The safest way to integrate AI into existing systems is in layers. First use AI alongside your systems, with people moving information between them. Then connect AI with read-only access, so it can see but not change anything. Only once that is proven should AI be allowed to write back or trigger actions, and even then with human approval for anything important. Each layer adds value while limiting what can go wrong.
Layer 1: Alongside (no integration)
People use an AI assistant separately, copying in what it needs and pasting results back after checking.
- Risk: low, because AI cannot touch your systems.
- Value: high for drafting, summarising and analysis.
- Main control: clear rules on what data people may paste in.
Many organisations get most of their early value here, with no IT project at all.
Layer 2: Read-only connection
AI can access selected information, such as documents, a knowledge base or CRM records, but cannot change it.
- Risk: moderate, mainly around who can see what.
- Value: AI answers questions from your real information.
- Main controls: permissions that mirror existing access rights; clear-out of outdated content; testing on sample questions.
Layer 3: Write-back with approval
AI drafts updates or actions, such as a CRM note, a ticket response or a scheduled task, but a person approves before anything changes.
- Risk: moderate to high, depending on the system.
- Value: removes manual data entry and speeds up workflows.
- Main controls: approval steps, audit logs, rollback options.
Layer 4: Automated action
AI acts within defined limits without approval for each action, for example routing tickets or updating standard fields.
- Risk: highest.
- Value: large time savings at scale.
- Main controls: tight limits, monitoring, exception handling, regular review.
Only move to this layer for well-understood, low-impact actions with a proven track record.
Practical safeguards at every layer
- Test in a sandbox or copy of the system before touching live data.
- Start with one workflow rather than connecting everything.
- Keep a manual fallback so work continues if the AI component fails.
- Log what AI does, especially once it can change records.
- Involve the system owner from the start; they know where things break.
Common integration mistakes
- Connecting AI to a shared drive full of outdated documents, so it confidently quotes old policies.
- Giving AI broader access than the users it serves.
- Skipping Layers 1 and 2 because Layer 4 looks more impressive.
- Having no owner for the integration once the project team moves on.
Who should lead integration
IT or a technical partner handles the connections, but a business owner should decide what AI is allowed to do and check that the new workflow actually helps. Our earlier article on integrating AI with existing software covers tool compatibility; this guide focuses on the staged approach. For the people side, see how to lead AI change without breaking what already works.
Our Leading AI Innovation & Organisational Transformation course covers simple, no-code ways to connect AI to tools you already use, alongside the leadership skills to roll changes out safely.
Frequently asked questions
Do we need IT to start using AI?
Not for Layer 1, beyond approving tools. For Layers 2 to 4, IT or a technical partner should be involved.
How do we stop AI seeing information it should not?
Make sure AI inherits the same permissions as the user, and test with sample questions before rollout.
What if an AI integration causes an error in our records?
That is why write-back should start with approval steps and audit logs. Keep the ability to roll back changes and a manual fallback process.