The honest answer most US teams want: a working AI agent prototype in 14 days, production-ready in 4–6 weeks. The dishonest answer is "it depends" — which is true but useless. Here's what actually moves the timeline.
A custom AI agent typically takes 14 days to a working prototype and 4–6 weeks to production. The single biggest variable is not the model — it's the cleanliness of your data and the clarity of your evaluation criteria.
The realistic 2026 timeline
Across the AI agent projects we've shipped at Zero Friktion in 2025–2026, the pattern is consistent. Here's what a typical schedule looks like:
- Days 1–2 — Map the friction. Sit with the operator. Watch the actual workflow. Write down the costliest 3 frictions in plain English.
- Days 3–10 — Prototype. One-week sprint to a demo-able agent that addresses the highest-leverage friction. Real data, real tools, no mocks.
- Weeks 3–4 — First production deploy. Ship to a small subset of real users. Start collecting eval data and observability.
- Weeks 5–6 — Hardening. Tighten the eval set, add retries, define cost ceilings, train the human-in-the-loop interface.
What actually slows it down
The model isn't the bottleneck — pricing has commoditized, and Claude, GPT-4-class, and open models are all production-ready. The slowdowns come from elsewhere:
1. Unclear evaluation criteria
If you can't tell us in week one what "good" looks like — measurable, with examples — we can't tell you when the agent is done. Teams that walk in with a written eval set ship 2–3× faster.
2. Messy or fragmented data
Retrieval-augmented agents only work if there's something to retrieve. If your operational knowledge lives in three CRMs, a Notion wiki, and the heads of two senior employees, expect to add a 1–2 week data-cleaning phase before anything useful.
3. Unscoped human-in-the-loop
Most production agents need a human approval step. The interface for that approval is a real product surface — not an afterthought. Plan for design and engineering time on it from week one.
4. Tool integrations with rate limits or auth quirks
Salesforce, NetSuite, custom internal APIs — these eat days. Not because they're hard, but because nobody owns them anymore inside the client org.
What does not slow it down (anymore)
A few things that used to be timeline-killers in 2023–2024 have stopped mattering:
- Model choice. Picking between Claude, GPT, or Gemini is a one-day decision, not a one-month evaluation.
- Hosting infrastructure. Vercel, Modal, or a small Fly app handles 95% of agent deployments. No Kubernetes required.
- Vector databases. Pinecone, pgvector on Postgres, or Turbopuffer — pick one in an hour, move on.
Three timelines for three real projects
Project A — Logistics dispatcher copilot
14 days to a working agent that drafts the night handoff. 9 weeks to production (including a custom mobile driver app). Slowed by GPS data normalization across two tracking vendors.
Project B — Healthcare patient-intake agent
10 days to a voice + chat triage prototype. 5 weeks to a clinic pilot, then 4 more weeks for HIPAA-compliant deploy hardening.
Project C — Financial reconciliation agent
21 days to first prototype (longer because the source data was in eight half-used SaaS subscriptions). 6 weeks to production.
How to compress your own timeline
- Walk in with the workflow recorded. A Loom of someone doing the job by hand is worth ten meetings.
- Pre-write 20 evaluation examples. Inputs and the output you'd accept. This is the single most leveraged thing you can do.
- Cut scope ruthlessly. One workflow. One user persona. Ship it. Then expand.
- Pick the human-in-the-loop early. Decide who reviews the agent's output and what their UI looks like before week three.
Bottom line: in 2026, the engineering for a production AI agent is not the slow part. The slow part is figuring out what you actually want it to do, with what data, judged by what standard. Solve that in week one, and 4–6 weeks to production is realistic.