Jim Keller's Tenstorrent Just Raised $1.5 Billion to Build the First Real Nvidia Alternative. The RISC-V Bet That Could Reprice Enterprise AI Infrastructure.

In June 2025, Gartner published a forecast that more than 40% of agentic AI projects would be canceled by the end of 2027. Gartner analyst Anushree Verma cited three failure drivers: escalating costs, unclear business value, and inadequate risk controls. She also estimated that only approximately 130 vendors out of thousands claiming to offer agentic capabilities deliver genuine agentic functionality — the rest are agent-washed products: chatbots, RPA tools, and AI assistants relabeled without substantive capability changes.

Sixteen months after that forecast, the deployment data is consistent with Gartner's concern. S&P Global Market Intelligence and McKinsey survey data puts the share of enterprises running at least one AI agent in active production at approximately 31%. Banking and insurance lead at 47%. BCG and Forrester research places the median time from agent deployment decision to measurable business value at 5.1 months. IBM's 2025 CEO study found only 25% of AI initiatives delivered their expected ROI. These numbers tell a coherent story: most enterprises are attempting to deploy AI agents, most are not succeeding at production scale, and the gap between project initiation and production deployment is not a technology problem. It is an activation problem — the process of getting from approved project to production system that employees use, trust, and depend on.

The Three Ways Agentic AI Projects Die

Gartner's three failure drivers — cost, unclear value, and risk controls — are the right categories, but they do not explain the mechanism. Understanding how each kills projects in practice is necessary for designing an activation strategy that avoids them.

How Costs Escalate from POC to Production

The cost problem in agentic AI is structurally different from the cost problem in traditional enterprise software. A traditional software deployment has a largely predictable cost structure: infrastructure provisioning, licensing, and a defined implementation timeline. An AI agent deployment has an open-ended cost structure where each agent action triggers model inference, tool calls, and potentially further agent spawning — creating a cost curve that increases with usage in ways that are difficult to predict from a proof-of-concept budget.

Enterprise finance teams that approved AI agent projects based on POC token budgets find that production deployments generate token volumes ten to fifty times higher than the POC suggested. This is predictable and consistently underestimated. POC environments use curated, simple tasks with known scope. Production environments expose agents to the full range of task complexity and edge cases, and agent behavior on complex tasks — additional searches, re-evaluations, recursive tool calls to resolve ambiguity — generates token volumes that scale nonlinearly with task difficulty.

The enterprise AI token cost paradox makes this worse: as per-token costs have dropped by 80-90% over the past eighteen months, enterprises have deployed agents in more workflows at higher usage rates, and total token spending has increased despite declining unit costs. An agent optimized to complete simple tasks efficiently is often the same agent that consumes far more tokens on complex tasks, because it has the intelligence to recognize that a simple approach will fail and to search for a better one — a behavior that is architecturally desirable but financially surprising at production scale.

The solutions that work do not reduce agent capability. They give agents explicit cost budgets per workflow execution, implement token ceilings that trigger escalation rather than continuation when exceeded, and instrument agent behavior to identify the specific action types that account for disproportionate token consumption. A document review agent that retrieves a full contract corpus before answering a specific question is dramatically more expensive than one that retrieves only the relevant sections. These are optimization decisions that require observability tooling — cost-per-workflow dashboards showing agent action logs alongside token consumption — that most enterprise deployments lack in their initial phases.

How Business Value Disappears at the Twelve-Week Review

The business value problem is primarily a measurement failure, not a capability failure. Enterprise leaders who authorized AI agent projects typically did so on qualitative commitments — "this will reduce manual review time" or "this will accelerate customer response" — without establishing the baseline metrics that would allow them to measure whether those outcomes were achieved.

When the twelve-week project review arrives and the question is what the enterprise received for its investment, teams that have not instrumented their pre-agent workflow cannot answer with data. The agent may be producing real efficiency gains — in many cases it is — but if the comparison baseline does not exist, the ROI claim cannot be substantiated to a CFO or a cost-reduction committee. In that environment, the AI agent becomes an undefended budget line that does not survive the next planning cycle regardless of what the agent is actually doing in production.

The enterprises that successfully demonstrate business value from AI agent deployments share three properties: they established baseline metrics before deployment using the same measurement methodology they will use to evaluate the agent, they defined a minimum success threshold as part of the project approval document, and they built dashboards that show agent behavior and business outcome metrics in the same view so the connection between what the agent did and what the outcome was is visible to leadership at every review cycle.

How Risk Controls Collapse Organizational Trust

The risk control problem takes two forms in enterprise AI agent deployments. The first is output risk: the agent produces incorrect, inappropriate, or harmful outputs that damage the enterprise or its customers. The second is scope risk: the agent takes actions beyond its intended scope — accessing systems it should not reach, committing to transactions that were not authorized, or making decisions that require human review and judgment.

VentureBeat's analysis of enterprise AI agent trust failures documented a consistent pattern: companies that deployed AI agents without explicit scope limits found that the same capability that made the agent valuable in its intended use case created liability when the agent encountered an edge case outside that use case. A procurement agent that can create purchase orders is valuable when it creates the right ones and dangerous when it creates the wrong ones without a human checkpoint. The danger is proportional to the capability — a more capable agent creates larger incidents when its scope is undefined.

The solution is not to reduce agent autonomy to the point of uselessness. A heavily supervised agent that requires human approval for every action is an expensive form of autocomplete, not an autonomous system, and it will not activate successfully because employees who must approve every agent action quickly discover that approving every action is more work than doing the task themselves. The solution is an explicit scope contract: a documented statement of exactly what actions the agent can take without human approval, what actions require approval, and what actions are completely out of scope. Enterprises that publish these contracts and build approval workflows for boundary cases find that risk control becomes an activation accelerator rather than a blocker — employees trust agents more when they understand precisely what the agent will and will not do on their behalf.

The Scale of Agent-Washing and Its Cost to Enterprise Buyers

One of the most damaging dynamics in enterprise AI adoption in 2026 is the confusion between genuine agentic AI systems and agent-washed products. Gartner's estimate of 130 genuine vendors out of thousands claiming the label implies a market where the majority of products marketed as "AI agents" have no genuine autonomous task-planning capability.

The enterprise buyer cost of agent-washing is primarily organizational and experiential, not financial. Most agent-washed products are affordable, and the procurement decision is not where the damage happens. The damage happens when enterprise teams deploy a product expecting autonomous workflow execution and receive a slightly improved chatbot. That experience — the gap between expectation and reality — produces organizational skepticism about the entire agentic AI category, and that skepticism creates resistance to the next deployment, which may involve a genuinely capable system from a vendor with real agentic capabilities.

Notion's 24-hour agent rebuild provides a reference point for what genuine agentic capability looks like in a production product: an agent that accepts a task, continues working on it autonomously for up to 24 hours without human prompting at each step, uses real tools to complete real work, and delivers a finished output rather than a status update requiring human guidance at each decision point. Enterprise buyers evaluating products marketed as AI agents should apply a simple practical test: ask the vendor to demonstrate a task that takes the agent more than five minutes to complete without human input. Most agent-washed products cannot pass this test because they are fundamentally question-and-answer systems with improved natural language interfaces, not autonomous task-completion systems with multi-step planning capability.

The Activation Gap: Why Proofs of Concept Don't Become Production

The most common failure pattern in enterprise AI agent deployment is not a technology failure. It is an activation gap — the distance between a successful proof of concept and a production-deployed system that employees use, trust, and depend on as part of their regular workflow.

Signal's analysis of the Agentforce enterprise activation gap identified a consistent pattern across enterprise AI deployments: proof-of-concept environments are controlled and curated, the demo workflows are selected because they work well, and the humans evaluating them approach the exercise with expectations calibrated toward success. Production environments are neither controlled nor curated. The workflows are messy, the data is incomplete or inconsistently formatted, and the employees using the system compare it against the tools they already know and trust rather than against an imagined ideal.

The activation gap has three specific dimensions that each require a different design intervention:

The workflow integration gap. POC environments typically run agents in isolation — the agent completes a task in its own interface, and the output is manually copied into the systems where employees actually work. Production activation requires that the agent operate inside the systems employees already use: the CRM, the helpdesk platform, the document management system, the communication tools. An agent that requires employees to navigate to a new interface for every interaction imposes a behavior change cost at every activation event. That cost is low for early adopters who are motivated to try new tools and high for mainstream users who are already proficient in their existing workflow and have no reason to add friction.

The trust calibration gap. Employees who are skeptical of AI outputs — and in most enterprise environments, skepticism is the organizational default for new tools — need to see the agent succeed at a task they understand before they will trust it with a task they care about. POC environments typically demonstrate agents on synthetic or simplified tasks that are specifically chosen to succeed. Production activation requires a graduated trust-building sequence: start with low-stakes, easily verifiable tasks where employees can check the agent's output against their own knowledge, let employees build a mental model of where the agent is reliable and where it needs verification, and expand the agent's scope only after that mental model is established through repeated successful experiences.

The feedback loop gap. POC environments rarely have correction mechanisms: when the agent gets something wrong, the evaluator notes it and the demonstration continues. Production environments require fast, low-friction correction pathways, and they require those corrections to actually feed into agent improvement over time. Without a feedback mechanism, employees who encounter agent errors lose trust faster than they build it, and adoption stalls at the early-adopter cohort rather than expanding to the broader organization. The correction pathway does not need to be complex — a simple thumbs-down button that routes incorrect outputs to a review queue and updates the agent's behavior is sufficient — but it must exist and must be visible to employees who use it.

Why Banking and Insurance Lead at 47%

The 47% production deployment rate in banking and insurance — sixteen percentage points above the enterprise average — is not accidental. It reflects structural properties of the financial services industry that the activation problem is easier to solve against, and that other industries can deliberately replicate when choosing which workflows to target first.

First, banking and insurance have the most clearly documented, most easily measurable workflows in the enterprise. Loan underwriting has a written process with defined decision criteria and a measurable average cycle time. Claims processing has a documented workflow with clear approval authorities, a measurable error rate, and outcomes denominated in dollar value. Fraud detection has historical labeled data and measurable outcomes. When the workflow is documented and the outcome is measurable, the activation playbook simplifies dramatically: wire the agent into the documented workflow, measure agent performance against the historical baseline using the same methodology, and iterate on the exception cases that fall outside the agent's scope contract.

Second, financial services enterprises have established governance frameworks for employee authorization that translate directly into agent scope contracts. A bank that has written policies for what a relationship manager can authorize without escalation to a senior approval authority has the organizational vocabulary and governance muscle memory to write a scope contract for an AI agent. The agent receives the same authorization scope as an employee at a defined seniority level. This is a governance exercise built on familiar infrastructure rather than an entirely new institutional design challenge.

Third, the ROI calculation is directly verifiable in financial services. An underwriting agent that reduces average decision cycle time from three days to four hours has a verifiable dollar value per decision — the labor cost saved, multiplied by the volume of decisions, minus the cost of the agent. A marketing agent that improves content engagement has a harder-to-quantify value that requires attribution modeling and longitudinal analysis to support. Measurable, defensible ROI is what keeps agent projects funded through the twelve-week review cycle and beyond.

The Five-Step Enterprise Activation Playbook

The 60% of agentic AI projects that reach production are not distinguished by better models, higher budgets, or more capable vendors. They are distinguished by a deployment methodology that differs systematically from the 40% that will be canceled. The methodology is repeatable, and it applies across industries, workflow types, and vendor platforms.

1. Define the activation moment before building the agent. The activation moment is the first time a new user experiences the core value of the agent without assistance from a champion or trainer. For a claims processing agent, it might be the first correctly processed claim the employee submitted independently. For a contract review agent, it might be the first non-standard clause the agent flagged that the employee verified and acted on. Define this moment precisely before starting implementation, and design the entire first-use experience around reaching it within the first work session. Every hour between a user's first interaction with an agent and their first genuine value experience increases the probability they will not return. Activation moments that require a week to reach produce adoption curves that look like product abandonment.

2. Start with a workflow that has a documented baseline. Do not start with a workflow you are inventing specifically for the agent. Start with a workflow employees already complete manually, where you can measure current performance using the same measurement methodology you will use to evaluate the agent after deployment. The baseline serves two purposes: it is your ROI calculation, and it is your trust calibration mechanism. Employees who can see that the agent outperformed the manual baseline on the same task they perform daily — in their unit of measurement, using their vocabulary — develop trust in the agent more quickly than employees who are told the agent is performing well on tasks they do not recognize or metrics they cannot verify independently.

3. Publish a scope contract before any employee uses the agent in production. Before the first production user, distribute a written document — ideally one page — that states exactly what the agent can do without approval, what it must request approval before doing, and what is completely out of scope. This document is not for legal or compliance purposes. It is for activation. Employees who understand what the agent will and will not do can calibrate their trust appropriately, delegate tasks within the agent's defined scope with confidence, and escalate intelligently when they encounter boundary cases rather than abandoning the agent after the first unexpected behavior.

4. Instrument the activation funnel, not only the outcome metrics. Your analytics should tell you not just what the agent completed but how many users attempted their first interaction, what share completed their first session successfully, what share returned within seven days of their first session, and at what specific point in the workflow users who abandoned did so before completion. The activation funnel analytics that consumer AI apps use to measure retention and diagnose abandonment are directly applicable to enterprise agent deployments. Week-one abandonment after a successful first session almost always indicates a workflow integration gap — the agent works but does not fit naturally into the employee's existing workflow. Abandonment before completing the first session almost always indicates a trust calibration gap — the employee did not reach the activation moment before encountering a failure or confusion that ended their session.

5. Design the escalation path so the worst case is as fast as the manual baseline. The most common reason employees stop using an agent permanently is not that it produces errors. It is that when it produces an error, the recovery process is slower and more disorienting than completing the task manually from the beginning. If correcting an agent error requires navigating to a different interface, understanding what the agent did before the error occurred, and redoing the work in a separate manual system, the employee's rational calculation is to stop using the agent and return to the manual process that has a known recovery path. Design the escalation path — what happens when the agent fails, exceeds its scope, or produces an output the employee cannot verify — so that the worst-case recovery time is comparable to or faster than the manual baseline. A well-designed escalation path converts agent failures from trust-erosion events into managed exceptions that do not disrupt the employee's workflow or their willingness to use the agent on the next task.

The Budget Reckoning Arriving in Q1 2027

Gartner's 40% cancellation forecast implies a specific budget consequence: a substantial portion of enterprise AI agent spending in 2025-2026 will produce cancellation decisions and limited production deployment in the planning cycles now arriving. The Q4 2026 and Q1 2027 annual planning cycles are asking every department to justify its AI spending against measurable returns. For projects that cannot show a documented baseline, a tracked activation funnel, and a verifiable ROI calculation, the planning cycle is the cancellation event Gartner predicted.

The enterprises that have activated successfully — that have a dashboard showing agent performance against a documented baseline, a growing pipeline of expanding deployments from workflows that succeeded first, and a clear methodology for selecting the next workflows to deploy — will emerge from this planning cycle with larger AI budgets and stronger organizational conviction. The enterprise AI budget reckoning is the selection event: enterprises that can show they know how to activate agents will be funded to deploy more. The 40% that cannot will face a longer path back to organizational trust before investing again.

The enterprises that cancel their first wave of agentic AI projects do not necessarily have worse leadership or lower capability than the 60% that ship. They made the same technology bet, on the same models, from the same vendors. The difference is process: they did not define the activation moment, did not establish baselines, did not publish scope contracts, did not instrument activation funnels, and did not design escalation paths. Each of these is a fixable organizational decision, not a technology limitation. The enterprises that understand this will not wait for the next AI technology generation to try again. They will apply the activation playbook to the projects already in flight and be in the 60% that ship rather than the 40% that cancel.

Takeaway

Takeaway: Gartner's forecast that 40% of agentic AI projects will be canceled by 2027 is an activation failure prediction, not a technology failure prediction. The three failure drivers — escalating costs, unclear business value, and weak risk controls — are solvable with process, instrumentation, and organizational design, not better models or higher infrastructure budgets. The five-step activation playbook that separates the 60% that ship from the 40% that cancel begins with defining the activation moment before building, establishes a documented baseline before deployment, publishes a scope contract that enables trust rather than creating bureaucracy, instruments the activation funnel rather than only outcome metrics, and designs escalation paths that make the worst case as fast as the manual baseline. Banking and insurance lead at 47% production deployment not because they have different technology but because they have documented workflows, established governance vocabulary, and directly calculable ROI. Every industry has workflows with those properties — the enterprises that activate successfully in 2027 will be the ones that start there, not with the workflows that are most exciting or most ambitious.