The integration is the project.
The agent is a fortnight of work. Getting it to talk to the systems you already have is the rest of it.
Lede
Clients are often surprised by where the hours go. The reasoning part — the bit that reads a document and makes a judgement — is frequently the smallest line in the estimate. The expensive part is everything around it.
This is not a complaint. It is the single most useful thing to understand before you budget for an agentic project.
The four hard integrations
The system with no API. Someone's ERP, a vendor portal, a mainframe with a green screen. There is always at least one. The options are all bad, they all cost weeks, and choosing between them is a business decision about acceptable fragility, not a technical one.
The system with an API and no access. Technically trivial, organisationally slow. The connection needs a service account, the service account needs an approval, the approval needs a security review, and the security review needs a person who is on leave. We now start this in week one of every engagement regardless of when we need it.
Identity. Who is the agent acting as? If it acts as a service account, your audit log says a robot did everything and you have lost the ability to answer "who approved this." If it acts on behalf of a user, you need delegation your identity provider may not support cleanly. This one is worth deciding deliberately rather than by accident.
The human handoff. Getting an exception in front of a person, in the tool they already use, with enough context to decide in thirty seconds. Nobody wants another dashboard, so this usually means writing into a queue, a ticketing system, or a channel — and that integration is where adoption is won or lost.
Scoping around it
We ask about integrations before we ask about the workflow. A painful workflow attached to two clean APIs is a straightforward engagement. A moderately painful workflow attached to a system nobody has credentials for is a research project wearing a delivery costume.
It also changes which workflow you should automate first. The obvious candidate is the one causing the most pain. The right first candidate is usually the one where the pain is real and the integration surface is boring — because shipping something that works buys you the organisational goodwill to attempt the hard one next.
The reasoning layer keeps getting cheaper and better without you doing anything. The plumbing does not. Budget accordingly.