Developing and Implementing an AI Strategy in 6 Steps
The myos playbook for AI roadmaps that genuinely create value: from the first lean pilot to cross-functional automation.
The mindset: start here first
Almost every leader currently feels the same pressure. Their own team, the advisory board, and customers all expect them to have an AI strategy and to roll it out across the entire company. Behind this sits a legitimate concern: whoever does not learn to use AI sensibly risks their own relevance over the medium term.
Introducing AI into a company often feels like having to eat an elephant all at once. You simply do not know where to take the first bite. This is exactly where most efforts fail, and across industries it is always for the same two reasons.
The two most common sources of failure
First: not starting small. Second: thinking too little creatively about possible use cases. Whoever avoids both has already set up the largest part of the work correctly.
The guiding principle throughout all six steps is therefore: do not try to boil the ocean all at once. AI gets distributed across the company step by step, value is proven early, and from these first successes comes the momentum for everything that follows. Crawl, walk, run, in exactly that order.
Step 1: Secure leadership backing
Without genuine backing, nothing happens in the end, and the responsibility for it lands on the driving person. Buy-in is therefore not a formality, but step zero of implementation.
Find the internal AI champion
You need a person with political capital, credibility, and the authority to rally teams behind them. In nine out of ten cases this is a leader. New employees rarely have enough trust and context, and people from the middle level quickly hit limits without real cover. Through the employee survey in step 2, you also often discover unexpected talents with high social capital who can carry the effort from below.
Practical tip: show the success first, then ask for support
A clever way to generate sponsorship is a visible success before asking for budget. Set up a lean pilot, for example a chatbot for customer service trained on the return policies, and compare the handling before and after. Such small proofs convince decision-makers faster than any presentation.
The decider and the enabler
Two roles need to be filled. The decider makes the directional decision. The enabler clears obstacles out of the way. If these are not the same person, your AI champion must have direct access to both.
AI adoption has stages
Before anyone pitches ambitious AI ideas, there should be clarity in the company about how AI adoption actually unfolds. It runs in four phases that build on one another:
- Single-seat tools: put ready-to-use AI tools into people's hands (such as ChatGPT for teams, coding-capable assistants for development). Train them and provide the best models.
- Single-seat processes: automate the recurring chain of tasks of a single person. In invoicing, for example, someone turns receipts into invoices, sends them out, and then chases payments. If this is automated, that person regularly wins back days of time.
- Within-function processes with multiple participants: the same principle, larger track. You take a workflow within a function (for example the handover from inside sales to field sales) and rebuild it end to end with AI.
- Cross-functional processes with multiple participants: the premier class, that is, true cross-functional automation (for example performance marketing and creative production that continuously deliver new assets).
Keep the order
Only tackle cross-functional processes (phase 4) once phases 1 to 3 are firmly in place. Whoever skips the order automates chaos instead of value.
Disarm two cultural opponents early
Two forms of resistance show up reliably. First, the fear of being replaced. Second, the disappointment of people who have already tried AI once without success. The latter almost always say the same sentence: this does not work, it takes longer than by hand. This belief only reverses when someone sees something that truly works. That is exactly why the early, visible pilot is so important.
Step 2: Get the company talking
Now you systematically look for problems. Employees do not always tell their managers what is broken, but they confide in an anonymous survey. The target group is experienced specialists, people from middle management, and operational staff. Aim for around 30 responses and scale this number with the size of the organization.
The most important questions in the survey
- How long have you been with the company? This helps to distinguish people with high internal influence from those with low influence.
- Which tasks do you handle weekly? In the age of AI, some tasks are simply no longer worth people's time.
- What is your most inefficient and time-consuming task, and why? This question opens the door to ask about earlier attempts to make the task more efficient.
Cultural change begins with experiments
Actively ask what individuals are already trying out with AI, and make these initiatives visible. This is how you reinforce exactly the behavior you want to see across the entire company.
Step 3: Interview people (with respect)
Interviews reveal what is broken, what works, and where AI creates leverage. Conduct 6 to 12 conversations of around 30 minutes each. One half with leaders, to understand business priorities, staffing mix, and organizational dynamics. The other half with deliberately selected operational employees who stood out in the survey.
How to run a perfect stakeholder interview
- Ask the right questions. Use the same core questions every time, individualized based on the survey: what do you do? What are you looking forward to with AI? What do you view critically?
- Transcribe everything. Use a tool for meeting notes. Clean transcripts later lead to considerably better recommendations during the synthesis.
- Probe until it gets concrete. A weak answer reads: I spend a lot of time chasing customers. What helps then is: how many hours? Which tool? What triggers the follow-up actions? Where does something get stuck? Walk me through a concrete example from this week.
- Quantify in every interview: frequency, time spent, cost of that time, impact (on outcome and strategic significance), and effort to fix.
Attention pays off
People reveal more when they sense that you really listened. If someone wrote in the survey that building reports is tedious, go deep in the conversation: which steps exactly? How many hours per week? Who else is involved? What breaks regularly?
Step 4: Turn chaos into structure
Now the collected material gets synthesized with AI. AI is excellent at turning unstructured material into clear patterns. That is exactly what you produced the raw data for in steps 2 and 3.
The workflow
- Set up a development environment (such as VS Code or Cursor) and an AI coding assistant within it.
- Load the entire context, that is, the results of the survey and the interview transcripts, into this environment.
- Have the AI summarize the pain points and opportunities.
- Critically question the result and then have a standalone report created.
Not in one shot
Do not run the synthesis only once. Let it run multiple times, add additional research with each round, and go deeper with each pass. The five dimensions from step 3 work well as an evaluation grid: frequency, time, cost, impact, and effort.
Step 5: Build a roadmap
Make your work visible. Transfer the synthesized data into tables, boards, and timelines. With the right connection, the AI can write the result directly into a structured database, that is, produce not just a document but a real project overview.
Even without a technical background
- Create a database in the form of a table. This is where your AI projects will live going forward.
- Connect your AI tools to this working environment.
- Insert your analyses and have the work structured as a roadmap (table, board, and timeline views).
- Evaluate and sort the opportunities by impact and effort.
Step 6: The cycle of implementation
Build the least that delivers the greatest value. Implementation does not mean adding AI everywhere, but deploying intelligence exactly where it unlocks the workflow.
Agentic workflows beat pure agents
The biggest mistake is jumping directly to fully autonomous AI agents without understanding how processes behave under load. Most processes are neither purely agent-driven nor fully automated. They are hybrids: a deterministic backbone with a few intelligent steps inside it. The rule of thumb: if a step can be formulated as an if-then rule, it is deterministic. If it can be described more like the task of a junior employee, it is a case for AI.
The AI automation spectrum
Processes can be placed on an axis from left to right: from fully rule-based (determinism) to fully reasoning (inference). A fixed field stands for a rule-based step, an AI field for a step that applies reasoning. The sweet spot for most processes lies not at the edges but in the middle: the agentic workflow, that is, a deterministic backbone with a few AI steps sprinkled in.
| Level | Character | Example on the same lead process |
|---|---|---|
| Workflow | Purely rule-based, all steps fixed | New lead, send follow-up, enter into CRM |
| AI workflow | One AI step in the fixed sequence | New lead, AI formulates the message, send follow-up, enter into CRM |
| Agentic workflow (sweet spot) | Fixed backbone with multiple AI steps and branches | New lead, AI researches the company, AI checks the knowledge base, enter into CRM, then the process branches: AI formulates the message and a Slack notification goes out, after which the follow-up is sent |
| Agent | Largely autonomous with instructions, knowledge, and tools | New lead and new email flow into the agent, which acts on its own: send follow-up, enter into CRM |
Build in checkpoints
Long-running agents without human or deterministic intermediate checks accumulate errors. Build in checkpoints every few steps so the system does not drift. If an agent runs for many hours and makes a one percent error early on, it multiplies until the result misses the target by a wide margin.
Break workflows into micro-steps
Break down every workflow on your roadmap into 5 to 15 micro-steps and then look for the most repetitive and time-intensive one. That is your AI entry point, not the whole workflow and not the glamorous step that leadership gets excited about. Narrow, surgically precise successes build momentum and give operational staff a clear sense of progress.
Build it yourself or buy it?
The cost of building software is dropping rapidly. Even so, your first reflex should be to buy. Buy when the step is generic or when speed counts. Build only when the workflow is deeply individual or when you can steer the agent as tightly as a junior employee. Cheap to build does not mean cheap to maintain. Use in-house developments only where they create real differentiation or strategic leverage.
How much AI is too much AI?
AI shines at turning unstructured chaos into structured clarity, not at performing exact arithmetic tasks. Keep sensitive steps and those with a high accuracy requirement in deterministic software and surround them with intelligent steps that accelerate everything else.
The cycle closes
Ship the first agentic workflow with checkpoints, measure the result, and feed it back into the roadmap. From every clean success comes the next one.
How myos applies this approach
These six steps are the operating logic behind a myos AI audit with a subsequent roadmap. We can run the entire process in your organization, that is, leadership alignment, employee survey, stakeholder interviews, AI-assisted synthesis, a living roadmap, and the first surgically placed agentic workflow. Alternatively, we accompany and enable an internal champion who carries the effort themselves.
The guiding principle stays the same across all six steps: start small, prove value early, distribute AI across the company step by step, and gain momentum from each success for the next one.
