In many companies, AI has already entered through the side door.
An employee uses ChatGPT to rewrite an email. A manager asks it to summarise a document. A marketing team tests ideas for social posts. A salesperson generates a follow-up outline. A business leader asks it for a summary, analysis or action plan.
All of this can be useful.
But it is not yet a transformation.
The problem is not that these uses are wrong. The problem is that they often remain individual, scattered and impossible to manage. Everyone experiments in isolation, with their own habits, prompts, tools, level of caution and understanding of the limitations.
The company may feel that it is progressing with AI, but it is not yet building a collective capability.
It is accumulating uses. It is not yet creating protocols.
A novelty produces a demonstration; a protocol creates a method
A novelty use case impresses quickly.
It shows that a tool can produce text, a summary, an idea, an image, an analysis or the beginning of a deliverable within seconds.
The demonstration is often convincing, particularly when it uses a simple example. Time is saved, the output looks polished and the potential is immediately visible.
But a demonstration is not enough to create a working method.
A protocol answers more demanding questions.
When should AI be used? With what data? To produce which deliverable? With what level of human control? According to which quality criteria? Within what limits? Who approves the result? Where is it stored? How is the real benefit measured?
Without these answers, AI remains a shiny box into which everyone places requests with varying degrees of control.
With them, it becomes an operational lever.
Useful AI begins with real work
The first instinct should be simple: begin with the work itself, not the tool.
Many AI initiatives fail because they start with a question that is too broad: “What can we do with AI?”
That question produces ideas, but rarely systems.
A better question is more concrete: “Which repetitive, slow, ambiguous or poorly standardised tasks consume too much attention today?”
These might include summarising meetings, preparing minutes, analysing inbound enquiries, qualifying leads, drafting replies, structuring proposals, comparing documents, preparing briefs, producing first versions of content or detecting inconsistencies in a file.
The starting point is not the magic of AI.
The starting point is an operational friction.
A good use case has an input, an output and a control
A robust AI use case must be framed as a process.
It needs a clear input: a document, note, transcript, client request, CRM record, brief, dataset or combination of sources.
It needs an expected output: a summary, matrix, draft, list of questions, analysis, plan, classification or recommendation.
Most importantly, it needs a control.
This third point is often missing.
AI can produce quickly, but it can also produce incomplete, approximate or contextually unsuitable information with great confidence. The challenge is therefore not merely generation. It is validation.
A useful protocol specifies what must be checked: accuracy, consistency, tone, compliance, sources, assumptions, risks, sensitive data, alignment with the offer and final human decision.
Without control, AI increases the speed of ambiguity.
With control, it increases the speed of useful work.
The prompt is not the system
Many companies still reduce AI to a collection of prompts.
That is a step forward, but it is not enough.
A prompt can help standardise a request. An isolated prompt, however, does not answer every methodological question.
Who uses it? In what context? With which documents? At what point in the process? What happens to the result? How do we know it is good? When should AI not be used? When must the work be escalated to a human?
The real asset is not merely the prompt.
The real asset is the complete protocol: context, data, instruction, expected format, control criteria, responsibility, storage and final use.
A good prompt without a protocol becomes a personal trick.
A well-designed protocol becomes a team capability.
AI does not replace judgement; it moves it
The most dangerous promise associated with AI is total substitution.
In most companies, the objective is not to remove people. It is to use their judgement more effectively.
A senior employee should not spend valuable time reformatting notes, copying information, producing a first outline or manually searching for scattered material.
Their judgement remains essential, however, to understand context, identify risks, make trade-offs, adapt, validate and decide.
Useful AI does not remove human judgement. It moves it towards the moments where it creates the greatest value.
AI can prepare, structure, compare, summarise and propose.
But decisions, responsibility and interpretation remain human.
A serious protocol must make this boundary explicit.
The real benefit is not always raw time saved
There is a great deal of discussion about saving time.
That is understandable. Some uses genuinely save thirty minutes, an hour or more on repetitive tasks.
But the most valuable benefit is not always raw time.
AI can also improve preparation quality, reduce omissions, standardise deliverables, make a method accessible to more people, accelerate learning or create a first version where nobody previously knew how to begin.
Within a company, these benefits may be more important than merely saving minutes.
For example, if AI helps a sales team prepare more effectively for meetings, the benefit is not limited to preparation time. It may result in better qualification, a stronger understanding of the need and higher conversion.
If it helps structure inbound enquiries, the benefit is not merely administrative. It is better prioritisation.
AI should therefore be measured against what it genuinely improves: speed, quality, clarity, consistency, decision-making and knowledge transfer.
Sensitive data requires discipline
Using AI in business also raises a question of caution.
Not every piece of content should be copied into any available tool. Not all client data can be processed without a framework. Not every internal document should leave the company’s controlled environment.
An AI protocol must therefore include confidentiality rules.
Which data is permitted? Which data is prohibited? Which tools may be used? What content must be anonymised? Which deliverables must be checked before distribution? Which uses are restricted to particular roles?
This framework is not intended to slow adoption.
It is intended to prevent adoption from becoming disorderly.
Trust in AI does not come from general enthusiasm. It comes from clear discipline.
How to create a first useful protocol
There is no need to begin with a large AI programme.
Choose one recurring, visible and low-risk use case.
For example: turning meeting notes into structured minutes, preparing an initial summary of an inbound call, generating qualification questions, rewriting a commercial proposal or creating a content outline from a brief.
Then write the protocol.
What triggers it? What data enters the tool? Which instruction is used? What format should be produced? Who reviews it? Which criteria determine approval? Where is the result stored? What happens when the output is poor?
Test it on a few real cases, adjust it, document it and train the people involved.
A useful AI protocol must be simple enough to use, clear enough to transfer and robust enough to prevent obvious errors.
Adoption depends on trust
A team does not adopt AI sustainably simply because management says it matters.
People adopt it when they understand where to use it, why it helps, how to control the result and within which limits they can act.
Otherwise, two behaviours emerge.
Some use AI everywhere, sometimes too quickly and without enough control.
Others avoid it, through caution, lack of confidence or because they cannot see how it relates to their actual work.
A protocol resolves this opposition.
It gives structure to the enthusiasts and reassurance to the cautious.
That is often the best starting point for healthy adoption.
The right question
The question is not: “Do we use AI?”
The question is: “Have we turned some AI uses into reliable working methods?”
If the answer is no, the company is still in the phase of scattered experimentation.
That is not a problem. It is normal at the beginning.
But lasting value requires the next step: selecting the right use cases, formalising protocols, training teams, measuring benefits and maintaining clear human oversight.
AI is not useful because it is impressive.
It becomes useful when it is integrated into a working system.
That is when it stops being a novelty.
And starts becoming an operational advantage.