5%.
According to The GenAI Divide: State of AI in Business 2025 from MIT’s NANDA initiative, that is the share of integrated GenAI pilots that deliver measurable value in production.
The other 95% get stuck somewhere between proof of concept, internal enthusiasm and operational reality.
- No structural impact.
- No measurable P&L improvement.
- No scalable use in day-to-day work.
In most cases, the problem is not that the model is not powerful enough.
The problem is that the organisation underneath it is not ready. The report describes a clear funnel:
- 60% of organisations evaluated an enterprise GenAI system
- 20% reached the pilot stage
- 5% reached production with measurable value
Between those 60% and 5%, a lot of budget disappears. But so does confidence.
So the question is no longer: “Should we do something with AI?” The real question is: “Are we building another isolated experiment, or are we building something that can actually run in production?”
This is not a technology problem. It is an organisational memory problem.
MIT points to what it calls the learning gap: tools that do not learn from the organisation, do not integrate properly into the workflow and do not connect to the way work actually gets done.
That is exactly what we see in practice.
Too many AI pilots are still built as a separate layer next to the organisation. A chatbot beside the business. A promising interface, but disconnected from the documents, access rights, metadata, versions, workflows and decisions that define the real process.
The tool can respond. But it does not truly understand the business context.
- It does not always know which document is the correct version.
- It does not always know who is allowed to see what.
- It does not understand the structure of the case file.
- It does not know which decision was already validated.
And too often, the user has to rebuild the full context manually in every new interaction.
That may work in a demo. It does not work for business processes where accuracy, compliance and traceability matter. AI can only create sustainable value when it works on the information that actually drives the organisation: documents, cases, metadata, permissions, workflows, validations and decisions.
Without that layer, you are not building a business process. You are building an experiment.
Meanwhile, shadow AI is already happening
The report also exposes another uncomfortable reality: the gap between official AI strategy and what employees are already doing every day.
In many organisations, approved enterprise AI initiatives are moving slowly. But employees are already using personal AI tools to summarise, rewrite, analyse and prepare work. In other words: AI is already inside the organisation.
- But often outside governance.
- Outside audit.
- Outside document control.
- Outside access management.
- Documents are being copied.
- Business context is being shared.
Sensitive information may be entered into tools that are not part of the organisation’s information architecture. The risk is not only that AI adoption is too slow. The risk is that AI adoption is already happening without control. And when that happens, the organisation loses exactly what it needs most: trust, traceability and governance.
What the 5% do differently
The organisations that succeed do not simply place a generic AI tool on top of existing chaos. They embed AI into the work itself. Not as an isolated pilot. Not as a disconnected copilot. But as a domain-specific solution, integrated into the workflow, powered by reliable information and governed by clear rules.
That is the difference. For AI to work in production, it needs to know:
- which documents are relevant
- which version is valid
- which metadata provides the context
- which user is allowed to access which information
- which action was performed
- which decision requires human validation
Only then does AI become usable in a professional environment. Not as a black box. Not outside the case file. But as a controlled, traceable agent inside the organisation’s system of truth.
From AI pilot to reliable process
That is the position SoftAdvice stands for.
AI only creates real value when the information foundation is right. When documents are no longer scattered across mailboxes, SharePoint sites, personal folders and disconnected applications. When permissions, metadata, version control and workflows are not treated as technical details, but as the foundation of the operating model.
With M-Files and Aino Custom Agents, this becomes concrete. The AI agent does not work on random information. It works inside the system of record. It respects the existing permission structure. It uses available metadata. It operates on controlled documents. And it keeps the human in the loop where validation is required.
That is the difference between an AI demo and a production process. A demo shows what might be possible. A production process shows what works reliably.
The real question for your leadership team
Is your organisation using AI as a standalone experiment to show that it is innovating?
Or is AI embedded into your information architecture, with clear governance, controlled access and measurable results?
Most pilots do not fail because the model is not good enough. They fail because the organisation underneath the model is not ready. The 95% add AI on top of chaos. The 5% build the foundation first.
That is where the real difference begins.