Adoption decides this, not architecture
The strongest argument for putting AI inside your existing systems is behavioural: people already open them. A new application, however capable, competes for attention with everything else and usually loses.
We have built excellent standalone tools that ended up unused because they were one destination too many.
Put it where the work happens
- Support answers inside the helpdesk, not in a separate window
- Document extraction feeding the finance system directly
- Drafting inside the CRM where the record already is
- Knowledge retrieval inside the chat tool the team lives in
The test: would using this require someone to open something they do not currently open? If yes, adoption is now your main risk rather than accuracy.
When standalone is right
- The workflow does not exist yet, so there is nothing to embed it in
- The users are outside your business — customers, suppliers, applicants
- The existing systems genuinely cannot be extended
- You intend to sell it as a product in its own right
The cost difference
Embedding into existing software is usually £15,000–£40,000. A standalone application with accounts, roles and its own interface starts around £40,000 and rises quickly.
That gap buys authentication, permissions, an interface, hosting and everything else the host system already provides.
The hybrid that often wins
Build the capability as a service with a clean interface, then surface it in the systems people already use. If a standalone application is needed later, it consumes the same service.
That way the expensive part — the pipeline, evaluation and correction loop — is built once.