AI productivity software is being rebuilt around a broader architecture than the chatbot. Product launches and research released through May 15 point to enterprise systems that combine foundation models with company data, planning, memory, application access, security controls and human oversight.
The shift matters because a capable model alone is not enough to complete consequential work across a business. Microsoft’s workplace-agent research has highlighted how sharply agent performance can deteriorate when tasks become concurrent and long-running, while IBM, Microsoft, Google, OpenAI and Databricks are all advancing pieces of the infrastructure required to make agents more useful—and more governable—in production.[1][2][3][4][5]
From assistants to coordinated systems
The emerging productivity stack treats AI as a software system rather than a single interface. A language or multimodal model supplies reasoning, generation, document understanding and tool-selection abilities. But enterprise value depends on what surrounds that model: governed access to business information, orchestration of multi-step work, reliable execution in existing applications, and controls that limit what an agent can read or change.
That framing was central to IBM’s May 5 announcements at Think. IBM expanded its watsonx Orchestrate, data and hybrid-cloud offerings and described an AI operating model built around four connected systems: agents, real-time data, automation and hybrid infrastructure. IBM Chairman and CEO Arvind Krishna argued that companies must redesign how work operates, rather than simply add more AI tools.[4]
Microsoft’s May 2026 Work Trend Index similarly suggests that workplace AI use is becoming more complex. The company said its privacy-preserving analysis of more than 100,000 Copilot chats classified 49% as cognitive work, including analysis, problem-solving, evaluation and creative thinking. Another 19% involved working with people, 17% producing work and 15% finding information. Microsoft said 16% of surveyed AI users were “Frontier Professionals,” using agents for multi-step workflows and multi-agent systems.[5]
These figures describe usage patterns, not a direct measure of business output. Still, they help explain why technology providers are investing in the connective layers between models and everyday work.
The core architecture of AI productivity
Grounded models and enterprise context
Foundation models remain the reasoning engine, but they need current, permissioned business context. That usually means retrieval-augmented generation, document parsing, semantic indexes, structured databases, event streams and permissions-aware search. Without this layer, agents are constrained by training data that may be incomplete, outdated or unrelated to the organization’s operations.
OpenAI said May 15 that Databricks would make GPT-5.5 available for enterprise agent workflows through its AI Unity Gateway, AgentBricks and Agent Supervisor API. On Databricks’ OfficeQA Pro benchmark, which tests work involving scanned PDFs, legacy files, parsing, retrieval and grounded reasoning, GPT-5.5 reached 50% accuracy and reduced errors by 46% versus GPT-5.4, according to the companies.[6] Those are vendor-reported benchmark results rather than independent measurements of workplace productivity, but they underline the importance of document understanding and retrieval in enterprise automation.
Planning, memory and orchestration
An agent that handles more than a simple request needs to turn goals into ordered steps, select tools or specialist agents, track dependencies, retry failed actions and determine when a person should intervene. This is the orchestration layer. It increasingly includes state management: task status, procedural knowledge, user context and tenant-specific boundaries that must persist safely across long-running work.
Microsoft Research’s CORPGEN project illustrates why those capabilities matter. The framework evaluates agents in simulated workplace environments where multiple interdependent tasks unfold over five hours. In Microsoft’s tests, completion rates for leading computer-use agents fell from 16.7% to 8.7% as concurrent task loads increased. CORPGEN used hierarchical planning, memory isolation and experiential learning, reaching completion rates up to 3.5 times higher than baseline systems across three agent backends.[1]
The findings are a warning against equating strong performance on isolated benchmarks with reliable operations. Microsoft identified memory overflow and interference between unrelated tasks as key failure modes. For enterprise deployments, keeping workstreams separate and preserving the right state may be as important as improving a model’s raw reasoning score.
Tools and execution environments
Agents create value when they can act, not merely summarize. That requires APIs, connectors, workflow engines and, in many organizations, computer-use interfaces for older software that lacks usable APIs.
Microsoft put Windows 365 for Agents into public preview on May 1 to address that problem. The service provides a managed execution environment in which computer-using agents can operate legacy and line-of-business applications through a virtual desktop. Microsoft positioned it alongside Agent 365, a control plane intended to manage agent identity, policy and governance.[3]
The approach recognizes an inconvenient enterprise reality: many critical processes are still embedded in applications designed for people clicking through screens. Computer-use agents can extend automation into those workflows, but they also make execution controls, auditability and error recovery substantially more important.
Security and governance are part of the product
Giving an agent access to business systems expands both its usefulness and its attack surface. A particularly important risk is indirect prompt injection, in which malicious instructions are embedded in an email, document or other content that an agent retrieves. The model may then treat hostile content as an instruction even though the user never directly submitted it.
Google outlined its response to that threat for Workspace with Gemini on April 2. Its defense-in-depth approach combines configuration fixes, synthetic attack data, machine-learning and LLM-based defenses, model hardening and end-to-end testing across products including Gmail and Docs.[2] The underlying principle is significant: retrieved content is data that may be adversarial, not trusted instruction.
A production architecture therefore needs least-privilege access, identity controls, approval gates for sensitive actions, audit logs, data-loss protections, tenant isolation and sandboxed execution. Governance cannot be bolted on after an agent has been granted broad access to email, documents and operational systems. It must determine what the agent is allowed to see, which tools it may invoke, when it must pause, and how its actions can be reviewed.
Evaluation and observability are equally necessary. Organizations need to measure groundedness, task completion, tool-call success, latency, cost, recovery from failure and policy compliance—not just whether a response sounds plausible. CORPGEN and OfficeQA Pro reflect a move toward testing realistic, multi-step work rather than standalone question-answering.[1][6]
Human review remains the accountability layer
The likely near-term model is supervised automation, not unrestricted autonomy. Microsoft’s Work Trend Index found that AI users ranked quality control of AI output and critical thinking as the two most important human skills in an AI-heavy workplace.[5] In practice, people remain responsible for setting intent, defining acceptable quality, approving high-impact actions and owning the outcome.
There is evidence that AI can improve output in bounded settings. A Stanford study of 5,172 customer-support agents found that a generative-AI conversational assistant increased issues resolved per hour by 15% on average, with larger gains for less experienced workers.[7] That result is meaningful, but it does not establish that general-purpose agents can safely manage complex, cross-functional workflows without close supervision.
Broader adoption data point to the same distinction. McKinsey’s 2025 survey found that 88% of respondents said their organizations regularly used AI in at least one business function, yet only about one-third said their companies had begun scaling AI programs. For agentic AI, 23% reported scaling an agent somewhere in the enterprise and 39% said they were experimenting; no individual business function exceeded 10% reporting scaled agent deployment.[8]
McKinsey also found that only 39% attributed any level of enterprise-wide EBIT impact to AI, with most of those respondents attributing less than 5% of EBIT.[8] Widespread experimentation, in other words, has not automatically produced broad financial impact. Data quality, process ownership, workflow redesign and change management remain prerequisites for measurable gains.
The productivity contest moves beyond model scores
As of May 15, the central development in workplace AI is architectural. Providers are assembling stateful, tool-using systems that connect models to data, software and business processes while attempting to constrain their risks. The strongest near-term opportunities are in knowledge management, IT operations, software engineering, document-heavy work, customer support, marketing and research—areas where tasks can be grounded in available data and checked against clear standards.
The competitive question is no longer solely which model writes the best answer. It is which platform can reliably retrieve the right information, maintain context across work, execute within existing systems, recover from errors, resist adversarial instructions and surface the right decisions for human review. For enterprises seeking productivity gains, that full stack is likely to matter more than any single benchmark result.
Editor’s Take
The useful unit of enterprise AI is no longer a chat window; it is a governed workflow that can retrieve the right records, make a plan, use narrowly scoped tools and leave an audit trail. That is harder to ship than a clever assistant, but it is where durable productivity gains will come from. The practical starting point is a small, high-volume process with measurable outcomes and explicit human approval points—not a broad promise to automate an entire department.
The most important signal in this story is that reliability declines when agents must manage concurrent, long-running work. Model demos still routinely hide this operational reality. Buyers should ask for evidence on task completion, tool-call failure rates, recovery behavior, permission boundaries and review queues under real workloads. A system that succeeds once in a clean demo but loses state or crosses data boundaries in production is not an efficiency tool; it is an expensive source of operational risk.
There is real upside here, especially for document-heavy and legacy-software workflows, but the market will reward vendors that make controls native rather than optional. Watch for products that treat identity, isolation, evaluation and observability as first-class parts of the agent runtime. Those are the components that turn promising model capability into something a business can safely depend on.
References
- Microsoft Research – https://www.microsoft.com/en-us/research/blog/corpgen-advances-ai-agents-for-real-work/
- Google Security Blog – https://blog.google/security/google-workspaces-continuous-approach-to-mitigating-indirect-prompt-injections/
- Microsoft Tech Community – https://techcommunity.microsoft.com/blog/windows-itpro-blog/windows-365-for-agents-now-in-public-preview-run-ai-agents-securely-at-scale/4513479
- IBM Newsroom – https://newsroom.ibm.com/2026-05-05-think-2026-ibm-delivers-the-blueprint-for-the-ai-operating-model-as-the-ai-divide-widens
- Microsoft WorkLab – https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization
- OpenAI – https://openai.com/index/databricks/
- Stanford Graduate School of Business – https://www.gsb.stanford.edu/faculty-research/publications/generative-ai-work
- McKinsey – https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai/
