The question we hear most often: how long does it actually take to go from 'we want to use AI' to a working system in production? The honest answer is 4 weeks, if you scope it right. The dishonest answer is 6 to 18 months, which is what happens when you don't.
Every fast AI implementation we've shipped follows the same structure: one workflow, a fixed 4-week window, real data from day one, and a binary go/no-go at the end. Here is the step-by-step.
Week 1: pick one workflow and map it completely
The single biggest reason AI projects stall is scope. Teams try to automate five things at once and end up with five half-built systems. Week 1 is about selection and mapping, nothing else.
Selection criteria for your Week 1 workflow: it runs at least 20 times per week, it currently takes a human 5 to 30 minutes, it has a clear input and a clear output, and the person doing it today would be relieved to hand it off. If you can answer yes to all four, you have a good starting candidate.
Mapping means documenting every step the human currently takes, every system they touch, every decision they make, and every edge case they handle. Do this before writing a line of code. The map reveals which steps are mechanical (automate these first) and which require judgment (automate these second, with a human-in-the-loop phase).
Week 1 deliverable: workflow map and data access
By end of Week 1, you need: a documented step-by-step map of the workflow, API credentials or data access for every system the workflow touches, a dataset of 50 to 200 real historical examples (inputs and correct outputs), and an agreed success metric (accuracy rate, time saved, tasks handled without human review).
Week 2: build the prototype on real data
Week 2 is pure build. The goal is a working prototype that handles the mechanical steps of the workflow end-to-end, with the AI steps handling the judgment calls. You are not optimising yet — you are getting from zero to something that runs.
Build sequence: connect the integrations first (API calls, data reads, system writes), then add the AI logic in the middle, then build the human escalation path for cases the AI can't handle confidently. Test against your historical dataset as you go — not at the end.
The most common Week 2 mistake is building the perfect system instead of a working one. Speed is the goal. You will tune accuracy in Week 3.
Week 3: validate accuracy and tune
Week 3 is testing and tuning. Run the prototype against your full historical dataset. Measure your success metric. Identify the failure modes — what kinds of inputs does it get wrong, and why?
Most failures in Week 3 fall into three categories: prompt issues (the AI is given ambiguous instructions), retrieval issues (for RAG-based systems, the wrong context is being injected), or edge case gaps (the workflow map missed an important scenario). All three are fixable in a few hours each.
Target accuracy before go-live: 90%+ on clearly scoped tasks, with a well-defined escalation path for the rest. If you hit 85% in Week 3, that is usually good enough to go live with a supervised phase.
Week 4: go live with a supervised phase
Go live does not mean remove humans from the loop. It means the AI is handling real production inputs, a human is reviewing the outputs, and you are measuring the system against your success metric in the real world rather than against historical test data.
The supervised phase typically runs 1 to 4 weeks post-launch, depending on volume and risk. As accuracy stays above threshold and the team builds confidence, the human review rate is reduced — from reviewing every output, to reviewing flagged outputs, to reviewing a random sample. By the end of the supervised phase, the system is operating autonomously with exception handling.
What makes a 4-week timeline possible
The 4-week timeline is only possible with three conditions: a single workflow (not five), real data access from day one (no waiting for a data team to prepare a clean dataset), and a go/no-go decision framework agreed in Week 1. Without all three, timelines expand to 6 to 12 months.
The fastest implementations we've shipped were not faster because they were simpler. They were faster because the scope was enforced. One workflow. Four weeks. Measurable result. Then expand.