Direct answer
Connect AI through a controlled workflow: a verified trigger starts the job, minimum necessary data is retrieved from approved sources, AI performs a bounded task, business rules and human review control the result, and the final record returns to the system of record with logs and exception handling.
System anatomy
Treat AI as one bounded component in a larger workflow
Trigger
A form submission, record change, message, schedule, or employee action starts the work.
Retrieve
The system obtains only the approved fields and source material needed for the job.
Reason or generate
A model classifies, extracts, compares, summarizes, drafts, or recommends within defined instructions.
Control
Rules, confidence conditions, and human approval determine what may happen next.
Write back
Approved output is recorded in the CRM, ERP, ticketing system, content system, or other source of truth.
Observe
Logs, outcomes, corrections, cost, and exceptions support monitoring and improvement.
Integration inventory
Define the role of every participating system
| Question | Why it matters | Evidence |
|---|---|---|
| What starts the work? | Prevents duplicate or accidental runs | Verified event and unique identifier |
| Where does truth live? | Avoids conflicting records | Named system of record for each field |
| What can be read? | Limits exposure | Field-level data and permission list |
| What can be changed? | Controls action risk | Approved write actions and rollback |
| What happens on failure? | Keeps work from disappearing | Queue, alert, owner, and manual fallback |
| How is activity traced? | Supports investigation and measurement | Logs connecting input, output, approval, and outcome |
Useful patterns
Choose the least powerful architecture that solves the job
- Read-only assistant: retrieves approved context but cannot change records
- Draft-and-approve: prepares work that a person verifies before sending or saving
- Rules with AI classification: AI interprets an input, while deterministic rules control the action
- Human by exception: routine low-risk cases proceed when conditions are met, with exceptions routed
- Tool-using agent: can take several bounded actions, requiring stronger permissions, monitoring, and recovery
Permissions
Use separate identities and minimum necessary access
Do not give an AI workflow the same broad access as an administrator or senior employee. Create a dedicated service identity where possible, restrict it to the required records and actions, and separate testing from production.
Credentials belong in secure secret storage, not prompts or source documents. Logs should record actions without unnecessarily copying sensitive content.
Failure design
Make incomplete work visible and recoverable
- Use idempotency or duplicate protection when actions may retry
- Define timeouts and what happens when a provider is unavailable
- Send uncertain cases to a queue, not into silent failure
- Preserve the original input and decision trace where appropriate
- Create rollback for write actions
- Assign an owner to integration changes and vendor updates
The value point
After this page, you should be able to decide:
Which systems must participate, which one remains authoritative, and where approval and exceptions belong.Your working output should be a trigger-to-record architecture, integration inventory, permission model, and failure path.
Questions business leaders ask
Frequently asked questions
Can AI work with our existing CRM or ERP?+
Often, yes, through supported APIs, automation platforms, exports, or controlled user interfaces. Feasibility depends on the specific product, plan, permissions, data quality, and required actions.
Do we need to replace old software before using AI?+
Not necessarily. A focused system can work around stable existing tools, but missing APIs, poor data, or unsupported software may limit reliability and increase maintenance.
Should AI write directly into the system of record?+
Only when the action is bounded, tested, reversible, logged, and appropriate for the risk. Draft-and-approve is often the safer starting pattern.
What is the biggest integration risk?+
Excessive permissions and unclear failure handling are major risks. A useful integration also needs a reliable trigger, source-of-truth rules, duplicate protection, monitoring, and ownership.
Research anchors
Primary and authoritative sources
Examples and planning ranges are clearly labeled. Source terms, provider behavior, and regulations can change; verify current requirements for your organization and jurisdiction.
Prepared and reviewed by the Future Made Useful systems editorial team. Material guidance reviewed July 17, 2026.