The 90-Day AI Automation Roadmap: From First Workflow to a Governed Program

Most stalled AI projects were not doomed by the model. They lacked a baseline, tried too many things, skipped controls or never had an owner. This roadmap is deliberately narrow: one workflow, taken all the way from selection to a measured result and a governance routine in 90 days, with a decision gate at the end of each phase. It is the practical companion to our complete guide to AI automation.

Roles you need from day one

Phase 1: Foundation (days 1 to 30)

Weeks 1 and 2: choose and baseline

• List 10 to 15 candidate processes with volume, owner and pain points.
• Score them with the process selection framework and pick one.
• Write a one-page charter: problem, users, success metrics with numbers, out of scope, owner.
• Start measuring the baseline: time per task, cost per task, error rate, cycle time, volume.

Weeks 3 and 4: design and prepare

• Map the current process step by step, marking where judgment is required and where rules suffice.
• Choose the simplest architecture: workflow with model steps unless the path truly varies. See agents vs workflows.
• Choose the platform: existing tools, workflow platform or custom. See n8n vs Zapier vs Make and build vs buy.
• Assemble 50 to 200 historical cases with known correct outcomes as the test set.
• Define permissions and data flows; list what the system may read, write and never touch.

Phase 2: Build and pilot (days 31 to 60)

Weeks 5 and 6: build the smallest real version

• Integrate with the actual systems, in a test environment, from the first version.
• Implement the model steps with structured outputs and validation between steps.
• Add logging of inputs, outputs, decisions and costs from day one.
• Set caps: maximum steps, spend per run and timeouts.

Weeks 7 and 8: evaluate, then shadow

• Run the test set, measure accuracy against the pass mark set in the charter and fix the top failure categories.
• Run in shadow mode: the workflow processes live items but a person still does the work and compares.
• Add human approval steps for anything customer-facing or irreversible. See human-in-the-loop design.
• Run adversarial tests for prompt injection and misuse; see AI agent security.

Phase 3: Scale and govern (days 61 to 90)

Weeks 9 and 10: controlled live operation

• Go live to a subset of volume, with reviewers sampling outputs daily at first.
• Track the metrics from the charter weekly and compare to baseline.
• Collect corrections in a structured way and feed them back into prompts, examples, rules or retrieval sources. Systems that never learn plateau, a central finding of MIT's GenAI Divide report.

Weeks 11 and 12: govern and decide

• Register the workflow in an AI inventory with owner, purpose, data, permissions and review date.
• Write the runbook: what to do when it fails, how to pause it, who to call.
• Set the review cadence and the thresholds that trigger a rollback.
• Produce the ROI summary using the method in how to measure AI automation ROI.
• Check regulatory duties such as disclosures; see the EU AI Act overview.

Budget guidance

Allocate roughly in thirds across build, evaluation and operation, and resist the urge to put everything into the build. Evaluation and ongoing monitoring are where the value is protected. Set a hard spend cap for the pilot and a per-run cost cap, and review both at each gate. The costs to model are: platform and model usage per completed task, integration effort, review time and maintenance.

Common ways this plan goes wrong

• Skipping the baseline because the team is eager to build. Then no result can be proven.
• Letting scope grow into adjacent processes before the first is measured.
• Reviewers with no time. If the frontline expert cannot spend an hour a day in week 9, the timeline is fiction.
• Treating Gate 2 as a formality. The gates are where projects should sometimes stop.
• No owner after launch, which turns a working workflow into decaying software.

These map closely to the causes discussed in why agentic AI projects fail. If you would rather have an experienced team build the first workflow and hand over the runbook, talk to Navelo Software or see our custom software development service.

Sources and further reading

• Gartner press release on agentic AI cancellations
• NIST AI Risk Management Framework
• Building Effective Agents, Anthropic Engineering