Scaling Business AI Requires More Than Better Models: Secure Access, Reliable Capacity and Measurable Demand

What Happened

The latest technology developments highlight three constraints on business AI adoption: limited demonstrated demand for autonomous transactions, growing resistance to infrastructure expansion, and security weaknesses that extend beyond the model itself.

• Agent adoption received a reality check. Uber disclosed that third-party AI agents arranged just four rides in the third quarter of 2026. Agents also placed roughly 500 Uber Eats orders during Halloween week. These are different reporting periods, but both figures show how early agent-mediated commerce remains on this platform. [5]
• Data center expansion faces political resistance. Election results included dozens of community measures restricting development. An NBC News Election Day poll found 76% of voters opposed data centers in their communities. These developments could constrain new capacity; they do not establish an immediate shortage at any particular cloud provider. [4]
• A major theft reinforced the importance of privileged-access security. Binance said attackers targeted $6.34 billion in cryptocurrency after compromising an engineer’s laptop. According to Binance, the operation involved months of preparation, social engineering and novel malware; the FBI and TRM Security attributed it to North Korean actors. Binance reported freezing or intercepting $2.85 billion. [3]
• Platform dependence became a governance issue. Google faced criticism over a plan that would turn off Search for schools restricting AI. For buyers, the issue is whether access to an established service can become conditional on accepting additional AI functionality. [1]

Announcements from the preceding day add context: CrowdStrike said it would acquire AI observability company Fiddler AI and AI security-testing company Dreadnode, while Anthropic’s proposed FlashProve system would let users verify which model produced a response. Neither development eliminates the need for application-level controls. [23][13]

Why It Matters to Businesses

Agent capability is not the same as customer adoption. Uber’s figures do not prove that agents lack business value, particularly in internal workflows. They do argue against forecasting transaction revenue from demonstrations alone. Businesses should measure completed tasks, exception rates and cost per successful outcome before expanding autonomous execution. [5]

Infrastructure planning now includes local politics and power availability. Resistance to construction compounds existing grid-access delays. Reports that some builders are using retired jet engines for power illustrate the pressure, but these less-efficient workarounds are not evidence that every cloud workload faces disruption. Buyers should validate regional capacity rather than assume unrestricted expansion. [4][15]

Consumer technology exposes enterprise procurement weaknesses. Google reported dismantling a botnet that reached more than 1.3 million Google-certified Android TV devices. Separately, researchers testing 12 TVs found that some continued transmitting data after tracking opt-outs. Certification and settings therefore need verification, especially for workplace displays, meeting-room equipment and other connected devices. The TV study’s small sample should not be generalized to every product. [24][10]

Kimbodo Engineering Perspective

We would prioritize bounded automation over broad autonomy. Start with workflows that have explicit permissions, observable outcomes and reversible actions. Drafting a support response or preparing an order is a different risk category from issuing refunds, moving funds or changing production infrastructure.

The Binance account makes endpoint compromise a particularly relevant design concern. An agent platform’s security boundary includes developer laptops, deployment credentials, tool permissions and approval systems—not just prompt filters. A compromised workstation should not provide an unrestricted path to high-value execution. [3]

Model identity verification would address a narrower problem: whether a response came from the claimed model. It would not establish factual accuracy, permission to act or lawful use of retrieved content. FlashProve remains a proposal without a stated launch date, so it should not be a production dependency today. [13]

Portability also requires judgment. Keep business rules, evaluation datasets and authorization policies outside provider-specific interfaces where practical. Do not duplicate every managed service merely to claim independence; invest in an exit path for dependencies whose failure or policy changes would materially interrupt operations.

How We Would Implement It

• Establish a baseline. Choose one workflow and record manual completion time, error rates and cost. Define success as a verified business outcome, not a model response or tool invocation.
• Separate reasoning from execution. Place models behind an application gateway. Route proposed actions through a tool broker with schema validation, tenant-aware authorization, spending limits and narrowly scoped credentials.
• Gate consequential actions. Require independent approval for payments, destructive changes and sensitive exports. Use idempotency keys, reconciliation and rollback procedures where the underlying systems support them.
• Instrument the full transaction. Record model versions, policy decisions, tool calls, approvals and outcomes. Redact secrets and sensitive content before storing traces. Evaluate observability and security-testing products against these requirements rather than acquisition announcements. [23]
• Validate infrastructure assumptions. Test regional quotas, inference latency, failover behavior and recovery procedures. Use queues and asynchronous processing where the business can tolerate delay; reserve capacity only where demand and service commitments justify it.
• Test suppliers and devices. Review update support, administrative access and network behavior. Segment connected equipment from sensitive systems and verify that configured privacy controls behave as expected. [24][10]

Risks, Costs and Security

Security controls add cost, but uncontrolled execution creates larger exposure. Approval queues, isolated execution environments, evaluations and audit storage all consume engineering effort. Track their cost alongside inference and integration expenses, using cost per successfully completed task as the operating metric.

Data access requires legal diligence. Publishers’ lawsuit against Brave alleges unauthorized copying and commercialization of news content through its search index. Those allegations are not a court finding, but they underline why buying a search API does not by itself settle downstream content rights. Review licensing, retention and redistribution terms for retrieval-based applications. [22]

Resilience has trade-offs. Multi-region deployments, reserved capacity and provider alternatives increase expense and operational complexity. Apply them according to recovery requirements and business exposure, not as universal defaults.

The practical priority is to prove that an AI workflow delivers value, constrain what it can do, and verify the infrastructure and suppliers supporting it. Expand autonomy only when measured performance and controls justify it.

Sources

1. [1] Google faces backlash for plan that would turn off Search for schools that restrict AI
2. [3] Binance says hackers who stole billions in crypto spent months planning heist
3. [4] Data centers lost the midterms
4. [5] Uber only fulfilled 4 rides from AI agents last quarter
5. [10] Does opting out of TV tracking work?
6. [13] Anthropic is working on a system to prove AI models’ identity
7. [15] Data centers are turning to old jet engines as the wait for power grows
8. [22] NYT demands Brave shut down its paywall blocker and end deals with AI firms
9. [23] CrowdStrike to shell out nearly $500M for 2 AI security startups as the market for securing agents takes off
10. [24] Google dismantles botnet secretly running on 1.3 million Android TV boxes