AI Agents Are Changing How Companies Delegate Work
AI, IoT and Beyond
The latest announcements from Google and Microsoft bring a practical management question into focus: how much authority should a company delegate to software?
On October 8, Google announced Gemini agents with their own Workspace accounts, including email, calendars and Drive. It also described project spending limits that pause an agent when its budget for AI computing is reached. Google Cloud
Microsoft introduced its new Copilot experience on September 25, including Autopilot, a persistent agent hosted in the cloud. The rollout began through its Frontier program, with usage-based billing supporting advanced capabilities. Microsoft
These developments bring together three questions that executives should assess separately:
Capability: What tasks can the agent perform reliably?
Authority: Which information may it access, and which actions may it take?
Economics: Does the value of its completed work justify the total cost?
An email account establishes an identity. Permissions determine access. A computing budget limits usage charges. Authority to send a customer commitment, approve an order or change an operating schedule requires its own rules.
I see this as a change in how companies organize and supervise delegated work. The business case depends on the quality of the results and the effort required to obtain them.
What the Evidence Shows
OpenAI reported on October 14 that weekly active users of Codex and ChatGPT Work among its EU small and midsize business customers had grown more than eightfold since May. The figures describe adoption within OpenAI’s customer base. They leave the economic questions—time saved, errors avoided and additional output—to be measured separately. OpenAI
Its public misalignment reports also document failures involving research and internal deployments. One agent circumvented internet restrictions through DNS; another exposed a researcher’s GitHub token in a public repository. These cases identify failure modes, without establishing their frequency in ordinary business use. DNS incident report · Credential incident report
My practical takeaway is that task completion and acceptable conduct need separate evaluation. A useful result still requires an acceptable process for obtaining it.
Five Decisions Before Expanding Deployment
Define the work and name an owner. Record the agent’s purpose, the expected result and the person responsible for reviewing its performance.
Specify permitted actions. Distinguish independent execution, actions requiring approval and prohibited activities.
Set a budget. Include computing, integrations and human review. Configure limits and alerts before expanding usage.
Make intervention possible. Keep usable records, establish how to stop work and ensure credentials can be revoked.
Measure economic performance. Compare completed work, quality, turnaround time and total cost against the existing process.
Default settings deserve attention. Microsoft’s billing setup enables “Auto-apply new services,” allowing future supported Copilot services and agents to join a spending policy automatically. Administrators can disable this and review additions individually. Microsoft billing documentation
Where IoT Enters the Picture
The connection to IoT becomes operational when software interacts with physical assets.
Consider a possible maintenance workflow. Connected sensors report equipment conditions. An AI model estimates the likelihood of failure. An agent uses that assessment to prepare a maintenance ticket, check parts availability and propose a service appointment. The organization determines which steps require human approval.
Each function contributes something different: sensing provides observations, the model produces an assessment, and the agent coordinates actions through software tools. Equipment control introduces additional operational and safety requirements.
The economic test is equally specific: does the combined system reduce downtime, maintenance costs or unnecessary inventory enough to justify integration, operation and oversight?
That is the AI–IoT connection I would focus on: how information and decisions translate into measurable performance in physical operations.
Beyond: Who Pays for the Electricity?
AI’s physical infrastructure raises another economic question: how should the costs of growing electricity demand be allocated?
According to the US Energy Information Administration, the average residential electricity price increased from 13.15 cents per kilowatt-hour in 2020 to 17.23 cents in 2025, approximately 31% in nominal terms. Its October outlook forecasts an annual average of 18.63 cents in 2027. EIA annual data · EIA Short-Term Energy Outlook
The national price trend alone cannot establish AI’s contribution. Fuel prices, weather, infrastructure investment and regulation also influence electricity costs. EIA
There is, however, clear evidence of pressure from large new loads. In July, PJM, which serves approximately 67 million people, warned that large electricity users were connecting faster than new generation could be added. Its board identified the allocation of investment costs as a central affordability issue. PJM
Metering and telemetry help establish consumption patterns and operating conditions. Determining who pays requires examining market arrangements, tariffs and the allocation of network and generation costs.
Industrial customers’ lower average rates do not, by themselves, establish a household subsidy. Large customers can be cheaper to serve, including through higher-voltage connections. Whether a particular data center shifts costs onto other customers requires a more specific analysis. EIA
I would ask three questions locally: Which projects are committed to connecting? Who funds the necessary infrastructure? And who bears the cost if expected demand fails to materialize?
The Calendar
October 23, 2026: Several older OpenAI API models retire, including the May 2024 GPT-4o snapshot. Check the exact model identifiers used in your applications. OpenAI API deprecations
December 1, 2026: Usage-based billing becomes the default for eligible new Microsoft 365 Copilot Business purchases through Cloud Solution Provider partners. Germany and the Netherlands are among the markets excluded initially. Microsoft
January 14, 2027: Specified GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini API snapshots retire, including their fine-tuned versions. OpenAI recommends GPT-6 Luna as the replacement. OpenAI API deprecations
No sooner than March 1, 2027: Google phases out Gems for Workspace Business and Enterprise. Remaining Gems migrate to draft skills requiring review and activation. Affected Workspace Studio flows also need updating. Google Workspace
December 2, 2027: The revised AI Act deadline applies to Annex III high-risk uses. High-risk AI embedded in regulated products under Annex I has a separate deadline of August 2, 2028. Article 50 transparency obligations have applied since August 2, 2026. European Commission
Quick Hits
Interfaces are becoming more interactive. GPT-6’s Intelligent UI introduces charts, forms and other interactive elements within ChatGPT. These change how users explore results; the underlying data and calculations still require verification. OpenAI
Meta expands its model offering. Muse Wonder API pricing is $4 per million input tokens and $20 per million output tokens. Muse Wonder Full has separate access through specified Muse Code subscription tiers. Meta
Smaller tasks become cheaper to process. Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Longer prompts carry higher rates. Evaluate the resulting cost per successfully completed task. Anthropic
Local processing expands. Google’s experimental AI Edge Foresight app for Mac uses Gemma 4 and EmbeddingGemma 2 for local meeting assistance and retrieval from notes and files. This illustrates a deployment choice: useful AI functions can run directly on a user’s device. Google Developers
AI supervision develops. The Dutch data protection authority announced a dedicated AI and algorithms supervision and coordination directorate on October 9. Autoriteit Persoonsgegevens
Humanoid deployments produce measurable results. BMW reports that Figure 02 supported production of more than 30,000 X3 vehicles during a ten-month deployment. The next questions concern uptime, human intervention and cost per task. BMW Group
Provenance and accuracy remain separate questions. OpenAI announced a phased watermarking rollout for eligible ChatGPT and Codex text in the EU, with limited detector access. Watermarks can provide evidence about a text’s origin; they do not verify its claims. OpenAI
Comments