Quick answer
A real US enterprise AI build runs 14-28 months end-to-end, not the 6-8 weeks most vendors quote. The full timeline: 2-4 weeks scoping, 2 weeks eval set, 4-8 weeks MVP, 4 weeks shadow mode, 8-12 weeks gated rollout, then 12+ months of scale and audit-prep. Deloitte's State of AI 2026 reports that 42% of US companies scrapped their AI initiative — almost always because they treated the 6-week MVP as the finish line and skipped shadow mode, gated rollout, and the eval pipeline. From 150+ enterprise deployments since 2023, the buyers who hit the 14-28 month plan ship products that survive an audit.
Why 42% of US AI projects failed in 2025
Deloitte's State of AI 2026 report tracked 1,800 US enterprises through their 2025 AI initiatives. 42% of those projects either ended without reaching production or were quietly retired in Q4. The single most common reason in their post-mortems: "the timeline was set by vendor demo, not by buyer reality."
In our own client portfolio — 150+ enterprise AI engagements across US, India, and EU since 2023 — we see the same pattern. The deployments that survived a 12-month audit cycle followed a 14-28 month timeline. The ones that quietly died followed a vendor's 8-week MVP plan and treated the MVP as the finish line.
The 6 real phases (and how long each actually takes)
- Scoping — 2 to 4 weeks. Workflow mapping, stakeholder interviews, risk register. Honest scoping kills 1 in 4 projects before any code is written. That is the point — better to kill the wrong project early than ship a wrong product.
- Eval set — 2 weeks. Your domain experts label 150-300 test cases with the correct expected behavior. We do not write a single prompt until the eval set is signed. Skipping this is the #1 reason vibes-tested AI ships broken.
- MVP build — 4 to 8 weeks. The thing most vendors call "the project." It is one of six phases, not the whole thing.
- Shadow mode — 4 weeks. Agent runs in parallel with humans for every real production input. We compare outputs daily. Disagreements drive prompt patches. This is where vibes-based AI gets exposed.
- Gated rollout — 8 to 12 weeks. 10% → 50% → 100% traffic with hard rollback criteria written before traffic flips. Material-threshold checkpoints require human approval at every step.
- Scale + audit — 12+ months. SLA monitoring, weekly prompt patches, monthly cost review, quarterly audit-evidence package. This is where enterprise AI either becomes durable or quietly dies.
Need an honest scoping call? Bring your use case + your audit constraints. We come back with a phased plan + an honest "this might not be a fit" assessment if it is not. Free 30 min.
Book a free 30-min call →What happens when buyers skip phases
| Phase skipped | What typically breaks |
|---|---|
| Eval set | No baseline to measure regressions against. Every prompt tweak is a guess. |
| Shadow mode | Production gets the first test. Hallucinations or wrong tool calls hit real users. |
| Gated rollout | Hard cut-over. If the first day goes badly, you have a public incident not a rollback. |
| SLA + audit-evidence prep | Year-2 SOC 2 audit fails. Or your healthcare AI gets OCR-flagged. Or you cannot answer a buyer's security questionnaire. |
The three timeline buckets we see in practice
- Focused use case (14-18 months): single workflow, single integration, 10-15 user count. Customer support agent, internal Q&A bot, prior-auth drafter. ~$80K-$180K total engagement cost.
- Multi-tool suite (18-22 months): 3-5 connected agents, 6-10 integrations, eval use shared across agents. AP automation + fraud detection + reporting. ~$180K-$300K total.
- Enterprise platform (22-28 months): Multi-entity scope, full audit logs, on-call rotation, dedicated US-based PM. SOX-compliant finance AI, HIPAA-scope payer AI. ~$300K-$600K total.
How to compress the timeline without skipping phases
We have shipped 14-month enterprise builds when the buyer brought three things to the table on day one: (1) a clean data source already in a usable API, (2) a dedicated internal owner for the eval set, and (3) executive air cover for the shadow-mode phase. Without all three, the timeline naturally drifts to the higher end of the range. With all three, you can shave 4-6 months off and still ship audit-clean.
FAQ
Can we ship in 8 weeks?
A demo, yes. A production-grade AI agent that survives a 12-month audit cycle, no. The 8-week MVP is one phase of six. If a vendor quotes you 8 weeks total, they are either skipping phases or planning to ghost you after launch.
What kills AI projects most often?
Skipping the eval set and skipping shadow mode. Both are unsexy. Both are where good AI is built. The 42% scrap rate Deloitte reported tracks almost perfectly with how many buyers were quoted a timeline that omitted these phases.
Should we use an off-the-shelf product instead?
If your use case is generic (vanilla support chatbot, vanilla document summary), yes. Off-the-shelf wins on speed and cost. Custom builds are worth the 14-28 months only when the workflow is specific enough that off-the-shelf has a clear gap.
Does this apply to small businesses too?
The phases apply. The timelines compress. A small business focused-use-case build runs 4-8 months instead of 14-18. Eval set and shadow mode still mandatory; just smaller eval sets and shorter shadow windows.
Honest scoping
Get a phased plan, not a sales pitch
Bring your use case + your audit constraints. We come back with a phased plan, a written timeline, and an honest "this might not be a fit" assessment if the use case does not justify the build. 30 minutes, free, no sales sequence.
Methodology
Project counts (150+), failure rates, and phase timelines compiled from Braincuber's enterprise engagement records 2023-2026 and cross-validated against the Deloitte State of AI 2026 report (n=1,800 US enterprises), the McKinsey State of AI 2025 (n=1,491 global firms), and the IDC FutureScape: Worldwide AI 2026 Predictions. The 42% scrap-rate figure is Deloitte 2026; our internal failure-by-skipped-phase analysis pulls from 23 documented post-mortems. Pricing reflects our published 2026 rate card.
Founder and CEO of Braincuber. Has scoped and shipped 500+ Odoo, AI, and cloud projects for US mid-market and global brands. Takes every founder call personally — no SDR layer between buyers and the people building the system.