Most businesses already run on software they do not want to replace. AI integration means connecting model capabilities to those systems: enriching CRM records, classifying ERP transactions, adding smart search to a portal, or generating documents from existing data. It is the fastest path to value for companies with established systems.
Integration patterns
- Enrichment: when a record is created or updated, a model adds a summary, category, sentiment or extracted fields.
- Assist-in-place: a button inside the existing UI drafts, summarises or explains using the current record as context.
- Search and Q&A: a retrieval layer over documents and records, exposed inside the tools people already use.
- Event-driven pipelines: documents or messages arrive, get processed by a model and land in the right system with confidence scores.
Technical considerations
Authentication and permission mapping between systems, rate limits and retries, idempotent writes, prompt and response logging, PII handling and data residency, versioning of prompts alongside code, and cost attribution per feature.
Choosing model providers
Choose per task. Extraction and classification often work with smaller, cheaper models; complex reasoning or long documents may justify larger ones. Design an abstraction so providers can be swapped as prices and capabilities change.
My perspective
Integration work is unglamorous and enormously valuable. A model that reads every inbound email and updates the CRM correctly will change how a sales team works more than any standalone AI tool.
Business use cases
- CRM enrichment: summaries, categorisation and next-step suggestions on every lead
- ERP and accounting: transaction classification and anomaly flags
- Website and portal search upgraded to natural-language Q&A
- Document generation from existing records (proposals, letters, reports)
- Email and messaging pipelines that extract and route information
- Legacy data migration with AI-assisted mapping and cleaning
Limitations to be honest about
- Legacy systems may lack APIs; integration effort can dwarf the AI component.
- Data quality problems surface immediately and must be addressed.
- Vendor terms around data use vary; review them before sending sensitive records.
- Silent failures in pipelines are dangerous — monitoring is not optional.
How I approach it
- I inventory the systems, their APIs and the data that flows between them, then select one integration with clear volume and pain. It is built with logging, retries and a feedback channel, and rolled out to a subset of users before expanding.
Frequently asked questions
Can AI be added to our existing CRM or ERP?
Usually yes, via APIs, webhooks or middleware. The AI layer sits alongside the system and writes results back into standard fields.
Which AI API should we use?
It depends on the task, data sensitivity and budget. A provider-agnostic design lets you choose per feature and switch later.
How do you protect our data when using AI APIs?
By minimising what is sent, redacting PII where possible, choosing providers with appropriate data terms and logging every request for audit.
Last updated 11 September 2026