# AI guardrails and governance that make a system safe to put in front of customers

Source: https://ai.prospicience.in/capabilities/ai-governance-and-guardrails

Every system we ship runs on four rules: answers come from your approved sources or the system says it has no answer, citations come from the source itself, the working is shown, and a named person approves anything client facing. A manufacturer's system with more than fifteen AI helpers logs every AI action and has a person confirm each one.

## What you get

- Answers are grounded in your approved sources, and when nothing matches the system says so, so customers never receive an invented answer.
- Citations are taken from the source passage itself, so every reference can be checked.
- The working is visible: the reporting chat shows its query and the document reader shows what it read before you confirm.
- A named person approves anything that leaves the building, so AI drafts and your team decides.
- Every AI action is written to a log that cannot be edited, and agents are built without permissions they do not need, so your audit trail is complete.

## Systems we have shipped for this

- B16 Packaging concepts, with pixel proof of what changed (Spice brand): https://ai.prospicience.in/work/packaging-concepts-with-proof
- E02 A library that answers questions and always shows its source (Specialist education institute): https://ai.prospicience.in/work/knowledge-library-with-citations
- E04 A card statement sorter for the accounts team (Company finance): https://ai.prospicience.in/work/card-statement-sorter
- E05 A WhatsApp admissions helper for parents, built on an audited brief (Private pre-school): https://ai.prospicience.in/work/whatsapp-admissions-helper
- E10 Brand rules written so AI tools follow them (Used across our client projects): https://ai.prospicience.in/work/brand-rules-for-ai-tools
- E11 A factory's complete business system, with more than fifteen AI helpers built in (Transformer and brass parts manufacturer): https://ai.prospicience.in/work/factory-erp-with-ai-helpers
- E17 Which video model should you trust? A vendor-neutral benchmark (Internal research): https://ai.prospicience.in/work/video-model-benchmark
- E19 An agent that drives a paid tool in the browser, with thirteen safety rules (Internal tooling): https://ai.prospicience.in/work/browser-driving-video-agent
- E20 Costing an agent system before building it (Internal research): https://ai.prospicience.in/work/agent-cost-model

## What businesses come to us for

- A promotions agency wanted an AI audit and roadmap, plus policies covering client confidentiality, intellectual property, consumer data, quality control and human oversight.
- A consultancy wanted a check layer proving AI generated reports matched the raw source exactly.
- A platform team wanted evaluation frameworks confirming AI output was accurate and testable before it reached production.

## Questions

### How do you stop AI making things up in a business system?

We make "no answer" the default. The system searches your sources first, and if nothing matches well enough it replies that it has no answer without calling the AI. Invented answers come from asking a model with nothing to support it, and our design never does that.

### Where should a human sit in an AI workflow?

On the last step before anything leaves the building or becomes a record: confirming what was read before it saves, approving a script before it is voiced, confirming a payment reminder before it sends. Early enough to be quick, late enough to catch what matters.

### What about prompt injection through documents?

All retrieved content is treated as information and never as instructions, with a clear boundary in every prompt, so a document cannot redirect the system.

### How do you handle consumer data in AI tools?

Personal data stays out of general purpose AI tools. Access is enforced per user at the data layer, phone numbers are hidden by default, and every AI action is written to a log that cannot be edited.

### Can you audit AI use across a business?

Yes, and it is a common starting point. We review your workflows, find the repetitive and error prone ones, rank them by impact against effort and risk, and recommend two or three pilots, delivered as a roadmap you own.