AI has reached companies. The results have not.
McKinsey reports that 71% of companies now use generative AI. BCG and the MIT Media Lab NANDA project indicate that only about 5% create value at scale. This essay examines the gap between those two measures and what must be built around the model to close it.
Since ChatGPT launched in November 2022, generative artificial intelligence has moved from a subject limited to laboratories and technology companies into ordinary business life. Today, any professional can test advanced models in a few minutes, while companies distribute licenses, create internal assistants, and run pilots with ChatGPT, Claude, Gemini, Copilot, and other tools.
Access to the technology has become simple. Turning that access into measurable operational gain remains more complex because it depends on processes, data, systems, people, and clear criteria for success. That gap is where we created CatechLabs.
The distance between adoption and results
Adoption has accelerated, even though financial returns remain concentrated in a small share of organizations. McKinsey reports that 71% of companies regularly use generative artificial intelligence in at least one function. At the same time, research from BCG and the NANDA project, associated with the MIT Media Lab, indicates that only about 5% manage to create value at scale or turn their pilots into measurable operational and financial impact.
| Number | What it measures | Source |
|---|---|---|
| 71% | organizations that regularly use generative AI in at least one function | McKinsey, The State of AI (2025) |
| 5% | companies reporting value from AI at scale, in a survey of more than 1,250 companies | BCG, The Widening AI Value Gap (2025) |
| ~5% | organizations that turned generative AI pilots into measurable operational or financial impact | NANDA Project / MIT Media Lab, The GenAI Divide (2025) |
The studies use different methodologies, but they describe the same situation: experimenting with tools has become common; embedding them into operations remains the exception.
Between buying a license and improving a business metric sits a substantial amount of work. A company must choose a relevant process, understand how it works, organize the information it uses, define permissions, integrate systems, prepare users, and establish a reliable way to compare before and after.
Many projects stop before that stage:
- The pilot works during a demonstration but depends on manually uploading files.
- The assistant answers questions but cannot access the most important sources.
- Automation saves time in one activity and creates additional work in another.
- Because nobody recorded the previous performance, it becomes difficult to demonstrate the gain.
In these cases, the technology entered the company while the operation remained largely the same.
Where projects usually stall
In the conversations and diagnostics we conduct, the initial request is usually broad: the company wants to "use AI." When the process is examined more carefully, much more specific problems appear.
- Documents spread across folders, email, and different systems.
- Activities that depend on spreadsheets and knowledge held by a few people.
- Rules that exist but have never been formalized.
- Repeated tasks because systems do not communicate with one another.
- Little management visibility into time, capacity, quality, and risk.
In this environment, choosing the model is only one part of the project. Even advanced technology will struggle to produce consistent results when data is fragmented, responsibilities are ambiguous, or the process has not been clearly understood.
A model can produce a convincing demonstration and still fail in daily use. The difference appears when documents are incomplete, formats are unexpected, information conflicts, exceptions arise, and decisions require human review. That is when a proof of concept must become a system.
MIT Sloan recommends breaking workflows into tasks before choosing the technology. This analysis helps identify where automation creates value, which information is required, how much it will cost to maintain the solution, and which metric should change.
The work starts with practical questions:
- Which activity consumes the most time?
- Where do delays happen?
- Which errors repeat?
- What information supports the decision?
- Which steps still require human supervision?
The answers make it possible to decide where artificial intelligence actually contributes.
In some cases, a smaller, faster, and more economical model is enough. In others, the main gain comes from organizing data, integrating systems, or rebuilding a step in the software. Value comes from the combination of these decisions, not only from the isolated capability of the model.
The system built around AI
An enterprise application must work in the same environment where the work happens. That includes authentication, documents, databases, permissions, history, processing queues, approvals, records, and integrations.
It also requires decisions that rarely appear in demonstrations: what happens when information is missing, when the model has low confidence, when two sources conflict, or when an action needs approval before it can take effect.
This collection of decisions turns a one-off experience into usable software.
At CatechLabs, we treat software, data, and artificial intelligence as parts of the same system. Software connects the solution to the workflow, tasks, documents, and rules of the organization. When a user has to copy information from one system, talk to a chatbot, and then record everything again, an important part of the gain disappears.
Data provides the context an application needs to work consistently. Documents, histories, policies, records, and rules must be available with the right permissions. Without that foundation, an answer can look convincing while being incomplete, outdated, or based on the wrong source.
Models interpret documents, locate information, classify data, support analysis, produce reports, and execute tasks. The choice depends on the problem, cost, speed, volume, and required level of precision. Because every case has different constraints, there is no single model that fits every situation.
How we start a project
CatechLabs develops software for operations that depend on large volumes of documents, fragmented information, and specialized work.
The first step is understanding the current process. We speak with the people who perform the activities, observe the tools they use, identify waiting points, and examine how the result is tracked. This diagnosis lets us define a use case that is relevant, technically feasible, and measurable.
Projects that are too broad tend to consume time before producing any concrete change. We therefore look for one part of the operation where a solution can be implemented, put into use, and evaluated through its effect on time, capacity, quality, or risk.
Today, this work happens mainly through Solon and Forward.
Solon
Solon is our platform for complex legal operations, starting with judicial reorganization and insolvency.
These processes bring together thousands of documents, creditors, spreadsheets, filings, deadlines, and reports. Important information is often distributed across different files and systems, increasing manual effort and making traceability harder.
Solon organizes this workflow and helps legal professionals locate information, analyze documents, structure data, and produce reports faster. The platform takes on repetitive activities and makes the necessary context easier to access, leaving more time for analysis, strategy, and decision-making.
In practice, this can mean locating in seconds information that was previously scattered across hundreds of pages. It can also allow a team to review a larger volume of documents or complete in days an activity that previously occupied weeks.
These results depend on the context of each operation and must be measured. Technology creates capacity, but the effective gain varies according to the process, data quality, and how the solution is incorporated into the team's work.
Forward
Forward is our diagnostic and implementation practice for companies that have a relevant operational problem but do not yet know how to turn it into software.
The work starts with a Design Study, where we map the process, available data, systems involved, bottlenecks, and expected outcomes. From that assessment, we structure the use cases, architecture, risks, feasibility, implementation plan, and success criteria.
This stage reduces the risk of investing in technology before understanding the problem. It also avoids long diagnostics that produce presentations but never reach execution. When we find technical and economic feasibility, we move into development and deployment.
Each project must generate results for the client and, at the same time, produce reusable knowledge for CatechLabs. That learning can appear in new components, integrations, implementation methods, or capabilities incorporated into our products.
What we are building
CatechLabs is still at the beginning. We are developing products, working with clients, revisiting hypotheses, and discovering which parts of our vision work in practice. Some ideas move quickly; others need to be reshaped or abandoned.
Our direction, however, is clear. The next stage of artificial intelligence in companies will depend on the quality of the systems built around the models. Organizations that can organize their data, integrate processes, establish controls, and measure results will have a greater advantage than those that merely accumulate tools.
This space will record what we learn during that construction. We will write about AI agents, data, integrations, governance, Legal Tech, software architecture, and implementation, always paying attention to what changes in practice, where the technology can be applied, and how to verify whether it produced a result.
To receive the next essays, subscribe to the AI in Operations newsletter.
If your company already faces this kind of problem, learn about CatechLabs' work.
References
MIT / NANDA Project, The GenAI Divide: State of AI in Business
Research on the gap between generative AI adoption and measurable business impact.
MIT Sloan, How to Find the Right Business Use Cases for Generative AI
An approach to breaking down processes, assessing costs, and selecting use cases.
MIT Executive Education, Beyond the Algorithm: Bridging the Last Mile of AI Adoption
Discussion of how context, processes, and organizational maturity influence value creation.
BCG, The Widening AI Value Gap
Research on the difference between companies experimenting with AI and those creating value at scale.
McKinsey, The State of AI: Global Survey
Global research on the adoption, expansion, and capture of value from artificial intelligence.
Deloitte, State of Generative AI in the Enterprise
Analysis of mature enterprise initiatives and the factors associated with generating returns.