By AsymiLink Meta · Published
The decision, not the buzzword
Custom software used to mean a team writing every workflow. Agents now take the mechanical layer — CRUD, extraction, first-pass tests, routine orchestration. The work that remains is architecture, integration, compliance, and the parts of the system that carry business risk. The companies that get this right do not buy “an AI agent” or “a custom build.” They decide which of those two the problem actually is.
Agent on existing systems
Use this when the system of record is fine and the work is repetitive. Helpdesk, CRM, scheduling, internal knowledge. The agent is a project with a runbook, not a seat licence. Production window is typically 5–12 weeks. Ticket deflection in this work has landed at 82% on production deployments.
Custom software with agents inside
Use this when the workflow is the product. A new operations platform, a CRM that does not exist yet, a compliance surface that cannot be rented. Agents write the first pass on standard layers. Engineers keep architecture, integrations, and regulated logic. You own the repository.
Offline and sovereign
When data cannot leave the building, the same two choices still apply. The model and the application run on your hardware. That is a deployment constraint, not a seventh service.