Skip to main content
Apragya AIFree quotation builderQuotation builderQuote builderno signup
AI governance

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.

A compliance officer using a sieve that catches red slips

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
Six policies, each with its own action, covered what compliance cared about.

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.

Letters ready to send, with four held back under a padlock
Four drafts were held back for repeating a customer's identity number.

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.

  1. Define a policy: Pick a type: personal data, blocked keywords, pattern, content, token or cost limit.
  2. Choose the check point: Apply it to the input, the output, or both.
  3. Choose the action: Block the run, warn, or just log the case.
  4. Attach to agents: Apply policies to the agents they are meant for.
  5. 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

FeatureInstructions in the promptApragya guardrail policies
EnforcementThe model may ignore themChecked outside the model
Indian identifiersNot specificPAN and Aadhaar patterns
ActionNoneBlock, warn or log
Cost controlNoneToken and cost limits
EvidenceNoneViolations 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

  1. List what an agent must never read, write or spend.
  2. Start with personal data checks on output, set to block.
  3. Add blocked terms from your compliance team.
  4. Set token and cost limits to warn before you set them to block.
  5. 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

Free
Up to 3 users
CRM, HR and Vendor Onboarding
INR 3,999
Starter, a month
Up to 10 users, CRM and HR plus a choice of apps
INR 7,999
Pro, a month
Up to 25 users and every enterprise app

AI guardrails questions

Do guardrails replace human review?
No. They catch known risks automatically. Human review covers judgement calls.
Can a policy apply to only one agent?
Yes. Policies can be attached to the agents they are meant for.
Can we add our own patterns?
Yes. A pattern policy can match formats such as your loan or policy numbers.
Do guardrails slow agents down?
The checks are simple pattern and limit checks, so they add little time to a run.
What does it cost to start?
The free plan covers up to 3 users with CRM, HR and Vendor Onboarding and 1,000 AI Credits a month. Starter is INR 3,999 a month for up to 10 users, with CRM and HR plus a choice of apps such as Sales and POS, Inventory, Payroll, Purchase and Payables, Support and Manufacturing. Pro is INR 7,999 a month for up to 25 users with every enterprise app, including Finance and Accounting, plus 5,000 AI Credits, single sign-on and audit logs.

Put rules around your first agent

Start with personal data checks, add your blocked terms, and read the log weekly.