MIT NANDA’s State of AI in Business 2025 is getting a lot of airtime. I read it end to end. The core claim is a “GenAI Divide” where adoption is high and business change is rare. The report says only about 5 percent of enterprise-grade tools reach production and 95 percent show zero return. I treat that as a useful signal, not a verdict. The method leans on interviews and a short ROI window.

At RSG we focus on what changes outcomes. We use three levers. The AI Loop that learns in production. The Effectiveness Model that finds the pillar blocking value. The RSG Reference Model that wires Generative, Insights, Decisioning, and Agent AI so workflows actually run. This piece maps MIT’s findings to those levers and shows where I agree, where I disagree, and what to do next.

What the report gets right

However much debated, the report documents some uncomfortable truths about the state of AI in the modern enterprise.

• Pilot-to-production is the main cliff. We have observed the same pattern amongst RSG’s clients. Most custom or embedded tools stall on workflow fit and a lack of learning memory to get smarter.  This leads to persistent and very frustrating, very manual correction work.
• Shadow AI is real. The report states that roughly 90% of firms’ workers use personal LLMs, while only around 40% of companies have purchased official subscriptions. That gap signals pent-up demand the official stack isn’t meeting. People reach for personal tools because the official stack is slow or brittle – or simply doesn’t exist, at least not yet.
• Spend allocation bias. Executives overweight Sales and Marketing. In reality, though, improvements to back-office processes (e.g., less dependence on agencies for content) often pay back faster via reductions in outsourcing.

Some Key Caveats

The method leans on interviews, a six-month ROI window, and composite indices – all useful, but also inherently limited – as the report points out. 

• The 95% failure headline is interview-driven with a short observation period. The report measures ~6 months post-pilot. Many enterprise programs need 9–18 months to show impact. Timing matters
• “Structural change” is scored by a composite index. Directionally helpful, but not a scoreboard

The report frames Agentic Web as an emerging direction to solve some of the issues that lead to GenAI failures. However, these are early days and evidence is sparse. 

In sum, the payoff with GenAI will take longer than many time-constrained execs realize.  This does not mean failure, per se, unless you think – as too many futurists do – that AI represents some sort of magic elixir that will disrupt your industry in the near-term.

RSG’s Take

At RSG, we have witnessed all of these frustrations in the realm of MarTech.  But we also think there’s a path to success with GenAI, for those thoughtful and patient enough to follow it. 

That path includes:

1. Closing the AI loop
2. Balancing your AI effectiveness across enterprise considerations
3. Clarifying an enterprise architecture that doesn’t over-index on the promise of Agentic AI

Let’s take a look at each.

The AI Loop

At RSG, we believe that real value emerges when three types of AI capabilities operate as a governed loop. Insights AI measures outcomes, Decisioning AI changes actions, and Generative AI refreshes what it learns from. That is how you turn output into impact.

• Insights AI measures outcomes and decides when we scale. It tracks lift versus a control, ranks content variants by segment and channel, and flags patterns such as rising unsubscribes after a campaign, higher refunds after a new offer, or conversion drops tied to missing product facts.
• Decisioning AI converts those signals into new behavior. For example, you can adjust eligibility rules, throttles, next-best action, and channel arbitration on a weekly rhythm rather than a quarterly one
• Generative AI creates variants and updates its knowledge based on what it learns. This can help you refresh templates, retrieval sources, product facts, tone, and locale on a schedule
• Governed memory makes learning safe. You persist case memory and content memory with consent, an audit trail, and a kill switch

This loop – and not the absence of Agents that the report discusses – is a critical missing piece in an enterprise value chain. Without the loop, you increase output, but you do not increase impact.

Having such a loop isn’t straightforward to build: You need tooling from several vendors and glue them together via decisioning, content and data foundational services (see reference architecture below); remove overlaps, bring them under combined governance and so on. Agent AI can help but it’s not there yet.

An AI Effectiveness Model

When working with large enterprises, RSG deploys a universal AI Effectiveness model buttressed by five pillars. When you see many pilots and little production impact, it usually means one or more pillars are lagging. And therefore, doubling down on existing tactics will not bridge the strategic gap. Instead, you need to move your focus to the blocker or lagging pillar.

The five pillars and what “good” looks like

• Strategy and Alignment. Goals are clear, tied to P&L, and owned by named leaders.
• Business Applications. A real workflow is selected, with a control group and success metrics.
• Data and Content. Facts, templates, and retrieval sources are governed, current, and owned.
• Technology and Execution. CI/CD, observability, cost limits, and runbooks are in place.
• Enterprise Governance and Ops. RBAC, audit trails, privacy controls, and a kill switch are implemented before scaling.

The MIT report points out what I’ve seen in real life: lopsided maturity in several key dimensions. If an enterprise comes in at Level 4 in Technology and Level 1 in Data, it will fail. Likewise for strong Governance with thin Business Applications. The imbalance kills momentum and leads to dead-end pilots. RSG’s model makes these trade-offs visible so leaders can choose where to shore up first.

Reference Model and Agent AI

The report explores “Agentic Web” suggesting that agentic applications could bridge GenAI workflow gaps. Some AI enthusiasts will tell you that agentic applications can enable you to surmount much broader AI challenges. In theory this makes sense; in practice, it’s not happening broadly today, in part because you need general AI maturity and success before you can even think about layering in Agent AI solutions.

At RSG, we link Generative, Insights, and Decisioning in a holistic Reference Model. Agent AI adds orchestration and controlled autonomy across those layers. However, agents help only when we treat them like software products. 

To that extent, citizen-developed agents will not solve your AI conundrum, the same way that skunkworked GenAI tactics have not led to strategic value. Do not buy the “anyone can build agents” claim. Once flows touch PII, multiple APIs, or cost ceilings, you are managing real software. Treat agents like products with CI/CD, observability, role-based access control, red-team reviews, and rollback before you scale.

Side note: MIT NANDA describes a world of large-scale agent cooperation. To quote from their homepage, 

“…a system where trillions of AI agents can collaborate, communicate, and transact across organizational boundaries without bottlenecks or security vulnerabilities.”

It sounds adjacent to Agent AI protocols like MCP and A2A. Standards are good when they interoperate in real stacks. They fail when they multiply without working code. Soon, we will have more standards than we can realistically support, which will defeat the whole purpose of having a standard in the first place.

Bottom line

GenAI can provide valuable benefits, especially in terms of productivity, if applied right.  The 95 percent “failure” shows brittle workflows and missing learning. Not weak models. If the system cannot show lift versus control, update policy weekly, and refresh facts on a schedule, impact will stall. 

We at RSG fix those three levers, then scale. If you want the checklist and the promotion gates we use, ask me. We will tailor it to one of your workflows.

What we will cover if you contact me

• Your loop: remember, measure, act, govern
• Your AI Effectiveness and Benchmarking
• Buying and standards stance
• Questions to ask your vendors

Frequently Asked Questions

Q. Where do Agent AI and interop claims fit in this model?
Agents are orchestration, not the destination. They must read outcomes from Insights, update Decisioning policies, and refresh Generative substrates under audit. Prove interop by running one end-to-end flow in your stack. Slides do not count.

Q. What usually blocks AI value first in enterprises?
Lopsided maturity. Strong Technology with weak Data and Content is common. The fix is ownership and refresh cadence for facts, templates, and retrieval sources. The Effectiveness Model points you to the weakest pillar so you spend in the right place.

Q. How do I keep AI costs in check while experimenting?
Set per-workflow cost ceilings and alerting. Track cost per action in the same “one view.” Gate promotion on both lift and cost. If lift rises while cost explodes, you do not scale yet.

Q. What if my marketing team wants to own AI Agents directly?
It can work if you treat agents as software. That means CI/CD, observability, RBAC, and rollback. If you cannot satisfy those basics, keep agents in a sandbox until you can.