From Generators to Doers

Mustafa Suleyman, a DeepMind cofounder and now CEO of Microsoft AI, argues that AI is evolving from passive content generation to active assistance that can take actions. He believes the future is about agents capable of planning, coordinating, and executing tasks—not just chatting with you. A recent MIT Technology Review interview emphasized this perspective, portraying generative AI as a stepping stone toward more advanced, interactive systems (The Verge). If you’ve followed Suleyman since his Inflection AI days and read his book “The Coming Wave,” this concept isn’t new—but the timing is different: the tools are finally advancing.

The next wave of AI will not just generate text or images. It will take actions for you, end to end.

This shift is already evident. Leading labs and platforms are developing agents that schedule meetings, file expense reports, triage customer emails, and even orchestrate complex workflows across various applications.

What is Interactive AI, Exactly?

Interactive AI refers to systems that do more than just respond with content. They manipulate tools and services to achieve goals, often requiring multiple steps. Imagine an assistant that can compare flights, redeem credits, book the right itinerary, and notify your team on Slack. It’s not simply a prompt-and-reply interaction; it’s an ongoing collaboration with memory, planning, and the ability to execute tasks.

How It Differs from Generative AI

• Goal-driven: You specify the desired outcome, and the agent determines the necessary steps.
• Tool-using: Agents can call APIs, browse, fill out forms, and control applications.
• Stateful: They retain context across actions and over time.
• Supervised: The most effective agents still operate with guardrails and require human approvals for sensitive actions.

Why This Phase is Different

Three significant developments have come together, making interactive AI more practical:

• Model Capabilities: Leading models have improved in reasoning, planning, and tool usage over the past year. Research like ReAct and Toolformer have paved the way for reliable action-taking (ReAct) (Toolformer).
• Agent Frameworks: Major vendors now provide tools for creating agents, including OpenAI’s Assistants API, Google’s Project Astra, and Microsoft’s Copilot agents (OpenAI) (Google DeepMind) (Microsoft).
• Enterprise Readiness: Features like identity management, permissions, logging, and monitoring are becoming standard, making it feasible to safely deploy agents in real workflows, not only in demos.

Real-World Examples You Can Picture

• Travel Concierge: You tell an agent, “Book me a one-day NYC trip next Thursday under $600.” It searches, applies credits, books flights and a hotel, then emails you the receipts.
• Customer Support Triage: An agent reads incoming emails, checks the CRM, drafts replies, updates tickets, and escalates complex cases to humans.
• Sales Operations: It pulls pipeline data, identifies at-risk deals, drafts outreach, and schedules follow-ups, all in accordance with your guidelines.
• Finance Automation: It reconciles invoices against purchase orders and flags any anomalies for review.
• Developer Copilot, Upgraded: Beyond offering code suggestions, an agent files bugs, opens pull requests, runs tests, and writes release notes. Early systems like Auto-GPT and SWE-agent hinted at this direction (Auto-GPT).

What the Big Players Are Building

Interactive AI is not a distant dream; it is a priority roadmap across various labs and platforms:

• Microsoft: After bringing Suleyman and the Inflection team on board, Microsoft has been rolling out agentic features within Copilot and Copilot Studio to assist in the creation of task-specific agents, complete with approvals and audit trails (Microsoft) (Docs).
• OpenAI: GPTs and the Assistants API enable developers to connect models to tools, files, and functions, forming the backbone of many emerging agents (OpenAI) (Assistants API).
• Google DeepMind: Project Astra demonstrates real-time, multimodal agents that can see, speak, and respond swiftly, a crucial component for interactive work (Google DeepMind).
• Anthropic: Claude models have enhanced tool usage and grounding, offering guidance for safer enterprise deployments (Anthropic).
• Adept: Early work like ACT-1 showcased agents that can operate software UIs for users (Adept).

The Guardrails Question: Can We Trust Agents to Act?

Allowing software to take actions raises important safety and security questions. The good news is that there’s an expanding framework for managing these risks. The NIST AI Risk Management Framework outlines practical controls for trustworthy AI, including transparency, accountability, and ongoing monitoring (NIST AI RMF). Governments have also set priorities with initiatives like the 2023 UK AI Safety Summit and the Bletchley Declaration (UK Government).

Practical Design Principles

• Least Privilege: Provide agents with the minimal permissions necessary for each task.
• Human-in-the-Loop: Require approvals for significant actions, such as purchases, data deletions, or customer outreach.
• Transparency: Log every action with clear explanations and links back to the underlying prompts and tools used.
• Safe Tool Use: Whitelist tools and APIs; verify outputs that involve money, privacy, or security.
• Recovery Plans: Ensure it’s easy to revert changes and revoke access if something goes wrong.

How to Get Started with Interactive AI in Your Organization

Teams that learn by doing tend to gain momentum. Here’s a phased approach that keeps risks manageable.

1. Select a narrow, well-defined workflow. Examples include expense report preparation, meeting scheduling, or drafting RFP boilerplates.
2. Introduce tools safely. Start with read-only access, then gradually expand to limited write actions with approvals.
3. Instrument everything. Keep track of prompts, actions, outcomes, and user feedback to enhance reliability.
4. Run in shadow mode. Let the agent suggest actions while humans carry them out, comparing results before granting full autonomy.
5. Set red lines. For sensitive data or regulated steps, maintain full human control.

Most modern technology stacks already support this approach. For instance, Microsoft Copilot Studio, OpenAI’s Assistants API, and cloud workflow tools can easily connect an agent to calendars, email, CRMs, and billing systems with audit log capabilities (Copilot Studio) (Assistants API).

Limits and Open Questions

While interactive AI is powerful, it’s not infallible. Here are some potential challenges to expect:

• Fragility in Long Chains of Actions: Even minor errors can disrupt a plan, so robust retries and checks are crucial.
• Cost-Performance Tradeoffs: Tool calls and monitoring add latency and expenses; it’s best to start with high-value workflows.
• Evaluation Gaps: Measuring agent reliability remains a research area, but new benchmarks are being developed.
• Policy and Consent Challenges: Acting on behalf of users across terms-of-service boundaries requires careful legal consideration.

However, the trajectory is clear. Each release cycle enhances planning, memory, and tool usage. That’s why leaders like Suleyman view generative AI as merely a phase leading to something more capable and valuable in everyday tasks.

Conclusion

Generative AI has demonstrated that machines can produce useful content. Interactive AI will show that they can deliver substantial outcomes. If you manage a team or a product, now is the time to experiment with narrow, auditable agents and build the organizational capabilities you’ll need. The transition from generators to doers has begun. Those who learn to harness it safely will secure a lasting advantage.

FAQs

What is the difference between generative AI and interactive AI?

Generative AI creates content like text, images, or code based on prompts. In contrast, interactive AI takes actions to achieve goals, utilizing tools, planning steps, and often requiring approvals for sensitive tasks.

Are AI agents safe to use at work?

They can be safe when deployed with appropriate guardrails, such as least-privilege access, human approvals for high-risk actions, clear logging, and continuous monitoring. Frameworks like NIST’s AI RMF provide practical guidance for trustworthy deployments (NIST).

Will interactive AI replace jobs?

AI agents will automate some aspects of workflows, especially repetitive digital tasks, potentially shifting roles towards oversight, exception handling, and higher-value work. The impact will vary by industry and role.

How close are we to reliable end-to-end agents?

Basic use cases exist today in controlled environments. More complex, cross-application workflows are rapidly improving, supported by enhanced models and stronger platform tools from companies like Microsoft, OpenAI, and Google.

What tools should I try first?

Starting points could include Microsoft Copilot Studio for enterprise agents, OpenAI’s Assistants API for developer-focused workflows, and vendor-specific integrations for your CRM, help desk, and billing systems.

Sources

1. MIT Technology Review interview summary via Google News
2. Microsoft: Mustafa Suleyman joins as CEO of Microsoft AI
3. The Verge: Mustafa Suleyman on interactive AI
4. OpenAI: Introducing the Assistants API
5. OpenAI: Introducing GPTs
6. Google DeepMind: Project Astra
7. Microsoft: Copilot Studio agents overview
8. Research: ReAct – Synergizing reasoning and acting in language models
9. Research: Toolformer – Teaching LMs to use tools
10. Auto-GPT open-source project
11. Anthropic: Tool use and agents
12. NIST AI Risk Management Framework
13. UK AI Safety Summit: Bletchley Declaration
14. Adept: ACT-1 for operating software