Published
7 MIN READ
AI implementation plan: what it needs to include
An AI implementation plan is the document that says which processes will get AI, in what order, with what data, who is responsible for each stage and how the result will be measured. A plan ready for execution also includes the pilot's acceptance criterion, the estimated recurring cost and the way back if something goes wrong.
The plan is what separates an AI project from a collection of loose tests. Without it, each department buys its own tool, nobody knows how much the operation will cost in the third month, and the pilot becomes production without anyone having decided that.
At Wama, the plan is the deliverable of the AI assessment for companies, the first stage of the journey. This article shows what the document needs to contain, section by section, so you can hold whoever writes it accountable, whether that's a vendor or your own team.
What does an AI implementation plan need to contain?
A complete AI implementation plan answers seven questions before any line of code. When one of them goes unanswered, the gap shows up later, almost always at a higher cost.
Prioritized use cases: which processes are in, in what order and why.
Baseline: how much time, error or rework each process has today, to measure before and after.
Data and access: what data the AI will read, who is responsible for it and what LGPD (Brazil's data protection law) and artificial intelligence require for that use.
Owners: who approves, who operates and who reviews what the AI produces.
Pilot with an acceptance criterion: what has to happen for the test to be approved.
Recurring cost: model usage, infrastructure and people after delivery.
Rollback plan: how to suspend the automation and go back to the previous process without losing work.
The order in which these stages happen in the project is in the guide to AI implementation in companies. The plan is the document that locks in the answers before execution starts.
How do you prioritize the use cases in the plan?
Prioritize by crossing value and feasibility, not by which department complains the loudest. Microsoft's AI planning guide, updated in October 2026, says to score each use case by business impact, technical complexity, required resources and alignment with goals, and only then build a roadmap with success metrics, timelines and resources defined for each project.
The Brazilian federal government published a similar filter that also works for private companies. The criteria for identifying and prioritizing AI solutions in GuIA, an annex from June 2026, organize the evaluation into seven groups, including the maturity of the challenge, the real need for AI, data feasibility and risks. Real need is the group that cuts the list the most: many processes are solved with rules and integration, as the article on AI process automation shows.
A good plan shows the full list of processes evaluated, not just the chosen ones. Seeing what was left out, and why, is the fastest way to check whether the prioritization was honest.
Pilot, acceptance and rollback: the part most often missing
The most overlooked part of the plan is the one that defines when the pilot passed and how to undo the automation. Without a written acceptance criterion, every pilot ends up approved, because nobody agreed on what failing would look like.
The Brazilian federal government's GuIA artificial intelligence guide calls for, in the structuring phase, acceptance criteria with limits defined before the test and, in the implementation phase, a tested rollback plan. Microsoft's guide recommends starting with internal projects, which don't reach the customer, to limit the risk of the first test.
In the plan, this becomes a simple table: the metric, today's value, the minimum value to approve and who signs off on the decision. It also covers who reviews what the AI produces, a rule that belongs in the company's AI usage policy. At Wama, every AI output gets a human read before it reaches the client.
How much of the plan is recurring cost?
A good part of the cost of AI shows up after delivery, and the plan needs to estimate it. GuIA's structuring checklist says to work out the estimated total cost by adding cloud, inference and staff, and to identify who will maintain the solution, with in-house technicians or partners. In the sustainment phase, the same guide asks whether infrastructure and token costs are still within budget.
A token is the unit in which AI models charge for the text they read and write, and usage grows with the volume of the process. The full math, with a worked example in reais (Brazil's currency), is in the article on how much it costs to implement AI.
A serious plan separates three numbers: the project cost, the monthly operating cost and the maintenance cost. IBM notes, in its eight-step guide to implementing artificial intelligence, updated in April 2025, that models need regular monitoring and adjustment so they don't lose performance. That work has an owner and a price, and both need to be in the document.
How Wama builds the plan
At Wama, the implementation plan isn't sold on its own: it's produced by the AI Assessment, the small, paid first step of the journey that moves on to the custom Implementation Plan and then to execution.
The assessment shows where AI speeds up the company's slow, time-consuming and error-prone processes and delivers, along with the plan, the baseline for measuring before and after. There's no boxed plan: each company gets its own.
When the client moves on to the plan, the assessment fee is deducted from the first installment. To compare this model with other vendors', use the questions in the guide on how to choose an AI implementation company.
Common mistakes
The most expensive mistakes in an AI implementation plan are the ones that look like details when the document is being approved.
A plan that starts with the tool. If the document opens with the name of a platform, the decision came before the process analysis.
A vague baseline. "Improve customer service" can't be measured. The plan needs today's number: time, volume or error rate.
Unlimited access. Connecting AI to the company's systems requires saying what it can read and change, as the article on MCP for companies shows.
A pilot with no decision date. A test with no deadline to approve or shut down becomes a permanent cost.
Forgetting decommissioning. GuIA includes deactivation as the last stage of the cycle, with dependencies mapped and resources released.
Frequently asked questions
What's the difference between an AI implementation plan and an assessment?
The assessment is the investigation: it maps the processes, measures how they work today and points out where AI helps. The plan is the result of that investigation, with prioritized use cases, owners, acceptance criteria, costs and a schedule. At Wama, the AI Assessment is the stage that produces the custom Implementation Plan.
How long does it take to put together an AI implementation plan?
It depends on the number of processes evaluated and how accessible the data is. A plan for a single well-documented process comes out faster than a plan for the whole company. Microsoft recommends using the pilot results to estimate the timelines of the following use cases, instead of nailing everything down at the start.
Does a small business need an AI implementation plan?
Yes, but at a smaller scale. In a small business, the plan can fit in a few pages: one or two processes, the current measurement, who reviews the result and how much the operation will cost per month. The size of the document changes, but the questions it answers stay the same.
Who should approve the AI implementation plan?
Whoever is responsible for the process that will change, together with whoever decides the budget. When the process involves personal data, it's worth including whoever handles data protection at the company. The plan should name these people, because a decision without an owner is the first thing that stalls an AI pilot.
AI implementation plan or AI consulting?
The two complement each other. Artificial intelligence consulting can deliver recommendations and studies, while the implementation plan is the executable document, with order, owners and criteria. Before hiring, ask whether the deliverable ends in a report of recommendations or in a plan someone can actually execute.
If your company has already decided to use AI but doesn't have this document yet, start with the process that takes up the most of your team's time. Talk to Wama about the AI Assessment and the plan that comes out of it.
