AI creates a strong temptation to move quickly.
It promises to produce, summarise, classify, answer, analyse, automate and connect tasks that used to consume a great deal of time.
Inside a company, however, the main risk is not a shortage of tools. It is automating a way of working that is not yet clear.
The real issue
A vague process does not improve because AI assists it. It simply becomes faster, more opaque and sometimes harder to correct.
Launching an agent, connecting a form or creating an automation is not a strategy. Without rules, clean data, ownership and human validation, AI becomes another layer of complexity.
The right response is therefore to slow down briefly before accelerating. Not to produce a theoretical analysis, but to understand where friction appears, where information gets lost and where decisions depend too heavily on individuals.
Serious problems in a growing company do not always appear as sudden failures. They often emerge through repeated small misalignments: a forgotten follow-up, a delayed decision, a misunderstood priority, a poorly qualified opportunity or available data that nobody uses.
Signal 1: nobody can describe the process precisely
If the company cannot explain how a task is performed today, automating it is dangerous. AI needs a framework. Without one, it interprets, improvises and hides blind spots.
In isolation, the signal may look normal. Repeated every week, it reveals a weakness in the system. This is exactly the kind of pattern that must be recognised before it becomes a ceiling on growth.
Signal 2: there are more exceptions than rules
When every case is treated as unique, the work is not ready for full automation. Standard cases, frequent exceptions and situations that require human judgement must first be distinguished clearly.
In isolation, the signal may look normal. Repeated every week, it reveals a weakness in the system. This is exactly the kind of pattern that must be recognised before it becomes a ceiling on growth.
Signal 3: final validation has not been defined
A serious AI project must specify who checks the output, what they check, when the check happens and which criteria apply. Without a validation loop, the company confuses saving time with surrendering control.
In isolation, the signal may look normal. Repeated every week, it reveals a weakness in the system. This is exactly the kind of pattern that must be recognised before it becomes a ceiling on growth.
What needs to be put in place
Before automating, write the protocol. What is the objective of the task? Which data enters the process? What output is expected? Which criteria define an acceptable result? Which cases must be refused or escalated?
The next step is to separate assistance from automation. AI can first help prepare, summarise or classify information before it is allowed to trigger an operational action on its own.
Finally, measure the result. A useful AI project should be assessed on time saved, output quality, the rate of human rework, errors avoided and the team’s ability to use it in practice.
The method does not need to be heavy to be useful. It does need to be explicit. A company becomes more mature when it can explain how it decides, how it transfers information, how it follows progress and how it corrects course.
The leader remains central, but the nature of the role changes. The objective is no longer simply to compensate for gaps, carry exceptions and respond to emergencies. It is to turn what already works into a system that other people can understand and operate.
Where to start
- Choose a frequent, low-risk task.
- Describe the current process in five to seven steps.
- Define the criteria for an acceptable output.
- Test AI as an assistant before automating any action.
This first step should remain deliberately limited. The aim is not to create a large internal transformation programme, but to achieve an observable improvement at one precise point. Once the mechanism is understood, it can be extended.
The right sequence: understand, frame, automate
A robust AI project rarely begins with complete automation. It begins with a precise understanding of the task. What enters the system? What transformation is expected? Which result is acceptable? Which errors are serious? Which cases must remain human?
This framing work may look less impressive than an autonomous agent, but everything else depends on it. AI performs well when it operates in an environment where instructions, data and validation criteria are sufficiently clear.
The healthiest approach is often to begin with assistance. AI prepares, summarises, classifies, proposes or compares. A human still validates the result. This phase reveals limitations, exceptions and necessary adjustments before autonomy is increased.
Only when quality becomes stable should the level of automation rise. Automation then becomes a consequence of control, rather than a gamble on the tool.
The key point
AI is not an excuse to avoid structure. It is a brutally effective test of it: the clearer the system, the more useful AI becomes.
A company that grows sustainably does not simply add more effort. It builds systems that make the right efforts more visible, more consistent and easier to transfer.