As AI reaches the point of being an everyday workplace tool, a stark divide is emerging between two very different classes of users — and the gap is already shaping who wins and who falls behind in productivity, security and competitive advantage. Martin Alderson’s recent essay calling out “power users” and “light users” crystallizes a pattern many IT teams and executives are watching closely: AI is not a single-button upgrade; it is an ecosystem problem that requires APIs, sandboxes, developer-friendly tooling and realistic governance to turn potential into measurable gains.
AI tools — from chat assistants to code generators and agent platforms — have moved from optional experiments to mainstream deployments inside enterprises, driven by optimistic forecasts of higher productivity and cost savings. Analysts estimate generative AI alone could add trillions of dollars in corporate value and materially lift labor productivity if deployed with complementary investments in skills and process change.
At the same time, rigorous field studies and user reports show the benefits are uneven and context-dependent. Controlled experiments with coding assistants have even found slower developer throughput in certain real-world maintenance tasks, while other longitudinal corporate deployments report significant efficiency lifts. The evidence is mixed, but the pattern is clear: outcomes vary by how AI is integrated into workflows, the maturity of underlying systems, and the organizational choices that govern tool access.
Martin Alderson’s piece — picked up and summarized by several sites — captures that divergence through two archetypes: the power user who embeds AI into dozens of routine workflows and the light user who queries chatbots occasionally. Alderson argues the most impactful productivity gains will come from bottom-up experiments by teams with access to internal APIs and the ability to run safe, sandboxed agents — not from top-down rollouts of a single canned enterprise product.
Alderson’s critique — that Copilot can feel like a “poor clone” of a chat interface and that some internal teams prefer alternative tooling — speaks to a broader point: the fastest route to industrial-grade productivity is rarely the single-button enterprise AI shipped in a suite, but rather a platform that supports code execution, API integration and governance. When companies restrict workers to one slow or inflexible choice, executives can conclude AI “didn’t work” and halt investment — even when other architectures would have delivered value.
Why APIs matter in practice:
What secure sandboxes provide:
Key structural problems in large companies:
Source: GIGAZINE As AI becomes more widespread, two types of AI users are emerging. How can AI improve productivity?
Background
AI tools — from chat assistants to code generators and agent platforms — have moved from optional experiments to mainstream deployments inside enterprises, driven by optimistic forecasts of higher productivity and cost savings. Analysts estimate generative AI alone could add trillions of dollars in corporate value and materially lift labor productivity if deployed with complementary investments in skills and process change. At the same time, rigorous field studies and user reports show the benefits are uneven and context-dependent. Controlled experiments with coding assistants have even found slower developer throughput in certain real-world maintenance tasks, while other longitudinal corporate deployments report significant efficiency lifts. The evidence is mixed, but the pattern is clear: outcomes vary by how AI is integrated into workflows, the maturity of underlying systems, and the organizational choices that govern tool access.
Martin Alderson’s piece — picked up and summarized by several sites — captures that divergence through two archetypes: the power user who embeds AI into dozens of routine workflows and the light user who queries chatbots occasionally. Alderson argues the most impactful productivity gains will come from bottom-up experiments by teams with access to internal APIs and the ability to run safe, sandboxed agents — not from top-down rollouts of a single canned enterprise product.
Two kinds of AI users: what the gap looks like
Power users: tooling as a multiplier
Power users are not just “more enthusiastic.” They are people and teams that:- Treat AI as an extendable platform, not a widget.
- Combine model prompts, code-execution tools, internal APIs, and lightweight automation to compress multi-step workflows into single artifacts.
- Iterate rapidly, share prompts and micro-workflows, and integrate model outputs with downstream systems.
Light users: the illusion of productivity
Light users typically interact with AI through a single channel — usually a chat interface — asking questions and receiving answers. This pattern creates two problems:- Surface-level gains that disappear when answers need verification, integration, or reshaping to fit business data.
- Hallucination risk: chat model outputs often contain plausible but incorrect content, which increases review overhead and error rates for downstream work.
The enterprise tool landscape and why it matters
Copilot and the limits of “built-in” AI
Microsoft 365 Copilot is widely adopted as an embedded AI experience across Office apps, and many organizations treat it as the approved corporate AI. Analysts and surveys show generative AI is the most frequently deployed AI in organizations, often via embedded apps like Copilot. But embedding does not equal capability: users and third-party observers have reported execution limits, timeouts and resource-constrained agent sandboxes that struggle with larger files and heavier tasks. Those resource limits, often implemented to ensure multi-tenant stability, can make agent workflows brittle and slow.Alderson’s critique — that Copilot can feel like a “poor clone” of a chat interface and that some internal teams prefer alternative tooling — speaks to a broader point: the fastest route to industrial-grade productivity is rarely the single-button enterprise AI shipped in a suite, but rather a platform that supports code execution, API integration and governance. When companies restrict workers to one slow or inflexible choice, executives can conclude AI “didn’t work” and halt investment — even when other architectures would have delivered value.
Evidence is mixed: when AI helps and when it hurts
Academic and industry studies paint a nuanced picture:- A randomized trial of experienced open-source developers found that using current AI coding tools slowed task completion by 19% on average, largely because of time spent prompting, waiting for outputs, and reviewing generated code. That study underlines the overhead of poorly integrated tools.
- Conversely, other case studies and vendor reports document double-digit productivity gains when tools are embedded into appropriate workflows with good governance — for instance, tailored code-review automation or tightly-scoped RAG (retrieval-augmented generation) systems. Broader market analyses from McKinsey show significant theoretical upside for generative AI across many functions if organizations invest in the supporting systems and reskilling the workforce.
The API imperative: why internal connectivity is the competitive lever
Alderson’s second big point — that companies with internal APIs will win — maps directly to what CIOs and analysts now say: APIs are the lingua franca that lets agentic AI act on business systems. Without APIs, agents are stuck with generic prompts; with APIs, agents can fetch precise, auditable data and invoke guarded actions.Why APIs matter in practice:
- They convert opaque prompts into deterministic function calls that return structured responses.
- They enable least-privilege access so agents can query data without exposing raw databases.
- They let teams build composable workflows by chaining endpoints, reducing the need for fragile prompt engineering.
Safe AI agents: sandboxing, resource limits and the role of isolated execution
Alderson urges that any internal AI agent deployment must be wrapped in secure mechanisms. This is not theoretical: industry best practices for code-executing agents have converged around sandboxed, ephemeral VMs or container-based microVMs that enforce resource limits, network restrictions and audit trails. Public cloud vendors and open-source projects now offer dedicated “agent sandbox” frameworks to isolate untrusted model-generated code.What secure sandboxes provide:
- Isolation: gVisor, Firecracker microVMs or container sandboxes separate agent execution from host systems.
- Resource control: strict CPU, memory and timeout limits prevent runaway processes and help multi-tenant fairness — but these same limits can break large-file tasks if not tuned.
- Controlled I/O: explicit allow-lists for network calls, mounted volumes and external APIs.
- Ephemeral state: sandboxes are created per session and destroyed, reducing persistence of potentially malicious artifacts.
Organizational structure, governance and the “small team” advantage
Alderson observes that small and medium-sized businesses (SMBs) are sometimes outperforming large enterprises because they can move faster: fewer approval layers, easier access to tools, and the ability to iterate on workflows without heavy procurement cycles. This is consistent with corporate reports showing implementation friction at scale — especially where engineering is outsourced or siloed.Key structural problems in large companies:
- Locked-down IT policies that block local interpreters, tooling and even basic scripting for security reasons, preventing experimentation.
- Legacy systems with few or no APIs and minimal integration points, so agents cannot act against core workflows.
- Isolated engineering teams or outsourced development that lacks deep domain knowledge to build effective automations.
Security, accuracy and compliance risks: what to watch for
AI introduces a layered set of risks that cannot be ignored:- Hallucinations and factual errors: Models can produce plausible but incorrect outputs, requiring human verification and structured validation.
- Insecure generated code: Recent analyses show a large share of AI-generated code contains security flaws unless guarded by static analysis and secure-by-design policies.
- Data leakage: Agents accessing sensitive systems must be strictly controlled by least-privilege APIs and monitored ingress/egress rules.
- Operational brittleness: Resource-constrained sandboxes can regularly fail specific workflows (e.g., decompressing and parsing XLSX), producing inconsistent user experiences and eroding trust.
Practical playbook: how organizations should approach AI to boost productivity
Below are actionable guidelines distilled from Alderson’s argument and industry best practices.- Prioritize API enablement
- Catalog internal APIs, instrument them for monitoring, and make them discoverable to non-developers via secure UIs. CIOs increasingly call APIs the cornerstone of agentic AI for good reason.
- Start with observability-first sandboxes
- Use ephemeral microVMs or gVisor-based sandboxes to run agent code, but ensure telemetry captures resource usage, errors and call chains so failures are diagnosable. Cloud vendor agent-sandbox docs and open-source frameworks provide blueprints.
- Treat internal AI as a product
- Define SLOs, error budgets and onboarding flows for non-engineers. Publish examples, templates and reusable “agent recipes.” Build a governance path that balances speed and risk.
- Empower domain teams to experiment
- Give small, cross-functional teams capacity to test agentic workflows against a staging environment and iterate quickly. Maintain a central catalog of proven automations for scaling.
- Bake security and validation into the flow
- Use automated static analysis for code outputs, schema checks for structured data, and human checkpoints for critical writes. Veracode-style analysis shows security flaws are common in AI-generated code unless remediated.
- Measure real productivity, not perceived speed
- Track downstream metrics (time-to-decision, defects, review cycles) rather than proxies like lines-of-code or number-of-queries. Mixed studies show perceived time savings can differ from measured outcomes.
Where the biggest gains — and failures — will happen
- Gains are most likely where the workflow is well-scoped, the inputs are structured, and the organization has APIs. Functions like customer support triage, first-draft marketing assets, and narrow developer tasks (documentation, unit-test scaffolding, repetitive refactors) are frequently productive first use cases. McKinsey’s mapping of 63 use cases highlights concentrated value in customer operations, marketing and software engineering when paired with governance and reskilling.
- Failures are likely where models must replicate tacit expertise or where integration points are sparse — legacy ERPs without APIs, complex multi-file engineering tasks in large codebases, or any context that demands deep institutional knowledge. Controlled field studies show AI can slow experienced contributors when that tacit knowledge dominates the work.
Critical view and caution about unverifiable claims
Some public statements about internal vendor behavior (for example, claims that a large vendor is “rolling out competitor models internally” or that a bundled enterprise tool is universally “the only allowed option”) are difficult to verify externally and often reflect insider anecdotes or selective corporate pilot programs. Alderson reports that Microsoft has made alternative internal tooling choices in some teams — a claim that is plausible given heterogeneous enterprise practices — but it should be treated as anecdotal unless independently confirmed by corporate disclosures or multiple reporting outlets. Readers and decision-makers should therefore treat such claims as signals requiring further verification from vendor statements or procurement records.Bottom line: what IT leaders should do next
- Recognize that AI is an ecosystem problem. Buying a bundled assistant is not enough.
- Invest in internal APIs, developer experience, and a secure, observability-first sandbox platform.
- Promote bottom-up experimentation while building the guardrails that let successful patterns scale.
- Measure the right outcomes — and be prepared for mixed, task-dependent results in the near term.
Conclusion
The AI productivity story in 2026 is not about whether AI works in theory — it does. The question is how organizations stitch it into reality. The divide between power users and light users is more than shorthand; it points to the architectural and governance choices that determine whether AI becomes a sustained productivity lever or a costly, ignored initiative. The most important investments are not the flashiest models, but the plumbing — APIs, secure sandboxes, observability and the social processes that let frontline teams build, validate and scale their own automations. Done right, AI will be a force multiplier. Done badly, it will be another set of expensive tools that fail to move the needle. The next year will tell which path most organizations choose.Source: GIGAZINE As AI becomes more widespread, two types of AI users are emerging. How can AI improve productivity?
