustrative example from a lending business.
An example: a Kochi gold loan company
Take an illustrative lender whose support agent drafts replies to customer emails. Customers often include PAN and Aadhaar numbers. The compliance officer wanted those kept out of replies, and a cap on how long a single run could go on.
The policies set on the support agent
- Personal data in outputPAN, Aadhaar, card and phone patterns.Checked
- Blocked termsPhrases compliance does not allow in replies.Checked
- Account number patternA custom pattern for the loan number format.Checked
- Token limitStops unusually long runs.Ready
- Cost limitA ceiling per run.Ready
- Weekly violation reviewCompliance reads the log every Monday.Pending
In the first month the policies blocked four drafts that repeated a customer's Aadhaar number, and warned on a handful of long runs. Each case was logged on the run, so the officer reviewed them in minutes.
How LLM guardrails work in Apragya
Step 1 of 5
Define a policy
Pick a type: personal data, blocked keywords, pattern, content, token or cost limit.
- Define a policy: Pick a type: personal data, blocked keywords, pattern, content, token or cost limit.
- Choose the check point: Apply it to the input, the output, or both.
- Choose the action: Block the run, warn, or just log the case.
- Attach to agents: Apply policies to the agents they are meant for.
- Review violations: Each case is recorded on the run with what matched.
Where unguarded AI goes wrong
Personal data leaks
Identity numbers copied from an email into a reply.
Wrong words
Promises or terms the business never allows in writing.
Runaway runs
A loop that keeps calling a large model.
No evidence
Nothing to show an auditor about how AI was controlled.
Choosing AI governance software in India
Indian businesses deal with PAN, Aadhaar and GSTIN in ordinary documents, so personal data checks must know those formats. Guardrails should also sit next to role-based access, human review and an audit log, not in a separate tool.
Prompt instructions against Apragya guardrails
| Feature | Instructions in the prompt | Apragya guardrail policies |
|---|---|---|
| Enforcement | The model may ignore them | Checked outside the model |
| Indian identifiers | Not specific | PAN and Aadhaar patterns |
| Action | None | Block, warn or log |
| Cost control | None | Token and cost limits |
| Evidence | None | Violations recorded per run |
Checked against public information on 16 September 2026.
What sets PII protection for AI apart in Apragya
Human review
Risky output can wait for a person to approve or modify it.
PII masking
Sensitive fields can be masked for roles that should not see them.
Role-based access
Who can change policies is controlled by role.
Audit log
User, role, action, record, time and IP for changes.
AI Credits
Usage is metered and allocated per user and app.
One place
Policies, runs and review live beside the records.
Guardrails work with human review and AI Credits. See how the audit trail compares in Apragya and Odoo.
Getting started with guardrails
- List what an agent must never read, write or spend.
- Start with personal data checks on output, set to block.
- Add blocked terms from your compliance team.
- Set token and cost limits to warn before you set them to block.
- Review the violation log weekly for the first month.
Start strict on personal data and gentle on everything else, then tighten from what the log shows.
What it costs to start
AI guardrails questions
Do guardrails replace human review?
Can a policy apply to only one agent?
Can we add our own patterns?
Do guardrails slow agents down?
What does it cost to start?
Put rules around your first agent
Start with personal data checks, add your blocked terms, and read the log weekly.