95% of AI Agent Projects Die in Pilot: What's Missing Isn't Intelligence, It's a Control Layer
AI has moved from generating content to acting on its own. 95% of agent projects are stuck in pilot and never reach production, because what's missing isn't intelligence but a runtime control layer.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The AI architecture itself has become the new risk surface
Manoj gives three reasons trust suddenly matters. First, AI has shifted from generating content to executing actions on your behalf, and those actions can run for minutes or for days. Second, enterprises are moving from single-model workflows to multi-agent systems, where frameworks like OpenClaw can amplify the blast radius to possibly a thousand times that of ChatGPT. Third, policy can no longer be enforced only at deployment time; it has to be checked continuously at runtime, on every tool call and every output. The consequence is that the AI architecture itself has become a risk surface, and it needs a new infrastructure layer to manage agents.
— Manoj SaxenaThe problem isn't building agents, it's controlling them
He cites the MIT report: 95% of agent projects cannot get from pilot to production. The reason is that writing an agent takes a few hours while shipping it takes six or seven months — risk and compliance has to validate it against internal and external rules, security has to keep it from becoming an insider threat, finance has to keep it from burning through the token budget. Manoj's conclusion: the problem is not building agents, it's controlling them, and nobody is compressing the time on that step. What TrustWise does is a pre-production ‘spell check’ for agents.
— Manoj SaxenaThe control layer isn't a cost burden, it's cost leverage
The token economics of agentic AI are entirely different: a single input can trigger 20 to 50 actions, consuming 20 to 40 times what generative AI did two years ago. TrustWise uses three kinds of shields — security, compliance, and cost-and-carbon — to intervene at runtime, and claims measured results of 83% lower cost, 40% better security, and 60% better latency. The control layer is not a drag; it is the cost lever that lets agents scale.
— Manoj SaxenaAI is not an application, it's an actor that logs in by itself
Manoj distinguishes two kinds of software. For 75 years we built applications: an application waits for human input and its rules don't change. AI is an actor — it can log into websites and databases on its own, and the same input may give a different result tomorrow. So the enterprise stack needs a new runtime control layer, sitting above the model and agent orchestration layer and below the user experience layer. He also predicts that in three years, 90% of the control tower's users will not be humans but agents.
— Manoj SaxenaThe main users of the internet are no longer people
In the month just past, agent-generated traffic on the internet exceeded human traffic for the first time. Manoj compares it to the moment data traffic overtook voice traffic on AT&T's network, and treats it as a landmark migration: the enterprise stack of the future will be operated more by agents, at the speed of electrons, than by people. That gives rise to a distinction between AX (agent experience) and UX, and a control tower product has to serve both humans and agents.
— Manoj SaxenaCompliance isn't one layer, it's six aligned at once
An agent acting in the world has to align with six layers simultaneously: a global rights layer such as the UN charter of human rights; national law (Singapore's AI Act, the US, Saudi Arabia all differ); industry rules (FINRA vs HIPAA); company values; the business unit and workload layer; and the customer SLA and interaction layer. TrustWise completes this alignment assessment in 10 to 300 milliseconds. A failure at any one layer can turn into a disaster, and after the fact you have to be able to demonstrate the full ‘toothpick through the sandwich’ path.
— Manoj SaxenaKnowledge graphs can't govern an agent's runtime actions
A knowledge graph only tells you how entities relate; a context graph adds situation; a world model adds the laws of physics — but none of them is enough to control an agent's runtime behavior. What TrustWise invented is a semantic action layer: it defines the action paths that a given agent is permitted to execute at runtime, aligned across every layer. It works like a precomputed Google map rather than a self-driving car recomputing the map every 10 meters, which makes AI deterministic, provable, and faster.
— Manoj SaxenaHallucination can be an asset: 2 right out of 10 is enough
Phase one of the Control Tower is Guardian Agents, which prevent bad things from happening. The newly released Harmony AI adds Genesis Agents, built on OpenClaw, letting a set of agents generate hypotheses in areas like revenue leakage and fraud — the way Deep Blue once saw 95 moves deep. Manoj calls this ‘beneficial hallucination’: maybe only 2 out of 10 are right, but those two can be commercial game changers. This is his starting point for vertical AGI and autonomous finance.
— Manoj SaxenaIn their own words · checked verbatim
At the end of the day, trust cuts down to does the AI and the model do what you intend it to do? That kind of the heart of is it aligned to your business and personal intent?
Manoj Saxena4:17
In fact, there was an MIT report that says 95% of agent projects are not able to move from pilots to production. And the reason is that you can write up an agent very soon, but risk and compliance need to make sure that it is following all the internal and external regulations.
Manoj Saxena20:56
we have demonstrated as much as 83% reduction in costs as we improved safety by 40% and latency by another 60%.
Manoj Saxena31:39
last month, the first time ever, the traffic on the Internet, agent traffic exceeded human traffic.
Manoj Saxena45:12
There is a layer beyond that that we have invented and innovated called the semantic action layer, which is what are the allowable action paths for that agent at runtime that aligns to all those seven layers.
Manoj Saxena53:23
So what we are launching is the first prototype version of an open claw-based system where a group of agents can act as those machines that can go in and look 95 moves deep in domains like revenue leakage into fraud.
Manoj Saxena58:35
Figures
| Failure rate of agent projects going from pilot to production, per the MIT report | 95% | 20:56 |
| TrustWise measured cost reduction | 83% (alongside 40% better security and 60% better latency) | 31:39 |
| Actions a single agent input can trigger | 20-50 | 32:49 |
| Token consumption multiple versus generative AI two years ago | 20-40x | 32:49 |
| Prebuilt controls and policies TrustWise ships | over 1,100, covering 17 regulations | 42:08 |
| Agent count at one large global company by year end | 100,000 | 7:25 |
| Business units involved in the Hitachi OEM | 650 | 27:25 |
Glossary
- runtime governance
- Checking and constraining an agent's behavior in the moment it acts, rather than only before deployment or in an after-the-fact audit.
- AI control tower
- A cross-vendor, cross-model meta control plane that manages compliance, cost, and security across many agents in one place.
- guardian agents
- Control-type agents with a human in the loop and deterministic outputs, responsible for supervising and constraining business agents.
- semantic action layer
- A record and enforcement of the action paths an agent is permitted, under alignment constraints at every layer, making behavior deterministic and provable.
- trust posture management
- An approach to agent governance that weighs security, compliance, cost, and carbon efficiency at the same time.
How to listen
Enterprise CTOs and CIOs, heads of risk and compliance, and AI founders and investors who want to understand where agent deployments get stuck.
Product integrations and the four deployment patterns (roughly 27:25-31:00) can be skimmed.