July 25, 2026 AI Strategy Enterprise AI 12 min read

The 90-Day AI Pilot Playbook: Week-by-Week for Enterprise Teams

Most enterprise AI pilots fail not because the technology doesn't work, but because the pilot was structured to explore rather than decide. This is the playbook that fixes that.

A 90-day window is enough time to produce decision-quality evidence on almost any enterprise AI use case. It is not enough time to build a complete system, resolve every integration, or satisfy every stakeholder concern. The teams that confuse those two objectives are the ones that find themselves in month seven with a pilot that never formally ended and no one willing to make a call.

What follows is a week-by-week structure for running a 90-day AI pilot that produces a clean go/no-go decision at the end. It assumes a motivated team of five to seven people, a defined use case, and an executive sponsor who has agreed to attend two meetings: a kickoff and a decision gate. Everything else is execution.

Before Week 1: The Three Prerequisites

No playbook survives contact with a pilot that skipped its prerequisites. Before the 90-day clock starts, three things must be true:

The use case is scoped to a single workflow. Not "improve customer service" but "reduce first-response time on tier-1 support tickets." Not "automate compliance" but "flag change-of-control clauses in vendor contracts during quarterly review." The narrower the scope, the cleaner the evidence at the end.

Success criteria are defined and signed off. This is the most important prerequisite. The executive sponsor must agree, in writing, on what numbers constitute success before the pilot begins. If you are still negotiating success criteria in Week 8, you do not have a pilot — you have a research project.

Data access is confirmed, not assumed. The most common cause of a stalled pilot is discovering in Week 3 that the data needed to train, fine-tune, or evaluate the system requires approvals that take six weeks to obtain. Confirm access before the clock starts. See the data strategy guide for the access questions to resolve upfront.

Phase 1 — Weeks 1–4

Scope, Baseline, and Data Readiness

Week Primary Deliverable Key Activities
Week 1 Pilot charter signed Kickoff with exec sponsor. Document use case, success criteria, team, timeline, and go/no-go thresholds. Freeze scope.
Week 2 Human baseline established Measure current process performance: time per task, error rate, throughput. This is your comparison baseline. Without it, you cannot claim improvement.
Week 3 Data audit complete Inventory all data sources. Confirm access, assess quality, identify gaps. Flag any PII or data governance issues that affect model training or evaluation.
Week 4 Architecture decision Choose your approach: prompt engineering only, RAG, fine-tuning, or agent with tools. Document the rationale. Begin environment setup. See fine-tuning vs RAG if this decision is not yet settled.
Common Phase 1 Failure

Teams that skip the human baseline in Week 2 cannot demonstrate improvement in Week 12. The exec sponsor will ask "better than what?" and you will not have an answer. Measure the current state before touching any model.

Phase 2 — Weeks 5–8

Build, Integrate, and Iterate

Week Primary Deliverable Key Activities
Week 5 Working prototype First end-to-end version running on real data. Ugly is fine. The goal is a complete path from input to output, not a polished product.
Week 6 Internal evaluation v1 Run the prototype against a held-out evaluation set. Score against the success criteria defined in Week 1. Identify the biggest gaps.
Week 7 Integration checkpoint Connect to production data sources and downstream systems. Surface latency and throughput issues. Confirm cost per transaction is within budget.
Week 8 Domain expert review Have subject matter experts from the business unit review 50 to 100 model outputs. Capture qualitative feedback on errors, edge cases, and trust. This surfaces issues that eval metrics miss.

Week 8 is the informal midpoint decision. If the prototype is not within striking distance of the success criteria, the team should surface this to the exec sponsor now rather than in Week 12. A course correction at Week 8 is manageable. A surprise at Week 12 is not.

Phase 3 — Weeks 9–12

Evaluation, Evidence, and Decision Gate

Week Primary Deliverable Key Activities
Week 9 Evaluation v2 on full dataset Run the complete evaluation suite. Compare accuracy, latency, and cost against the Week 2 baseline and the Week 1 success criteria.
Week 10 Risk and compliance review Legal, security, and compliance sign-off on data handling, model outputs, and audit trail. For regulated industries, this must happen before any go decision. See the AI risk management framework.
Week 11 Pilot report drafted Document results against every success criterion. Include failure cases and edge case inventory. Draft the go/no-go recommendation with supporting evidence.
Week 12 Executive decision gate Present to exec sponsor. Go: define production scope, budget, and timeline. No-go: document what would need to change. Conditional: define the specific gaps and re-pilot scope.
Week 12 Decision Gate

What the exec sponsor needs to decide

The Team You Actually Need

A 90-day pilot does not require a large team. It requires the right five people with sufficient time allocation.

The exec sponsor attends Week 1 and Week 12. That is all that is required. More frequent exec involvement tends to shift the pilot toward stakeholder management and away from execution.

The Success Criteria Template

Every pilot charter should include four categories of success criteria, all defined and signed off before Week 1:

1. Task accuracy

Compared against a human expert performing the same task. Define the evaluation method, the dataset size, and the acceptable threshold. Example: "Correctly classify contract risk level with precision above 88% on a 200-document held-out set, compared to legal team baseline of 91%."

2. Latency

The system must respond within the time constraints the workflow requires. A contract review system that takes 45 seconds per document may be acceptable. One that takes 8 minutes is not. Set the threshold before the pilot.

3. Cost per transaction

Define the maximum acceptable inference cost per unit of work. This constrains model selection and architecture. See the full cost and ROI framework for how to build this estimate.

4. Business outcome signal

At least one metric that connects to a business result: hours saved per week, error rate reduction, throughput increase, cycle time compression. This is what the exec sponsor cares about. The technical metrics above are necessary but not sufficient.

Scope Discipline: The Hardest Part

The most common way a 90-day pilot fails is not technical failure — it is scope expansion. A pilot that started as "flag change-of-control clauses in vendor contracts" becomes "flag all compliance risks across all contract types" by Week 6. The original success criteria no longer apply. The evaluation dataset no longer covers the full scope. The Week 12 decision gate arrives with no clean evidence to present.

Scope changes during the pilot must go through the exec sponsor. This is not bureaucracy — it is the mechanism that keeps the pilot decision-ready. Any change to scope resets the success criteria and, if significant, resets the timeline.

If a genuinely important additional use case surfaces during the pilot, log it. Build it into the production roadmap if the pilot goes. Do not build it into the pilot itself.

What a Go Decision Actually Unlocks

A go decision at Week 12 is not a production deployment. It is authorization to build toward production. The gap between pilot and production is real and consistently underestimated. Reliability requirements, monitoring, human oversight, audit trails, and change management all expand significantly in production.

The pilot produces the evidence that the use case is viable, the technology works at the required quality level, and the business case justifies investment. The full enterprise pilot guide covers what happens after the go decision, including the production scoping, team scaling, and governance structure that a successful pilot requires.

What goes wrong after a go decision

Teams that receive a go decision often misread it as permission to scale immediately. Production requires monitoring infrastructure, fallback handling, human-in-the-loop workflows for edge cases, and a change management plan for the people whose work is changing. Build that plan in parallel with the pilot's final weeks, not after the decision gate.

The Week-by-Week Checklist

Print this. Put it on the wall. Check it every Monday.

Phase 1: Scope and Baseline (Weeks 1-4)

Phase 2: Build and Integrate (Weeks 5-8)

Phase 3: Evaluate and Decide (Weeks 9-12)

The 90-day structure works because it forces the decisions that pilots typically avoid: what does success look like, can we access the data, does the system actually perform at the required level, and does the business case hold. Teams that run pilots without this structure tend to drift toward exploration rather than decision. The 90-day playbook is not a constraint on ambition. It is the structure that turns ambition into a decision executives can act on.

Running a pilot and want a second opinion on scope or success criteria? A 15-minute conversation is usually enough to identify the structural issues before they surface at Week 12.

Schedule 15 minutes →