Ninety days is long enough to learn whether an AI pilot can become a production workflow. It is not long enough to transform the company, replatform every system, or settle every AI governance debate.
That constraint is useful. It forces the company to pick one workflow, one owner, one metric, one context problem, and one production gate. Mid-market companies do not usually fail because they lack AI ambition. They fail because they try to turn a pilot into a platform before the first workflow has earned trust.
McKinsey’s 2025 State of AI shows the gap clearly: 88% of organizations report regular AI use in at least one function, but about two-thirds are not scaling enterprise-wide and only 39% report EBIT impact. DORA’s 2025 AI-assisted software development report puts the operating lesson in sharper terms: AI amplifies the organization’s existing strengths and weaknesses, so returns come from the work system around the tools. Metacto’s Opportunity Mapping, Context Engineering, and AI Agents & Workflows map naturally to a 90-day production path: pick the workflow, engineer the context, and ship governed execution.
The 90-day goal is a production decision
At the end of 90 days, leadership should know whether to operate, expand, narrow, or stop the workflow. The goal is not a bigger demo.
Days 1-15: choose the workflow and baseline it
The first two weeks decide whether the rest of the plan will be useful.
Do not begin with a platform selection. Begin with workflow selection. Interview the sponsor, process owner, reviewers, and technical owner. Pull five to ten recent examples. Map how the work starts, what systems it crosses, where judgment enters, what action closes the loop, and which metric would prove improvement.
The deliverable is a decision package:
- workflow definition
- current baseline
- process owner and technical owner
- context sources and gaps
- approval and risk boundaries
- first-release scope
- success metric and review cadence
If this package cannot be produced, the 90-day plan should pause. More build time will not fix an unchosen workflow.
Days 16-35: engineer the context and controls
The next phase turns the pilot from a neat interaction into a workflow system.
Build the context contract: which CRM fields, ERP records, tickets, documents, emails, call notes, policies, examples, and spreadsheets matter. Decide source-of-truth rules. Define freshness expectations. Decide what the agent can read, what it can draft, what it can recommend, and what it cannot do without approval.
This is also when the team writes the first eval set. Include ordinary cases, edge cases, stale-data cases, policy conflicts, and examples where the right answer is escalation.
Days 36-65: build the controlled release
The first production version should be narrow. It should run inside or beside the real workflow, but it should not receive unnecessary autonomy.
For many teams, the right release is human-approved. The agent prepares a renewal brief, triage recommendation, invoice-exception packet, lead route, or policy answer. The reviewer sees the evidence, approves or edits, and the approved result lands in the right place.
This is where Metacto’s AI Agents & Workflows approach matters: a production workflow is not only a model response. It includes source integrations, human review, write-backs, an eval framework, monitoring, dashboards, and runbooks. In Metacto’s renewal-brief work, for example, the workflow reclaimed 4 hours per CSM while keeping actions 100% human-approved; that is the right shape for a first controlled release.
Days 66-90: launch, measure, and decide
The last phase is not a victory lap. It is the measurement window.
Launch to the first user group. Watch adoption, acceptance rate, edit rate, rejection reasons, escalation patterns, cost, latency, context misses, and the business metric. Hold a weekly review with the process owner. Capture incidents and near misses. Add failed cases to evals. Decide what must change before expansion.
Launch path
Use this sequence as the partner acceptance path. The engagement should move from artifacts to a narrow production proof before any broad expansion.
Step 1
Demand artifacts
Ask for the workflow map, eval set, permission model, runbook, and measurement plan.
Step 2
Fund narrow
Contract around one workflow, one owner, one metric, and a clear production gate.
Step 3
Verify production
Verify review, write-back, logging, monitoring, and support before calling it done.
Step 4
Expand on evidence
Use evidence from the first workflow to choose the next build, pause, or refactor.
The launch path above is framed as a partner acceptance path, but it works for internal teams too. The important idea is that artifacts, production proof, operations proof, and expansion evidence arrive in order.
The 90-day plan by gate
Use the gate below as the executive tracker. It keeps the plan honest without turning it into a giant project-management artifact.
90-day AI pilot-to-production plan
A 90-day plan should not promise transformation. It should promise a real production decision about one workflow.
Window: Days 1-15
- Primary output
- Workflow map, baseline, owner model, context inventory, risk boundary, and first-release scope.
- Executive gate
- Is this the right workflow to fund?
Window: Days 16-35
- Primary output
- Context contract, source-of-truth rules, permission model, review rubric, and initial eval set.
- Executive gate
- Is the foundation strong enough to build?
Window: Days 36-65
- Primary output
- Controlled production release with review surface, logs, write-back path, support route, and monitoring.
- Executive gate
- Is the workflow safe enough to launch narrowly?
Window: Days 66-90
- Primary output
- Usage, quality, cost, incident, reviewer, and business-metric evidence from the first launch group.
- Executive gate
- Operate, expand, narrow, prepare more context, or stop?
What leadership should not do during the 90 days
Do not add a second workflow because the first one is going well. Finish the proof.
Do not expand autonomy because the outputs look polished. Expand authority only when review, logs, evals, and rollback are ready.
Do not replace the metric with anecdotes. If the baseline was cycle time, measure cycle time. If the value was quality, measure corrections and rework. If the value was revenue, connect the workflow to pipeline behavior.
Do not let ownership drift to the technical team. The process owner owns the workflow outcome; the technical owner owns the system’s behavior.
The day-90 decision
At day 90, make one of five calls:
- Operate: the workflow is useful and should enter continuous operations.
- Expand: the workflow is stable enough to add users, cases, or an adjacent workflow.
- Narrow: the workflow works, but the scope was too broad.
- Prepare: context, data, permissions, or ownership need more work before production.
- Stop: the workflow did not justify further investment.
A successful 90-day plan is not always a green light. Sometimes the best result is discovering, early and cheaply, that the pilot was not the right production bet.