How to Think About AI Transformation

The biggest change AI brings is not that software can suddenly write text, generate images, or have conversations.

The much more important shift is in how software can make decisions and execute work.

For most of software history, the logic of a task had to be defined before that task was executed.

If X happens, do Y.

If the customer has status A, apply rule B.

If a value exceeds a certain threshold, follow path C.

A developer had to predict the possible situations in advance and define how the system should behave.

AI changes that.

Part of the logic no longer has to be fully defined before the process starts.

A system can receive context, interpret it, and decide during execution:

what it is dealing with,
what information matters,
what information is missing,
which tool should be used,
what decision should be made,
what the next step should be.

That is probably one of the biggest reasons AI matters.

A simple example

Traditional order-processing software works very well when the input looks like this:

Product: SKU-123
Quantity: 20
Address: X
Delivery date: Y

Every variable is predefined.

But real customers do not always communicate like databases.

They write:

Send me another 20 of the same valves we ordered last time, but this time deliver them to our Kraków warehouse. If those aren't available, something similar is fine as long as it can arrive by Friday.

Traditional software struggles with that because the logic becomes extremely difficult to define through fixed rules.

An AI-enabled system can:

understand the intent,
identify the customer,
find the previous order,
identify the product,
check inventory,
search for alternatives,
verify delivery dates,
prepare the order,
send it to a human for approval.

The important part is not that AI can write a reply.

The important part is that some variables and decisions can now be determined dynamically during execution based on context.

That opens up an entirely new class of processes that can be automated.

Two types of AI transformation

When I think about AI transformation, I see two fundamentally different categories.

1. Internal AI transformation

The first is using AI to change how a company operates internally.

You start with an existing workflow.

Then you look for processes that are:

time-consuming,
repetitive,
expensive,
difficult to scale,
error-prone,
dependent on a few specific people,
or simply not being done because the company does not have enough capacity.

Then you ask whether that workflow can be redesigned using:

traditional automation,
an AI workflow,
an agentic workflow,
an AI agent,
or some combination of all of them.

The important part is that the goal is not to replace an existing process one-to-one with AI.

The goal is to redesign the process.

Example: RFP preparation

Today, an RFP process might look like this:

RFP arrives
↓
someone reads it
↓
extracts requirements
↓
searches for similar projects
↓
asks different teams for information
↓
collects answers
↓
creates a first draft
↓
prepares an estimate
↓
manager reviews it
↓
proposal gets sent

The wrong question is:

How do we build an AI agent that handles RFPs?

The better question is:

Which parts of this workflow should be performed by AI, which by traditional software, and which should remain with humans?

AI might:

read the RFP,
identify requirements,
find relevant previous projects,
search internal knowledge,
prepare a draft.

Traditional software might:

pull rates,
calculate pricing,
check team availability,
generate the final document.

A human might:

validate assumptions,
adjust pricing,
approve the final proposal.

The process could potentially go from six hours to thirty minutes without making the entire thing autonomous.

That is what I mean by internal AI transformation.

Not:

Let's add AI to the company.

But:

Let's redesign how work gets done and use AI where traditional software previously could not handle the complexity.

How do you identify good internal AI use cases?

The best opportunities usually share a few characteristics.

The workflow happens frequently

If something happens five times per year, automating it may not matter.

If it happens 20,000 times per month, even a small improvement can create significant value.

Volume matters.

The workflow consumes a lot of time

This is the obvious one.

If an employee spends 20 minutes completing something that an AI-enabled system can prepare in 30 seconds, there may be a strong business case.

The workflow is expensive

Sometimes volume is relatively low, but the people doing the work are expensive.

If senior engineers spend hours every week investigating repetitive incidents, even partial automation may create significant value.

The company lacks capacity

This is one of the more interesting cases.

AI does not always replace work that already happens.

Sometimes it allows a company to do work that previously wasn't economically possible.

Imagine a company with 50,000 customers.

Account managers might only have enough time to deeply analyze the top 500 accounts.

An AI system could continuously analyze all 50,000 and tell humans where their attention is actually required.

No person was replaced.

The organization gained a completely new capability.

The workflow requires interpretation

This is often the strongest AI trigger.

If a task is simply:

take X and copy it into Y

you probably do not need AI.

Traditional automation is enough.

AI becomes interesting when the task looks more like:

read
understand
classify
search
compare
reason
decide
prepare

These are exactly the kinds of workflows that were historically difficult to automate with rigid rules.

2. AI transformation inside a product

The second category is completely different.

Here, the company is not trying to improve an internal workflow.

It is trying to improve the product its customers use.

The core question becomes:

Can AI help the user do something faster, do something better, or do something that was previously impossible?

Imagine a financial analytics platform.

Previously, the user might need to:

open a report,
find the correct table,
export the data,
analyze it manually,
compare it with previous periods.

With an AI-powered feature, the user might simply ask:

Why did our margin decline the most in Europe last quarter, and which three products caused most of the change?

The system can:

interpret the question,
retrieve the relevant data,
perform the analysis,
compare periods,
explain the result.

That is not internal automation.

That is a new product capability.

AI inside a product can create three types of value

1. Help users do an existing task faster

Instead of building a report manually for 30 minutes, the user can describe what they need in natural language.

2. Help users do an existing task better

AI can analyze more information, find patterns, surface anomalies, or provide recommendations that would otherwise require significant manual work.

3. Allow users to do something that was previously impossible

This is probably the most interesting category.

A product that previously only stored information might now:

analyze it,
identify anomalies,
explain what is happening,
recommend actions,
execute some of those actions.

At that point, AI is no longer just another feature.

It can fundamentally change the value proposition of the product.

AI workflow, agentic workflow, and AI agent are not the same thing

These terms are used almost interchangeably today, but I think it is useful to separate them.

AI workflow

An AI workflow is mostly a predefined software workflow where AI performs some individual steps.

For example:

Email arrives
↓
LLM classifies it
↓
system retrieves information from CRM
↓
LLM prepares a response
↓
human approves it

The path is predefined.

The AI handles specific tasks inside it.

Agentic workflow

An agentic workflow gives the system more freedom.

The system can decide which step should happen next depending on the situation.

For example:

customer request arrives
↓
system determines what is needed
↓
decides whether to check CRM, documentation, or ERP
↓
uses the relevant tools
↓
decides whether it has enough information
↓
if not, takes another step
↓
prepares the result

The exact path is not necessarily identical every time.

AI agent

An AI agent receives:

a goal,
context,
available tools,
constraints,

and has more autonomy in deciding how to achieve that goal.

For example:

Resolve this support ticket.

The agent might decide to:

read the customer's history,
search documentation,
inspect logs,
call an internal API,
investigate possible causes,
prepare a solution,
request approval before performing a sensitive action.

The more autonomy you give the system, the more important things like these become:

evaluations,
observability,
permissions,
guardrails,
fallbacks,
human-in-the-loop.

An agent is therefore not automatically better than a workflow.

If a problem can be solved with a simple and predictable workflow, that is usually the better solution.

When should you not use AI?

This is just as important as identifying where AI is useful.

When the problem follows clear rules

If the logic is:

if X happens, do Y

and the rules can be described precisely, traditional software is often better.

It is cheaper, faster, more predictable, and easier to test.

You do not need an LLM to calculate VAT.

When an API already solves the problem

Sometimes companies build agents for things that could be handled with one normal API call.

If a user asks:

Where is my package?

you do not necessarily need an autonomous agent reasoning through five different systems.

You may only need to:

identify the request
↓
retrieve the shipment status
↓
respond

The less AI required to solve a problem reliably, the better.

When there is no real business problem

"We should do something with AI" is not a use case.

If nobody can clearly explain:

What improves if we build this?

then development probably should not start yet.

When the cost of failure is too high

Imagine an AI system that works correctly 97% of the time.

That sounds excellent.

But across one million operations, it means 30,000 failures.

So "accuracy" alone tells you almost nothing.

You need to know:

what types of errors happen,
how often they happen,
what each error costs,
whether those errors can be detected,
what happens when something goes wrong.

The goal is not maximum autonomy

This is another important point.

Human-in-the-loop is often treated as if it means the AI system is incomplete.

I think that is the wrong way to look at it.

A very good system might work like this:

AI completes 95% of the task
↓
human reviews the result for 20 seconds

instead of:

human completes the entire task manually for 20 minutes.

That is still a massive transformation.

The objective should not be:

remove the human from every process.

The objective should be:

design the best possible process.

Sometimes the best system will be fully autonomous.

Sometimes it will be AI-assisted.

Sometimes AI should only prepare a recommendation.

How should AI transformation actually start?

I do not think it should start with:

What kind of agent can we build?

It should start with the process or product.

1. Map the current workflow

How does the work happen today?

Step by step.

Not at a high level.

Precisely.

2. Find the constraint

Where are we losing:

time,
money,
capacity,
quality,
speed?

3. Decide what actually requires AI

Every part of the workflow can be classified as:

human,
traditional software,
AI,
some combination of them.

4. Calculate the value

How frequently does the workflow happen?

How long does it take?

How much does it cost?

What would a meaningful improvement be worth?

5. Redesign the workflow

Do not blindly automate the existing process.

Build a better one.

6. Decide the level of autonomy

Should AI:

prepare something?
recommend something?
perform an action after approval?
operate autonomously?

7. Define failure modes

What happens when the system is wrong?

How do we detect it?

When does a human take over?

8. Only then build

This is the difference between AI experimentation and AI transformation.

Experimentation starts with technology:

Let's see what an agent can do.

Transformation starts with the business or product:

Let's see how this could work fundamentally better.

AI is the tool.

It is not the objective.

The fundamental shift

For decades, software mainly executed processes that humans had already described in advance.

AI changes this because software can increasingly interpret situations and decide parts of how a task should be completed at runtime.

That means a huge category of work that previously could not be reduced to:

if X → Y

can now be partially or fully automated.

But that does not mean everything should become an agent.

The best AI transformations will probably combine:

humans + traditional software + AI

in exactly the right places.

And that is where the real opportunity is.