Hard results amidst AI hype 

Despite billions in investment, 95% of organisations are seeing no measurable return on their AI projects, with only a rare 5% of pilots delivering millions in value. In other words, most enterprise AI initiatives never scale beyond flashy proofs-of-concept. The reasons are telling: brittle workflows, lack of contextual learning, and poor integration into day-to-day operations cause these pilots to fizzle out. 

The report calls this gulf between promise and impact the “GenAI Divide.”

Yet, crossing that divide is possible, and a small group of vendors and companies are doing it by changing their approach. Success isn’t about having the biggest model or budget; it’s about building AI that learns, adapts, and fits seamlessly into real workflows. CohesionX and its platform VectorMind exemplify this new mindset by connecting the dots literally, demonstrating they are part of the unicorn 5%. In this article, we explore how CohesionX’s product philosophy aligns with MIT’s findings – and how it helps clients get ROI from AI, not just more pilot projects. We’ll also look at some rolled out products as solutions (Iris, Synapse Search, Atlas, Knowledge Nodes) that are delivering results beyond the GenAI hype cycle.

Why most AI projects stall and how to break the cycle 

The MIT NANDA study clarifies why so many AI pilots stall out. 

Enterprises often chase trendy tools without a clear business problem in mind. It’s common to start with a cool demo and only later hunt for a use case – resulting in expensive prototypes, vague ROI decks, and panicked CIOs.  

No surprise, then, that analysts predict up to 90% of AI projects will be discontinued within two years. The MIT researchers similarly found that 60% of companies evaluated GenAI tools, but only 20% even reached pilot stage – and a mere 5% made it to production. The rest remain stuck in perpetual POC purgatory. 

Key patterns define this GenAI Divide: 

• Limited disruption: Only 2 of 8 major sectors show any meaningful change from AI – most industries have high AI adoption but low transformation. 
• Enterprise paradox: Big firms run the most pilots yet struggle to scale them up. Smaller teams often deploy faster. 
• Investment bias: Budgets gravitate to flashy, front-office AI (think sales or marketing) while high-ROI back-office areas are overlooked. 
• Implementation advantage: Projects built with external partners double the success rate of in-house efforts, thanks to better alignment and speed. 

The core issue behind these patterns isn’t a lack of tech or talent – it’s a lack of learning. Most GenAI systems today “do not retain feedback, adapt to context, or improve over time”. They’re static deployments that might impress in a demo but never evolve into something enterprise teams trust day-to-day. According to MIT, even enthusiastic users of ChatGPT said they draw the line at mission-critical tasks, because “it doesn’t retain knowledge of client preferences… it repeats the same mistakes and requires extensive context input each session”. In fact, for complex or long-term work, employees preferred a human 9 times out of 10 (not because AI isn’t smart, but because it lacks memory and adaptability). This is the learning gap keeping organizations on the “wrong side” of the GenAI Divide. 

How do we bridge this gap?  

The MIT report’s answer is simple: build or buy AI that learns and fits. The standout successes are “adaptive, embedded systems” that start with narrow, high-value use cases and integrate deeply into workflows, improving continuously via feedback.  

In practice, that means prioritising business context, user feedback loops, and human oversight over generic one-size-fits-all bots. It also means focusing on unglamorous but critical processes where AI can quietly save millions. As the report notes, some of the biggest ROI comes from automating “ignored” functions like operations and finance: replacing outsourcing and manual drudgery, rather than chasing a moonshot AI that magically “disrupts” the whole business. (In case you need convincing: companies in the study saved $2–10M a year by using GenAI to handle customer service and document processing, cut external content agency costs by 30%, and saved $1M in risk management – all without major layoffs.) 

CohesionX was essentially born to close this divide. From day one, the team’s philosophy has been to avoid AI theatre and focus on tangible outcomes. “AI is a practical tool integrated into products, much like networking or security – not just a showpiece,” says Xander Coetzee, Chief of Delivery & Architecture. Rather than start with technology and look for a problem, CohesionX starts with real operational friction – the boring, repetitive, costly tasks that bog teams down – and only then applies AI in a targeted way. By beginning with a “clean-slate discovery” of where employees waste time sifting through data or doing rote work, they ensure any AI solution will immediately address a visible pain point (e.g. hours saved, errors reduced, faster decisions). This approach guards against the prototype-to-nowhere syndrome. Every feature is tied to an agreed KPI before a line of code is written – “if it can't pay its way, it doesn't ship”, as Coetzee puts it. Through this lens, CohesionX steers clients away from AI vanity projects and towards the “low-hanging fruit” that delivers value sooner rather than later.  

Learning systems that blend into workflows 

One of the report’s loudest findings is that adaptability and context matter more than raw model power. The best performers “embed themselves inside workflows, adapt to context, and scale from narrow but high-value footholds”. This is precisely CohesionX’s design philosophy, which they call an “intelligence fabric.” The idea is that AI should be woven into everyday operations (invisible but invaluable) rather than sitting on the side as a gimmicky chatbot. Three principles guide this fabric: 

• Invisible augmentation: AI works behind the scenes to assist work without forcing users to change habits. “Users should never have to pause and think, ‘I’m using AI now’”. Xander continues. Whether someone is filing an insurance claim or drafting an email, the intelligence is just there, quietly boosting speed and quality. This resonates with MIT’s point that workflow integration trumps flashy UX – people adopt tools that make their existing processes easier, not ones that make them do things differently. 
• Domain-specific expertise: Generic AI often falls short because it doesn’t speak the company’s language or know its rules. CohesionX’s “assistants” are extensively trained on each client’s own data – policies, product info, terminology, you name it. That means answers and actions feel authentic, like they came from an in-house expert, not a robot spouting Wikipedia. This deep context is a form of memory: the system learns your business so it can truly act on your behalf. (It’s no coincidence the report found 66% of executives want AI tools that learn from feedback, and 63% demand they retain context.) 
• Human-in-the-loop oversight: Far from handing everything over to black-box AI, CohesionX builds transparency and control into every layer. Every recommendation comes with a rationale or source. Every automated decision can be reviewed and tweaked by a human, with a clear audit trail. In practice, this means users (and auditors) can trace any output back to its origin – “you can trace any API call within your business processes, pinpointing exactly where logic might break down”.  

This design addresses one of the biggest enterprise concerns: trust. People will trust AI when they can see how it works, correct it, and ensure its following policy. CohesionX’s approach ensures the AI augments employees rather than running amok. 

Under the hood, CohesionX’s VectorMind platform is built to support this adaptivity and oversight at scale. It’s a secure, cloud-native stack that plugs into the client’s environment (your data stays your data). It comes with modular “domain assistants” – essentially pre-built skill modules for common functions like claims processing, HR queries, compliance checks, agronomy advice, etc. – which can be quickly customised to an organisation’s needs. Crucially, these assistants operate through agentic workflows: instead of rigid scripts, they use intent-driven chains that interpret user input, retrieve and transform information, and take actions by interfacing with business applications. This means the AI can execute multi-step processes (with appropriate permissions and checks) rather than just answering questions. An always-on evaluation layer watches over these agents, enforcing guardrails (e.g. data privacy rules, accuracy checks, no unauthorized actions) and logging every decision for accountability. In short, VectorMind is built for adaptability (it can slot into any workflow, any app) and continuous learning (feedback loops are built-in), all under strict governance. Little wonder the vendors succeeding in the MIT study were those “aggressively solving for learning, memory, and workflow adaptation” – exactly what VectorMind was engineered to do. 

Safe, controlled AI with Business GPT 

Along with adaptability is the need for safety and control. Another often unspoken reason many AI pilots don’t get past experimentation is that enterprise IT and compliance teams get nervous – understandably so. Public AI tools can leak sensitive data or produce wild inaccuracies (so-called hallucinations) that are unacceptable in a business context. The MIT report flags this “shadow AI” problem: 96% of employees admit to using AI tools at work, often without approval. That means company data is potentially being fed into who-knows-where, and decisions are being made on AI outputs that no one vetted.

CohesionX directly tackles this issue with its Business GPT solution. Business GPT is an enterprise-safe large language model – a closed-loop generative AI app that gives employees the convenience of an AI assistant without the risks. All the AI processing happens in a secure, private environment (on the VectorMind platform), so nothing leaks out to third-party servers. Moreover, Business GPT comes with real-time AI firewalls that catch those “hallucinations” before they reach the user. If the model starts to generate something that looks dubious or non-factual, the system can flag or correct it using the company’s knowledge base and policies. By locking down data and keeping a human (or at least a human-designed policy) in the loop, Business GPT ensures AI adoption is both fast and safe.

This kind of approach is what will finally let AI scale beyond small pilots – because

it gives CIOs and CEOs confidence that AI can be unleashed across the workforce responsibly,

not in a rogue free-for-all. CohesionX is helping organizations put proper guardrails around AI use, so they can embrace these tools enterprise-wide rather than ban them or relegate them to sandbox experiments. 

Real AI solutions in action 

Perhaps the best way to illustrate how CohesionX aligns with the “state of AI in business” findings is through its actual solutions. Each of the following examples shows a previously unsolvable (or overlooked) business challenge that CohesionX tackled with GenAI – and crucially, each moved beyond pilot into real operation, delivering ROI in the field. 

Knowledge Nodes – On-demand expertise  

Decades of best-practice guides and research often sit locked away in binders or PDFs. Knowledge Nodes turns this into an expert-in-your-pocket. By embedding domain-specific documents into a retrieval system, it lets users (like South African farmers) ask questions in natural language via WhatsApp and get instant, context-rich answers. It’s always available, always relevant, and ensures critical know-how is usable at the moment it matters. 

Iris – Clarifying chaos from unstructured data 

Iris solves the nightmare of unstructured documents. From food safety labs to insurers, it extracts structured, reliable data from PDFs, scans, and forms – regardless of layout. Instead of staff manually copy-pasting results, Iris adapts to new formats, validates data, and outputs clean entries ready for dashboards or compliance. The result: faster workflows, fewer errors, and dramatically reduced risk. 

Synaptic Search – Conversational queries for corporate data 

Traditional business search is clunky and rigid. Synaptic Search makes it conversational: users ask in plain language (“Show me hybrid SUVs under 300k”) and the system translates that into precise database queries. By handling synonyms, typos, and casual phrasing, it surfaces the most relevant results – boosting customer engagement and helping staff access internal systems without technical friction. 

Atlas – Mapping the enterprise information abyss 

When millions of documents and SharePoint sites bury vital information, Atlas maps the chaos. It orchestrates semantic, metadata, and keyword searches, then connects results into a knowledge graph. Instead of sifting through haystacks, users see the right documents with context and relationships intact. Atlas saves time, increases confidence in decisions, and transforms knowledge management from a liability into an asset. 

Crossing the divide to an agentic future 

The MIT report concludes on an optimistic note: the GenAI Divide can be bridged, and doing so positions organisations for the next revolution in enterprise tech – something the authors call the Agentic Web. In the Agentic Web, AI isn’t just a tool responding to prompts; it’s a network of intelligent agents that learn, remember, and autonomously collaborate across systems. Business processes will be carried out by swarms of AI assistants talking to each other (with human oversight), APIs negotiating deals in real-time, and workflows that self-optimise beyond the static, siloed apps we use today. This might sound futuristic, but it’s already emerging – and it’s exactly where CohesionX is headed. By insisting on learning-capable, workflow-integrated AI now,  

CohesionX is effectively future-proofing its clients for this agentic era.  

The VectorMind platform’s emphasis on agents, integration, memory, and interoperability is laying the groundwork for clients to plug into that Agentic Web as it takes shape. In fact, CohesionX’s highest-tier offering is literally called “Neo Cortex – Agentic level”, hinting at the autonomous agent abilities on the horizon. 

What’s clear from both MIT’s research and CohesionX’s on-the-ground experience is that the winners in the next wave of AI will not be those chasing the flashiest algorithms, but those focusing on systems that learn, remember, and adapt. It’s no longer about who has the biggest model; it’s about who has the smartest approach. CohesionX is helping enterprises make the “fundamentally different choices” needed – favouring custom-fit solutions over generic tools, and partnership over go-it-alone development – so they end up on the right side of this divide. The early results speak for themselves: clients are seeing measurable value, not just tech demos, and they’re gaining the confidence to expand AI from pilot projects to core operations. In doing so, they’re not only catching up to the present state of AI in business but also preparing for its future. The GenAI hype cycle is giving way to practical, sustainable AI deployment – and CohesionX is proud to be leading enterprises across that gap, one successful project at a time. 

In the era of the Agentic Web, where dynamic AI agents might negotiate deals and drive workflows, those enterprises that have already woven AI into their fabric will be a step ahead. CohesionX’s vision of AI as a “fabric, not just a feature” is exactly what’s needed to turn today’s AI pilots into tomorrow’s competitive edge. The divide is real, but with the right philosophy and partners, it’s one any business can cross. 

Depart on your AI journey with trusted experts:  

Info@cohesionx.co.za