An artificial intelligence system that doesn't read the company's operational context or write to its systems requires a human intermediary: copying the data, pasting it, reading the response, and re-entering it where it belongs. This human intervention consumes much of the time the tool promised to save, and explains why so many implementations receive positive user reviews but have no impact on costs.
Integration is not the final phase of an AI project. It's the part that determines whether there is a project at all.
Level | What does the system do? | What does the person do? |
|---|---|---|
1. Isolated | It responds in its own interface. | Copy, paste, reintroduce |
2. Contextual | Read real data from the systems | Validate and execute the action |
3. Operation | Read and write with permissions and registration | To decide in cases that require it |
The value leap is between level 1 and 2, not between 2 and 3. When the system knows the actual status of the order, the customer's history, and the applicable conditions, the quality of what it produces changes category, even if it still doesn't execute anything.
Many organizations try to jump from 1 to 3 and fail; almost none consider staying at 2, which is where the best value-risk ratio is.
Because it's not a problem of connectors, it's a problem of agreements. Connecting two systems requires answering questions that no one has ever had to answer in writing:
What is the source of truth? If the customer status is recorded in the CRM and ERP with different values, integration requires deciding which one takes precedence. This is a business decision and is often left unmade.
What identity is used to access the site? The system should consult the user's permissions, not a service account with full access, for the reasons we explained in AI can also leak secrets [internal link].
What happens if the other system doesn't respond? Retries, queuing, controlled degradation. A workflow that assumes constant availability fails the first day it's not there.
Who maintains the contract between them? An integration without a versioned contract breaks down every time one of the two sides changes.
These four questions are what turn a two-week demo into a three-month project. They're also what make the result useful.
It is advisable to assign a number to level 1, because it usually presents itself as a temporary situation and becomes permanent.
If an employee spends two minutes per case copying context and reintroducing the result, and the process handles two hundred cases daily, that's more than six hours of work per day dedicated exclusively to acting as a connector between two systems that don't communicate.
This work also has three side effects: it introduces transcription errors, makes it impossible to reliably measure the process, and generates resistance to the use of the tool, because users perceive—rightly so—that they are working for the system.
In August 2026, Reuters analyzed why established European companies like SAP, Capgemini, Sopra Steria, and OVHcloud have become unexpected winners in the AI cycle. The analysis's conclusion is precisely the thesis of this article: the challenge for companies is no longer choosing the best model, but rather making AI work with the software, data, and processes they already have.
This is consistent with what McKinsey identifies as the decisive factor in the impact on EBIT: the deep redesign of workflows, which only 21% of generative AI adopters had undertaken by 2025. You cannot redesign a workflow without integrating the systems that support it.
Total integration is a poor goal. Sufficient integration for a specific process is achievable in weeks. Four steps:
Each iteration builds a reusable piece of the integration layer. After three or four processes, the company has infrastructure, not isolated solutions. This is the composable approach we describe in The bottleneck is once again architecture
When a provider presents a solution, it's important to look at where the user's hands are: Does anyone ever copy information from one window to another?
If the answer is yes, then what we're seeing is a Level 1 upgrade, regardless of how good the model is. And the cost of upgrading it to Level 2 is rarely included in the price that was just announced.
Because a system that doesn't read the operational context or write to the company's systems forces a person to copy and re-enter information. This human intervention consumes the time the tool promised to save and explains why many implementations receive positive feedback but have no impact on costs.
Three: isolated, where the system responds through its interface and the user copies and pastes; contextual, where the user reads actual data from the systems and validates and executes it; and operational, where the user reads and writes with permissions and logging. The biggest difference in value is between the first and the second.
Not the connectors, but the agreements: deciding what the source of truth is when two systems disagree, with what identity the data is accessed, what happens if a system does not respond, and who maintains the integration contract between them.
If an employee spends two minutes per case copying context and re-entering results in a process of two hundred cases per day, that's more than six hours a day acting as a connector. Furthermore, it introduces transcription errors, hinders process measurement, and creates resistance to its use.
No. Full integration is an unattainable goal; sufficient integration for a specific process is achieved in weeks. It's best to choose a limited process, identify the three or four pieces of information it needs, document them in a versioned contract, and start with reading before writing.
Observing the user's hands: if someone copies information from one window to another, what we're seeing is an isolated system, regardless of the model's quality. The cost of connecting it to real-world systems is rarely included in the advertised price.
Is your AI tool connected, or is someone acting as a bridge? We designed the integration layer by contracts, starting with a specific process and leaving reusable infrastructure. Let's talk → |
Español Español 
English
Français
العربية
Português
Deutsch
Italiano
