.svg)
An executive calls us. He wants to “get into AI.” He’s read up on it, his competitors are making moves, and his teams are asking for ChatGPT. The first question we ask him is: what exactly is it for? Nine times out of ten, the answer mixes three needs that have nothing to do with each other.
Because “AI” doesn’t mean anything. Behind the term lie three categories of projects. Three budget ranges, from a few euros a month to several hundred thousand. Three timeframes, from a single day to a year and a half. Three levels of risk.
The most costly mistake isn’t diving into AI. It’s aiming for the wrong level. Paying for a transformation program when a €20 subscription would have sufficed. Or deploying a generic tool when what was needed was an agent connected to your data, seeing the mediocre results, and declaring that “AI doesn’t work for us.”
Here is the framework we use to identify a need before even discussing technology or price.
The three levels, at a glance
Source: IT Systems framework.
The table covers the basics. But it’s the nuances that make the difference between good and bad decisions. Let’s go through it step by step.
Type 1 — Off-the-shelf AI tools
You pay a subscription fee, deploy it in a few days, and your teams can start using it the very next day. Copilot for Microsoft 365, ChatGPT Enterprise, Claude Enterprise, Notion AI. It’s self-service AI.
These tools are excellent for boosting individual productivity on general tasks: writing, translating, summarizing a document, searching through files, or polishing a draft. Anything that doesn’t really depend on your data or your processes.
The problem can be summed up in one sentence: they don’t know your business. They don’t know who your customers are, what your latest contract says, or how your approval process works. And all they can do is respond—they can’t open a ticket, update a CRM, or trigger any actions. The benefits are real, but they remain tied to the individual. They never trickle up to the process level.
If you just need to boost your personal productivity on everyday tasks, Type 1 is all you need. And above all, don’t pay anyone for it. A subscription, two hours of training, and you’re all set. That’s exactly what we tell some prospects: a standard Copilot is all you need—you don’t need us.
Type 2 — Business AI Agents
This is our turf, so let's take our time.
A business AI agent isn’t just a smarter chatbot. It’s a combination of elements: a language model, a layer that searches your data (RAG), your business rules, and, most importantly, integration with your information system. The agent understands your context. It searches your documents, emails, and database. It applies your rules. And it takes action: it opens a ticket, updates the CRM, generates a quote, sends a response, or escalates the issue to the right colleague with the file already prepared.
While Type 1 assists you, Type 2 takes over an entire aspect of your business. It’s not just a better version of the same tool. It’s something else entirely.
Some of the tasks a Type 2 agent handles: a Level 1 help desk that independently handles forgotten passwords and VPN access; sorting and pre-processing accounting requests; responding to RFP requests by building on your winning proposals rather than starting from scratch; analyzing contracts; and serving as a client memory that knows everything that has been said and delivered.
The appeal of Type 2 lies in the balance between its cost and the time it takes to deliver results. A project delivered in four to twelve weeks. A return on investment measured in months, not years. The code and intellectual property remain yours. And the ability to start small, with a single process, then scale up once the value has been proven.
The trade-off: a business application requires configuration and integration. It doesn’t deploy with a single click. It needs a usable knowledge base and dedicated access to your IT system. It’s a short but real project that deserves to be managed seriously. We’ve fixed enough low-code DIY projects that work in demos but break in production to know where that leads.
For most companies, it is at Level 2 that the value of AI becomes apparent the fastest. It is also the level that most service providers are least skilled at, because it requires both an understanding of your business and the ability to properly integrate your tools.
Type 3 — Foundational AI Projects
Type 3 is the program that brings about a fundamental transformation of part of the company: a predictive model covering the entire supply chain, a multi-criteria decision support system, and a redesign of the data architecture so that the entire company can finally make decisions based on reliable, up-to-date data. We’re talking about a timeframe of twelve to eighteen months and significant investment.
When it succeeds, the impact is significant and the competitive advantage is lasting. When it fails, it’s almost always for the same reason: the data. A Type 3 initiative launched on the basis of poor, incomplete, or scattered data will yield nothing of value, regardless of the budget. Data quality isn’t just a technical detail of the project—it’s essential to its survival.
That’s not the right way to start an AI transformation. In almost all cases, we recommend proving your worth with a few Type 2 projects that deliver results quickly, building data and cultural maturity along the way, and tackling Type 3 projects once the groundwork is in place.
The most costly mistake: choosing the wrong level
Here are the two ways to make a mistake. They are symmetrical, and both are costly.
Aiming too high. A company wants “the big AI project that changes everything.” It launches an 18-month program costing several hundred thousand euros. The real need behind the ambition was to automate the processing of a stream of repetitive requests—a Type 2 agent that could be delivered in two months for a fraction of the price. The common outcome: the program gets bogged down, the budget is spent, and the solution never makes it to production because it was too complex for what was required of it.
Aiming too low. This is the opposite mistake, and it’s more insidious because it masquerades as caution. You deploy a generic off-the-shelf tool, realize it knows nothing about your company’s context, can’t interact with your systems, and its responses go in circles. And you draw the wrong conclusion: “AI isn’t for us.” The problem wasn’t the AI. It was the level of sophistication. The need called for an agent connected to data and processes, not a general-purpose assistant. This mistake comes at a double cost: wasted time, and the mistrust that takes root internally, delaying the adoption of a solution that would have worked by several months—sometimes more than a year.
The question we ask right at the start of the scoping process—and which helps avoid both pitfalls— is: What is the least ambitious solution that actually solves your problem? If a Type 1 solution is sufficient, we stop there. If a Type 2 solution is necessary, we build it. If only a Type 3 solution addresses the challenge, we say so—after verifying that your data is up to the task.
Identify your needs in four questions
Just four questions are all it takes to get you on the grid.
Does your need depend on your data and processes? If not, consider Type 1. If so, you fall under Type 2 or 3.
Do you want the AI to take action, or just respond? If you want it to open tickets, update tools, or trigger actions: Type 2 minimum.
A specific process, or a broad transformation? A clearly defined process falls under Type 2. A fundamental strategic transformation falls under Type 3.
Is your data reliable, structured, and accessible? If the answer is no, a Type 3 implementation is premature. Start with a targeted Type 2 approach and take the opportunity to structure your data along the way.
How do you choose the right level, in practical terms?
On paper, the framework is simple. In practice, however, the right level isn’t always obvious, because a single need expressed by a manager can actually encompass three different projects. Here’s the method we use to make a decision.
We always start with the problem, never with the technology. Not “do you want AI?”, but “what is costing you time, money, or customers right now?”. Until we have a concrete understanding of the problem, talking about Type 1, 2, or 3 makes no sense.
Next, we look at three things. Volume: Does the task occur often enough to make automation worthwhile? A tedious but infrequent task doesn’t justify an agent. Data dependency: Does solving the problem require knowledge of your customers, your contracts, or your history, or is general knowledge sufficient? And the action: Does it just involve generating text, or does it require taking action within your tools?
From there, the level practically determines itself. Generic task, no specific data, no actions to trigger: Type 1, and we’ll let you purchase it on your own. A specific process, grounded in your data, with actions to be executed in your IT system: Type 2. A broad strategic issue, dependent on high-quality data across the entire scope: Type 3, provided that data maturity is in place.
The final step—and the one most often skipped in failed projects—is defining how the project will go into production and how we’ll measure its success, before writing a single line of code. A well-chosen use case without a production plan or metrics is bound to end up in the POC graveyard.
A few concrete examples, by profession
To give you a concrete idea of what the grid looks like, here are some common needs we hear about, along with the corresponding level.
Most of the practical needs of an SME or mid-sized company fall into Type 2. This makes sense: these are specific processes, rooted in the company’s data, that require action—not just a response. That’s also why we focus our efforts there.
The Essentials
AI isn’t a single thing. It consists of three distinct categories of projects that operate on different levels. Type 1 saves everyone time—quickly and inexpensively—but its impact is limited to the individual level. Type 2 takes over an entire business process in just a few weeks, delivering measurable results—this is where you see the fastest return on investment. Type 3 brings about profound transformation, but requires time, resources, and solid data.
The key is to get it just right. Not heavier than necessary, nor too light for the desired outcome. Getting the scale wrong is what separates an AI project that delivers in two months from one that quietly fizzles out after eighteen.
Frequently asked questions
What is a business AI agent?
A business AI agent is a system that combines a language model, a data search layer (RAG), your business rules, and integration with your information system. Unlike a generic tool, it understands your context and can perform actions within your tools: open a ticket, update a CRM, generate a document, or escalate the issue to the appropriate contact. This is what we call a Type 2 project.
What is the difference between Copilot and a custom AI agent?
Copilot (or ChatGPT Enterprise, Claude Enterprise) is a Type 1 tool: ready-to-use, excellent for individual productivity, but it doesn’t know your business and can’t interact with your tools. A custom agent (Type 2) is built around your data and processes and performs actions within your IT system. One assists each employee; the other handles an entire process. For many simple needs, Copilot is sufficient, and in that case, we’ll let you know.
How much does an AI project cost for a business?
It depends entirely on the level. A Type 1 tool is billed on a subscription basis, per user per month. A Type 2 solution is priced as a flat fee based on scope, with a return on investment measured in months. A Type 3 strategic project represents a significant investment over 12 to 18 months. Confusing the levels is the leading cause of misallocated budgets. We detail the price ranges in our article dedicated to pricing.
How long does it take to deploy an AI agent?
A Type 1 tool can be deployed in a matter of days. A Type 2 agent goes live in four to twelve weeks, depending on the scope and quality of the initial data. A Type 3 project takes twelve to eighteen months. The factor that most affects the timeline for a Type 2 project is not the technology, but the state of your knowledge base and access to your IT system.
Should we start with a POC?
A POC only makes sense if it is designed from the outset to be rolled out into production. The leading cause of failure in AI projects is a brilliant proof of concept that no one knows how to scale up afterward. It is better to have a concrete, limited but real-world use case that is deployed directly into production within a controlled scope than an appealing POC that ends up as a one-off demonstration with no follow-up.
Is my company ready for AI?
For a Type 1, yes, it’s all about individual productivity right off the bat. For a Type 2, you need an identifiable process to automate and a knowledge base—even if it’s imperfect. For a Type 3, the real question is the quality of your data: a transformative project launched on poor-quality data is doomed to fail, regardless of the budget. If your data isn’t ready, start with a targeted Type 2 project and structure it as you go.
What level should I choose to start with?
In the vast majority of cases, we recommend starting with a Type 2 project for a specific use case: ambitious enough to deliver tangible value, yet narrow enough to deliver quickly and demonstrate a return on investment. This is the best way to build internal confidence before moving on to larger projects.
The next step
Not sure what your needs are? That’s exactly what we look at in the quick audit. A 30-minute video call, three questions, and a detailed recommendation: Type 1, Type 2, or Type 3, including the scope and budget. It’s free, no-obligation, and useful even if you decide to work with someone else or on your own.
.svg)