Direct answer
A useful 90-day AI plan does not start by buying a tool. It starts with a business problem, a baseline, an owner, data limits, and decisions at each gate.
Days 1 through 15 diagnose the work. Days 16 through 30 prepare data and owners.
Days 31 through 60 test 1 pilot. Days 61 through 90 produce a limited rollout or a stop decision.
This article is a decision plan, not a delivery promise. Management still selects targets, pass thresholds, vendors, budgets, and acceptable risks.
Make a 90-day plan, not an activity list
An activity list names meetings, demos, and implementation. A decision plan names the required output, its owner, the evidence to review, and the condition to continue.
Start with 1 process that has volume, inputs, outputs, and a human checker. Avoid unstable processes, sensitive decisions, or work without a business owner.
The earlier chapters explain business readiness in the AGI era and how to move from separate data to an agentic workflow. This chapter turns that work into a reviewable 90-day decision.
Days 1-15: diagnose and set the baseline
The business owner selects 3 to 5 candidate tasks. Each candidate must state its trigger, input, human steps, output, monthly volume, cycle time, common errors, and impact of a wrong output.
Measure the baseline from real work samples. Record work time, rework rate, backlog, escalation, and points where people change or reject an output.
Do not use estimates created after selecting a solution. A weak baseline can make savings look large without a testable comparison.
Use-case selection rubric
Use this rubric to compare candidates. It is not a validated model and does not give an automatic score; the owner selects weights and thresholds for the business risk.
| Dimension | Owner question | Early signal |
|---|---|---|
| Risk impact | Who is affected if the output is wrong, and what needs recovery? | Start with effects that people can reverse and check. |
| Data readiness | Is the input complete, lawful to use, and traceable? | Use data with an owner and access rules. |
| Process clarity | Does the team use similar work steps? | Select low variation for the first pilot. |
| Operating value | Do volume or waiting time matter? | Measure capacity, quality, and rework together. |
| Human control | Who approves the output and next action? | Require approval before external action. |
This rubric does not prove that a use case will succeed. It helps management reject candidates with unready data or uncontrolled risk.
Days 16-30: prepare data, owners, and limits
After selecting 1 candidate, write a short work contract. State the pilot purpose, allowed inputs, prohibited data, output, process owner, reviewer, evidence location, and actions the agent cannot take.
Separate read, proposal, approval, and execution rights. In an early pilot, AI can read approved sources and prepare proposals; people approve every send, data change, or customer action.
If marketing is selected, audit measurement before implementation. Data Health Audit checks GA4, Pixel, CAPI, UTM, and events for marketing data gaps. It is not an audit of the whole company database.
If advertising signals and CRM data do not connect, Attribution Bridge can help organize GA4, Pixel, CAPI, server-side events, CRM, and dashboards. Use it when budget decisions need fuller funnel data.
Days 31-60: run a pilot that you can stop
A pilot does not prove that the whole company is ready for AI. It tests 1 hypothesis within a narrow boundary, with approved data and available human review.
Set a comparison group when the process permits it. Compare pilot output with the baseline for the same work type, not with a success story from another week.
Record every rejection, correction, escalation, and review time. This evidence is more useful than prompt counts, chat counts, or document counts.
Capacity simulation, not a results claim
Illustrative simulation: a team handles 1,000 cases each month. The baseline takes 12 minutes per case. After a process change, work takes 7 minutes and includes human review.
The difference is 5 minutes per case, or 5,000 minutes each month. That equals 83.33 hours of returned capacity. Capacity does not automatically become cash, revenue, or staff reduction.
Record 83.33 hours as returned capacity. Do not record it as cash benefit before a cost reduction or cash increase is proven.
Gates: continue, correct, or stop
Each gate needs pass criteria approved by the owner before the pilot starts. Criteria can include sample accuracy, rework limits, review time, data compliance, integration stability, and no external action without approval.
- Continue: move to a limited rollout when the owner accepts evidence on quality, control, and operating load.
- Correct: correct prompts, data, SOPs, or review when the team can isolate the problem and manage the risk.
- Stop: stop when data is unauthorized, output fails the owner threshold, operating cost is unreasonable, or human control fails.
A stop gate is a healthy result. It prevents larger procurement for work that is not ready for automation.
Days 61-90: limited rollout or clean closure
A limited rollout adds users, volume, or input variation in steps. Do not add every division when the pilot passed for 1 process and 1 user group.
Write a runbook for inputs, review, escalation, manual fallback, access, logs, and recovery. Set who receives weekly reports and who can change scope.
If the pilot stops, keep the decision, evidence, discovered limits, and conditions that must change before another test. This closure prevents the team from repeating the same experiment without new data.
Management acceptance record and handover
Before a limited rollout, the process owner writes an acceptance record. It states the workflow version, data sources, access limits, checked samples, result against the threshold, open findings, and the owner decision.
For example, the process owner accepts a pilot for summaries of 1 request type only. The record states that people must approve customer sends and must escalate cases outside the approved list.
This record is not empty administration. It gives operations, finance, and procurement the same evidence when they assess scope, cost, and the next change.
Capacity value and cash cost
Capacity value differs from cash benefit. Capacity has value when the team uses it to reduce a queue, improve review, or serve volume that it can actually handle.
Net monthly cash benefit = verified cash benefit − monthly operating cost.
Estimated payback months = initial cost ÷ net monthly cash benefit. Use this formula only when that benefit is positive and stable.
Operating cost can include API, hosting, support, QA, and review. Initial cost can include setup, integration, and training.
Choose services by stage, not as mandatory packages
Marketing needs matter when the tested process uses campaigns, leads, content, or online sales data. Product and engineering needs matter when the pilot needs a digital asset or recovery of an existing application.
| Stage | Need | Service to consider |
|---|---|---|
| Diagnosis | Map use cases, risk, and vendor decisions. | AI Consulting or AI Implementation Roadmap. |
| Marketing data | Check funnel, events, and dashboards. | Digital Marketing Setup and Attribution Bridge. |
| Advertising operations | Campaigns need tracking, decision logs, and review. | Setup Sistem Meta Ads dengan AI. |
| Content and discovery | The team needs auditable SEO and AI Search work. | Pelatihan AI SEO & GEO. |
| Online sales | The business needs an owned storefront and checkout. | Jasa Pembuatan E-commerce, when online sales are in scope. |
| Application and technical team | An application needs recovery or developers need internal capability. | Vibe Code Rescue or Private AI Software Engineering Apprenticeship, when internal developers exist. |
| Maintenance | A passed scope needs review cadence and care. | OS Care Retainer. |
Do not treat all these services as prerequisites. Select a service only when the need exists at that stage and the owner approves the scope.
Choose support for your needs
If the plan still relies on assumptions, AI Consulting can help select a work path. When management has approved targets and decision gates, use AI Implementation Roadmap for resources, governance, and milestones.
For landing pages, tracking, campaigns, lead management, and an initial dashboard, see Digital Marketing Setup. For campaign operations that need monitoring, creative review, CAPI, and decision logs, see Setup Sistem Meta Ads dengan AI.
When the need is a content calendar, briefs, approval, and reporting, use OS Module: Content Ops.
When the team needs auditable SEO and AI Search assets, see Pelatihan AI SEO & GEO.
When online sales are in scope, Jasa Pembuatan E-commerce can build a storefront, catalog, checkout, and tracking. When an AI-built application needs stabilization, choose Vibe Code Rescue.
When the company has internal developers and wants capability through real work, consider Private AI Software Engineering Apprenticeship. After a scope passes, OS Care Retainer supports monitoring, audit-log review, and model or prompt updates.
For recurring advertising operations, consider Kelola Meta Ads Bulanan after a review. Scope, operator, and capacity are confirmed after the needs review. This service does not guarantee an active operator before that confirmation.
Brief risk and SEO notes
NIST AI RMF Core offers a voluntary framework: Govern, Map, Measure, and Manage. Use it to organize risk decisions and evidence; it is not certification or Indonesian law.
Google Search Central says that core SEO still applies to AI features. Google does not require special schema for AI Overviews or AI Mode, and it does not guarantee indexing, ranking, or display.
Next step
If management cannot yet select a candidate or owner, begin with AI Consulting. If the decision already has targets, data, owners, and gates, use AI Implementation Roadmap to sequence implementation, resources, governance, and quick wins.
Frequently asked questions
Is 90 days always suitable for AI transformation?
No. A 90-day plan fits only when data, access, and owners are available. Adjust the sequence or stop a stage when its decision gate shows that requirements are not met.
Does returned capacity always reduce cost?
No. Returned capacity becomes business value only when the owner uses it for backlog, quality, service, or costs that can change.
When should a company stop an AI pilot?
Stop when data is unauthorized, quality fails the owner threshold, human control fails, or cost and risk are unacceptable.
