AI strategy does not start with AI
For artificial intelligence to deliver results, we must first understand which business problem we are trying to solve. Buying new technology without understanding our own processes is a bit like installing a smart management system in a garage where no one can find the tool they need.

When technology comes before the problem
Today, practically every large company is discussing artificial intelligence. Leaders set implementation goals, departments look for use cases, IT teams test models and employees receive training.
At first glance, everything seems right: the company is learning to use a promising technology.
But I increasingly encounter the same problem. The discussion begins with the question, “Where can we use AI?”
It should start with a different question: “Where is our business losing money, time or quality?”
These are fundamentally different approaches.
In the first, we look for an application for a tool that has already been chosen. In the second, we seek the best way to solve a specific problem. AI may be one of the options, but it is not always the best one.
AI will not fix organizational problems for a leader
Imagine a company where customer requests arrive through email, messaging apps and direct messages to employees.
Some requests get lost. Others are processed twice. Managers do not know how long it takes to resolve a problem or who is responsible for it.
An idea arises: introduce an AI assistant for customer support.
But what data will it work with? Where is the consolidated history of requests? How do we identify who is responsible? What happens if the agent cannot solve the problem?
If these questions remain unanswered, AI will not eliminate organizational disorder. It will simply become another participant in an unmanaged process.
Sometimes a standard request-tracking system, consistent handling rules and clear accountability produce a much greater initial effect.
Only then do automatic request classification, answer retrieval and AI agents begin to deliver measurable benefits.
Automating disorder does not turn it into order.
Start with the economics
For me, the first step in an AI strategy is defining a measurable business problem.
For example:
- An excessively high cost of handling customer requests.
- Lost sales due to slow responses to inquiries.
- Errors in stock planning.
- Excessive effort spent preparing documents.
- Prolonged equipment downtime.
The next step is to establish the current metric and the desired outcome.
Suppose handling one customer request costs a company 300 rubles. We want to bring that cost down to 200 rubles without compromising service quality.
Now we have a problem that can be solved in different ways.
We can redesign the process, implement an off-the-shelf system, automate routine operations or apply AI.
We can then compare those solutions in terms of cost, time, risk and expected impact.
The figures are illustrative, but the principle is entirely practical: we are investing not in the implementation of a model, but in improving a business metric.

There is no need to fix the entire company at once
There is another extreme.
After hearing that AI requires high-quality data and processes, a company launches a massive program to put every system in order. Huge projects emerge to rebuild the architecture, create data warehouses and document every process.
A year later, the infrastructure has become more complex and the budget has been spent, but there are still no business results.
I do not believe the entire company needs to be in perfect condition before its first AI implementation.
What we need is to prepare the specific area in which we intend to achieve a result.
For the chosen scenario, it is sufficient to understand:
- How the process works and who is responsible for its outcome.
- What data is required and how much we can trust it.
- How to measure performance before and after implementation.
- How the solution will interact with existing systems.
- What to do when AI makes an error or cannot complete a task.
This makes it possible to launch small, manageable projects and gradually develop an overall implementation architecture.
A pilot is a test of a business hypothesis
I think one of the most common mistakes is to consider a demonstration of a working model a success in itself.
AI answers questions impressively, prepares documents or classifies requests. Executives are impressed. The team declares the pilot successful.
But what exactly improved in the business?
How much time was saved? How did the cost of the operation change? Was quality maintained? How much does it cost to support the solution? How often is human intervention required?
If there are no answers, we have tested the capabilities of the technology, but we have not demonstrated the project's value.
A pilot therefore needs clearly defined boundaries, baseline metrics and success criteria.
And, especially importantly, it must be allowed to produce a negative result.
If the experiment shows that conventional automation solves the problem more cheaply and reliably, that is also a good outcome. The company has avoided unnecessary investment and gained useful knowledge.
From experiments to a system
Individual successful pilots do not yet constitute an AI strategy.
A systematic approach emerges when a company can regularly identify suitable problems, assess their economics, run experiments and scale solutions that have demonstrated their effectiveness.
This requires clear rules for setting priorities, business accountability for results, a common technology architecture and transparent performance evaluation.
IT provides the tools, integrations, security and reliability.
The business defines exactly what needs to improve and what constitutes a successful outcome.
Company leadership ensures that priorities are aligned and resources are allocated accordingly.
Artificial intelligence becomes part of the company's operating model, rather than a separate experimental laboratory.

Where I recommend starting
I would suggest a fairly simple sequence of actions for a business leader.
First, compile a list of the most costly and painful business problems. Not a list of ideas for using AI, but a list of actual problems.
Then evaluate them according to expected impact, the complexity of changes, data readiness and the ability to measure the outcome.
Select several bounded scenarios and compare alternatives for each: process redesign, an off-the-shelf product, conventional automation or AI.
Then launch a pilot, measure the economics and decide whether to scale it.
Only when such decisions begin to recur should the company establish an overarching AI implementation model.
A good AI strategy starts not with choosing a neural network, but with understanding the business.
Because a leader's job is not to make the company sound technologically advanced. It is to make it more effective. And technology should help achieve that.