What Today’s AI News Means for Enterprise Agents: Control Costs, Data Access and Human Oversight

What Happened

Meta and Microsoft are reducing employees’ use of Anthropic’s Claude while promoting their own AI tools. Meta’s Claude Code user count reportedly fell from about 60,000 to 30,000 [1][3]. Cohere launched North 2 with cross-session memory, revised agent orchestration, cloud and on-premises deployment, and administrator token-spending caps [24][28].

OpenAI made text watermarking available as an opt-in API setting and plans to add it for ChatGPT and Codex users in the EU [14]. Separate testing found watermark detection fell from as high as 95% to 17% when a quarter of the words were replaced [9]. Meanwhile, a survey found that only 11% of 396 businesses could forecast AI spending [32].

Why It Matters to Businesses

These stories point to three linked procurement questions: Can we switch providers, predict operating costs and govern what agents do? A survey of 300 executives found that agent projects reach production at an average rate of 34%; fragmented data was the most-cited obstacle to expanding agents’ knowledge access [18]. Provider consolidation may simplify purchasing, but tighter coupling to one model or platform can make later changes expensive [1][3].

Oversight also needs more than an approval button. Researchers warn that overloaded reviewers can end up rubber-stamping agent decisions [17]. That concern becomes more consequential as AI moves into regulated actions: a Utah pilot plans to let AI diagnose acne and prescribe medication without direct human oversight [15].

Kimbodo Engineering Perspective

We would treat model choice as a replaceable component and the agent’s data access, permissions and audit trail as the durable system. Cross-session memory can improve continuity, but it also creates retention and access-control obligations [24]. Token caps are useful guardrails, not complete budgets: cheaper-priced models cost more on 32% of over 6,800 tasks in a Microsoft analysis [32]. Measure cost per successfully completed task, including retries and review.

How We Would Implement It

• Build a controlled knowledge layer: expose approved records through permission-aware retrieval and APIs; test answers against source documents before granting broader access [18].
• Use a model gateway: route by task, log model and prompt versions, enforce structured outputs, and attribute tokens and retries to each workflow [32][37].
• Separate recommendation from action: give agents narrowly scoped tools, require explicit authorization for consequential operations, and present reviewers with evidence and alternatives—not just an “approve” button [17].
• Evaluate before and after release: combine deterministic checks with judged assessments and traced production cases to catch regressions in multistep workflows [36].

Risks, Costs and Security

Budget for retrieval infrastructure, evaluation, monitoring and human review as well as inference. Limit persistent memory by purpose and retention period; enforce tenant-level permissions at every retrieval and action. Do not treat text watermarking as reliable proof of origin after editing, given the reported drop in detection [9]. For regulated or safety-critical decisions, establish an accountable escalation path and validate the workflow against applicable requirements before allowing autonomous action [15][17].

Sources

1. [1] Meta and Microsoft pull back from Claude as Anthropic transforms from partner into competitor
2. [3] Sources: Meta and Microsoft are working to cut their employees' use of Claude; Meta employees using Claude Code have dropped to ~30K from ~60K earlier this year (The Information)
3. [9] OpenAI will watermark ChatGPT text in the EU but makes it optional for API users worldwide
4. [14] To comply with the EU AI Act, OpenAI plans to add text watermarking for ChatGPT and Codex users in the EU and an opt-in setting for API customers globally (OpenAI)
5. [15] In a first-of-its-kind pilot in the US, Nolla Health will use AI to diagnose and prescribe acne medications to Utah patients without direct human oversight (Annika Inampudi/Bloomberg)
6. [17] Attempts to Keep Humans in the AI Loop May Actually Push Them Out
7. [18] Connecting AI agents to enterprise knowledge
8. [24] Cohere launches North 2, an update to its enterprise agent platform with cross-session memory and a redesigned harness, available across cloud and on-premises (Sean Michael Kerner/VentureBeat)
9. [28] Cohere unveils North 2 AI agent platform with rebuilt orchestration and token spending caps
10. [32] Survey: only 11% of 396 businesses could forecast AI spending; Microsoft finds lower-priced models cost more than higher-priced ones on 32% of 6,800+ tasks (Wall Street Journal)
11. [36] Presentation: Building Reusable Evaluation Frameworks for Agentic AI Products
12. [37] Article: The Platform Engineering Playbook for Production LLMs