It depends mostly on the depth you need, not on the technology. A read-only agent can be live in weeks. An agent that drafts for a person to send usually needs two to three months, because shadow and certification must cover a full business cycle. Acting alone within bounds takes longer, because it is earned on production evidence.
What drives the timeline?
Building the agent is rarely the slow part. A capable team can describe the job and connect a model in days. The time goes elsewhere:
- Access to systems. Read-only credentials to a core system can take weeks of security review.
- Process clarity. If nobody can say what a correct outcome is, certification cannot start.
- Past cases. You need a few hundred real cases with known outcomes to certify against.
- The business cycle. Shadow work must include a month-end, a billing run or whatever peak your process has.
- Approvals. Data protection, risk and sometimes a works council or a regulator need to be satisfied.
- Target depth. Every step up needs fresh evidence.
What does a typical plan look like?
| Stage | What happens | Who leads | Plan for |
|---|---|---|---|
| Scope and owner | One process, one owner, success criteria | Process owner | 2 to 4 weeks |
| Access and sources | Credentials, connections, source provenance | Platform owner | 2 to 6 weeks, often in parallel |
| Brief and case set | Brief written, past cases and answer keys gathered | Practitioner and analyst | 2 to 4 weeks |
| Shadow | Agent at depth 1 on live work | Process owner | 4 to 8 weeks |
| Certification | Replay against the floor, fixes, rerun | Process owner and reviewers | 1 to 3 weeks |
| Production at depth 2 | Drafts, a person sends | Process owner | 4 to 8 weeks before any widening |
| Depth 3 for a slice | Acts within bounds on routine cases | Process owner and risk | After fresh certification |
These are planning ranges, not promises. A process with clean data, a strong owner and a short cycle will move faster. A process that touches money, customers' rights or several systems will move slower, and should.
Why do timelines differ so much by function?
Because the work differs. A support agent answering policy questions can often reach depth 2 quickly, since the facts sit in a policy library and the outcome is easy to judge; the specifics are in how long it takes to deploy an AI support agent. A finance agent posting journals or matching payments meets month-end controls, auditors and ledger access, which stretch every stage; see how long it takes to deploy an AI agent in finance.
What does a realistic timeline look like in practice?
Say an insurer's broker services team wants an agent to answer brokers' questions about policy status and draft endorsement requests. Week one, the owner agrees scope. Access to the policy administration system takes five weeks of security review, during which the practitioner writes the brief and the analyst assembles 300 past queries with answers.
Shadow runs for six weeks to cover a quarterly renewal peak. Certification finds the agent misreads endorsements on older policy forms; after a brief change and rerun it clears. The agent goes live at depth 2 in week 14. After two months of low edit rates, a fresh certification moves policy-status answers to depth 3, around week 23. Endorsement drafts stay at depth 2, which is where the owner wants them.
What should you have ready before you start the clock?
- A named owner with time set aside.
- Agreement on what a correct outcome is.
- A route for people to take work back from the agent at any stage, tested before go-live.
- A data protection screening.
- A calendar showing your next business peak, so shadow can include it.
The full sequence is in our guide to deploying an AI agent.
Where Pryme fits
Agent Studio turns a plain-language job description into a candidate brief, which shortens the drafting stage. It does not shorten the evidence. Attest replays your own cases, and depth follows the result and the role's ceiling, whichever is lower. That part takes as long as your process needs.
Questions people ask
Can we go faster by skipping shadow?
You save weeks and lose the only view of the agent on live work before it affects anyone. Shorten it if your process is simple, but do not drop it.
Does a vendor's pre-built agent deploy faster?
It can shorten the brief and connections. It does not shorten certification on your cases or shadow on your work, which is where most of the time goes.
When should we expect the first benefit?
At depth 2, once reviewers trust the drafts enough to edit lightly. That usually comes a few weeks after go-live, not on the first day.