PRACTICAL AI / ARCHITECTURE AND GOVERNANCE

Assume AI will try every door.
Design so the wrong ones never open.

AI instructions are not security controls. Decide what people, applications, automations and AI can reach, then enforce those decisions in the architecture. A request that should be refused must still fail regardless of how it is worded, where it comes from or which model makes it.

  • Public and private separated
  • Identity before retrieval
  • Controlled delivery
01 / DEFINE THE CONTROL POINTS

Draw the boundaries before you build

Security architecture should answer a simple question: how could someone, or something, reach this information?

Don't just think about the app people normally use. Include AI, automations, APIs, reports, exports, shared links and admin tools. For every route, decide who or what can use it, what information it can reach and whether it really needs that access.

Public routes

Websites, forms and public AI can collect information and start processes without needing access to your private business records.

Stop and think: does this public component really need access to sensitive data, or can it simply pass the request into your protected environment?

Internal routes

Staff, internal apps and AI assistants should work through the identities and permissions your organisation already manages.

Stop and think: adding AI shouldn't give someone access to information they couldn't otherwise reach.

System routes

Automations and integrations can quietly create powerful routes into your systems. Give each only the access required for its job, and be particularly careful with anything exposed publicly.

Stop and think: can this route expose sensitive data? If it genuinely must, strong authentication, access controls, monitoring and audit become essential. If your team doesn't have that expertise, get specialist help or keep the data inside the protected environment.

MAKE THE SAFER CHOICE EASIER

Use the security environment you already trust

KJM builds business applications and records inside our established Microsoft environment, using Microsoft identity and access controls as the security foundation. The same principle can apply in a well-managed Google environment.

This doesn't remove your responsibility for security. You still decide who should have access, configure the permissions and design your application correctly. But it means you aren't asking AI to invent authentication, identity management and data security from scratch.

For most vibe-coded applications, start inside the protected environment and only open a route out when there is a clear reason to do so and you are confident about doing it safely.

If a feature needs considerably more security expertise than your team has, treat that as an architecture decision, not a prompt-engineering problem.

02 / PUBLIC AI

Keep public AI deliberately limited

A public AI assistant should usually need public information and very little else. Every additional connection increases what could be exposed or misused if something goes wrong.

Limit what it can access

Give public AI an approved source of public information. Don't connect it to internal business records simply to make the answers richer.

Stop and think: does the AI genuinely need access to this information?

Protect what people give it

Public AI may only retrieve public information, but visitors can still enter names, contact details or sensitive information themselves.

Stop and think: what happens to that information after it is entered, and is another copy being kept unnecessarily?

Limit what it can do

A public assistant might answer questions, collect an enquiry or start a process. Give it only the connections needed for those jobs.

Stop and think: could someone use this route to make the AI do more than you intended?

A safer pattern is to let public AI request an action, not control the result. For example, KJM's Kira can request information be sent to a client email already held in the protected system. Kira AI never retrieves the sensitive record or chooses the recipient.

If sensitive information must be returned directly through a public route, stronger security design and expertise are required.

THE SIMPLE RULE

Public AI should have very little to lose

Keep what it can reach and what it can do deliberately small. Treat information people give it as potentially private, even when the AI itself only uses public knowledge.

If the public AI genuinely needs access to sensitive information or powerful actions, stop treating it as a simple public AI service. That design needs the stronger security controls and expertise described in the previous section.

03 / INTERNAL AI

AI should only see what the user can see

Signing in establishes who someone is. It doesn't decide what the person is allowed to see or do. Their role and permissions do that. Internal AI should work inside exactly those same boundaries.

Apply permissions before AI sees the data

Don't give AI broad access and rely on instructions telling it what not to reveal.

First determine what the signed-in person is allowed to access. Only then should that information be available to AI.

Stop and think: could AI retrieve something the person couldn't open themselves?

Apply the same rule to actions

The same principle applies when AI does something rather than simply answering a question.

If someone's role cannot approve, delete, export or change something, asking AI to do it shouldn't create another route around that restriction.

Stop and think: could asking AI achieve something the person couldn't do through the application?

Access changes must apply to AI

If someone's role changes or access is removed, AI should immediately use the new permissions for anything it retrieves or does from that point onwards.

Stop and think: after access is removed, could the AI still retrieve fresh information through an old conversation or another route?

THE SIMPLE RULE

AI inherits permission. It doesn't grant it.

Internal AI can make information easier to find and work with. It should never make more information available.

The security system decides what the person can access. AI works within that decision.

Test both sides: confirm it works when access is allowed and fails when it isn't. If an AI instruction is the only thing stopping someone reaching restricted information, the security design is wrong.

04 / VERIFY THE DATA LIFECYCLE

Know where your data ends up

It's easy to design where information should go. The harder question is where copies may be left behind. Forms, AI conversations, automations, logs and other services can retain information without it being obvious to the user.

Know what becomes a business record

An enquiry or approved update may need to become a permanent record. The conversation around it usually doesn't.

Stop and think: what actually needs to be kept, and what should disappear once the process is complete?

Look for unnecessary copies

Information can be retained in more places than the system of record. Ask AI to help identify everywhere the information passes through and where another copy might be stored.

Stop and think: if the master record was deleted, where else might the information still exist?

Treat sensitive data differently

The more sensitive the information, the more important it is to understand where it travels, what is retained and for how long.

Stop and think: are you comfortable explaining to a client where their information goes and where copies may remain?

THE SIMPLE RULE

Keep the data once, wherever possible

Design the process to eliminate unnecessary copies of information. Where another copy is genuinely required, minimise what it contains, how long it is kept and who can access it.

A conversation, log or intermediate service shouldn't quietly become another store of sensitive information just because it is convenient.

For sensitive data, verify the actual services and settings you use. The goal is not simply to know where copies exist. It is to prevent unnecessary copies existing in the first place.