From Pilot to Production.

Short answer: a typical Teragrid Agent pilot reaches live production work in four calendar weeks. The timeline below is what we actually run with Malaysian enterprise clients — the deliverables per week, the decisions you (the buyer) need to make, and the failure modes that push timelines from four weeks to twelve.

Week 0 — Scoping (5 working days, before contract)

Before any contract, we run a free two-conversation scoping. The output is a written one-pager describing the proposed agent persona, the systems it will touch, the human escalation path, and the success metric. If scoping reveals the engagement is the wrong shape — too broad, too unbounded, too uncertain — we say so. Saying no early is the discipline that keeps the four-week timeline honest.

Week 1 — Identity, scope, and audit shell

Deliverables by end of week 1:

• Named service principal created in your IdP for the agent.
• Read-only access to source systems wired and tested.
• Per-action audit trail writing to your chosen storage.
• Empty agent shell deployed to a staging environment.
• Provenance manifest (which fact came from which source).

Failure mode this week: IT can't produce a service principal because nobody has done it before. Mitigation: name the IdP owner in the contract and put their sign-off on the critical path.

Week 2 — Tool wiring and supervised first runs

Deliverables by end of week 2:

• Tool catalogue (the functions the agent can call) typed and validated.
• Read-only tools producing reasoning traces for the supervisor.
• First end-to-end run on synthetic data — full provenance, no writes.
• Reasoning-bound configuration set (max loop depth, drift tolerance).

Failure mode this week: an upstream system's API returns inconsistent types under load, which destabilises tool calls. Mitigation: pin tool inputs at the type layer; treat upstream as untrusted even if it's internal.

Week 3 — Live data, human-in-the-loop

Deliverables by end of week 3:

• Agent running against live data, every action surfaced for human approval.
• Outcome telemetry feeding the supervisor dashboard.
• First refinement of agent persona based on observed misses.
• Write tools enabled but gated behind explicit human approval per action.

Failure mode this week: stakeholders "decide" late that the agent should also handle adjacent work. Mitigation: scope change → scope change document → not a free addition. Most pilots fail here from soft scope creep.

Week 4 — Autonomous cutover

Deliverables by end of week 4:

• Approval gates loosened on low-risk action classes.
• Human supervisor reviewing samples rather than every action.
• Production audit trail with cryptographic provenance.
• Runbook for incident response, rollback, and escalation.
• Sign-off document for go-live.

Failure mode this week: the supervisor reviews every action anyway, defeating the productivity gain. Mitigation: define "what's a sample" in advance and put a percentage on it.

What week 5+ looks like

We stay engaged. Agents drift, source systems change, regulations move. Most clients keep AITG on a thin retainer for ongoing operating partnership — tuning, audit support, persona evolution. We are deliberate that licence-and-leave is not the model.

When the four weeks become twelve

Three patterns extend pilots beyond a month: (1) procurement requires a tender, (2) the IdP setup takes longer than week 1, or (3) source-system access is restricted by an unrelated security review. None of these are agentic-AI problems specifically; they are enterprise-IT realities. We flag them in scoping and adjust the start date — not the deployment shape.

What to ask before kicking off

1. Who is the agent's named human escalation contact, by name?
2. Which two source systems will be wired first?
3. What does the supervisor see on day 1, day 7, day 30?
4. What is the rollback procedure, and who can invoke it?
5. Where do audit logs go, and who has read access?

If you can answer five questions before week 1 starts, the four-week timeline holds. If you can't, scoping needs another conversation before the contract.

Curious whether your use case fits the four-week shape? Start a scoping conversation — we'll tell you honestly whether it does, or what would need to change for it to.