Buying AI services
How long does it take to build an AI system?
Short answer
A focused AI system reaches production in 6–12 weeks: roughly 1–2 weeks to scope and assess, 2–3 weeks to a working prototype on your real data, and 3–7 weeks to harden it into something that runs unattended. Timelines slip past this almost entirely for non-technical reasons — data access approvals, security review, and no single decision-maker — rather than because the model work is hard. Treat any proposal quoting under four weeks to production as skipping evaluation, and anything quoting over six months as a programme rather than a project.
Last updated August 14, 2026 · Bitfumes AI consultancy
The realistic schedule
| Stage | Duration | What exists at the end |
|---|---|---|
| Assessment & scoping | 1–2 weeks | A costed plan, a chosen use case, a success metric |
| Prototype on real data | 2–3 weeks | Working software, honest accuracy numbers, an evaluation set |
| Hardening to production | 3–7 weeks | Monitoring, fallbacks, access control, cost controls, a runbook |
| Steady state | Ongoing | Evaluation on every change, drift monitoring, iteration |
The prototype stage is the one people expect to be slow and the hardening stage is the one they forget to budget. Getting to an impressive demo is genuinely fast in 2026; making it safe to leave running is where the engineering is.
What actually decides the schedule
- Data access — not data quality. Waiting three weeks for credentials to a system is the single most common delay, and it is entirely a calendar problem.
- Number of approvers. One decision-maker ships in eight weeks; a steering committee ships in eight months, building the same thing.
- Whether an evaluation set exists. Without it, every review meeting is an argument about vibes and the schedule stops being predictable.
- Integration surface. Reading from one system and writing to one system is a project; orchestrating six is a programme.
How to compress it
- Get data access provisioned during the scoping week, before engineers are waiting on it.
- Name one accountable owner who can approve scope without a committee.
- Ship to 5% of users behind a flag rather than waiting for a company-wide launch.
- Cut the second use case. Two half-built systems take longer than two sequential ones and prove nothing.
- Agree the success metric in writing before the build, so 'is it good enough' is a measurement rather than a debate.
Frequently asked
Can we get an AI prototype in two weeks?
Yes, if the data is already accessible and the use case is scoped — a working prototype on your real data in two to three weeks is normal. What you will not have in two weeks is monitoring, fallbacks or evidence it holds up on edge cases.
Why do AI projects take longer than expected?
Because the demo is fast and the hardening is not. Teams estimate against the prototype they saw in week three and are then surprised by the evaluation, monitoring, access control and cost work that turns it into something that can run without supervision.
How long does a RAG system take to build?
Two to six weeks for a production internal RAG assistant with senior engineers. The variable is document access and permissions, not the retrieval code.
What is the fastest path from idea to production AI?
A one-week paid assessment to pick the use case and prove the payback, then a single-use-case build behind a feature flag with an evaluation set from day one. Most delay comes from building the wrong thing carefully.
Related answers