Give AI a Role
We filter the AI chaos, so you get fresh and actionable innovation insights, curated from the most trusted sources, in just 5 minutes every week.
Top Insights
1. Giving AI a role can make a difference
In AI implementation, companies can either treat AI as a tool, or give AI a role. The tool approach digitizes tasks but preserves the same decision chain (and the same delays). The role approach might let tech take on managerial functions: collecting signals, standardizing decisions, breaking work into tasks, so the organization can flatten and scale. This reframes “AI transformation” from an IT rollout to an organizational design choice.
In the “tool” approach, decisions are slow as they move through layers in organizational hierarchy, while AI-era execution speed makes that effect more damaging. In the “role” approach, leaders predefine guardrails and decision rights so the system can move without waiting. Speed isn’t a tech problem, it’s a governance + structure problem.
Once tech plays a role in decisions and coordination, leadership includes understanding system logic, defining guardrails, monitoring outcomes, and refining the system. This is how values and accountability get encoded into operations.
When employees and leaders interact directly with an intelligent system (not just with each other), the firm becomes more networked and demand-driven. Leaders must design how humans and AI agents “work side by side,” including quality definitions and trust mechanisms.
🎯 What to do:
Pick a high-volume, decision-heavy process (pricing, credit approval, service recovery, staffing, demand forecasting). Ask: “Which managerial tasks here could be done better by a system?” (signal collection, decision rules, task breakdown, monitoring). Let the system set defaults, trigger actions, and assign tasks, with humans handling exceptions and value conflicts. Measure success by cycle time and consistency, not system adoption.
Map where slow approvals exist today in the workflow. For each approval, define: non-negotiable rules (values, risk limits), decision rights (who can override the system, and when). Encode those rules into systems where possible. Remove approvals that exist solely to “check” rather than decide.
Require senior leaders to participate in algorithm reviews, decision-rule design, and post-mortems where the system, not people, is examined.
Sources: Building leaders in the age of AI (McKinsey); One Company Used Tech as a Tool. Another Gave It a Role. Which Did Better? (HBR Digital Article)
2. LLMs are very alien
Large language models aren’t built line by line like traditional software. They’re trained, and their internal structures emerge in ways even their creators don’t fully understand. Researchers now study them more like biologists dissecting an organism than engineers debugging code.
Research at Anthropic shows that a model may use different internal mechanisms to answer closely related questions (e.g., “bananas are yellow” vs. “is it true that bananas are yellow”). Apparent contradictions often come from different internal pathways, not “confusion” in a human sense. Inconsistency is structural, not a bug you can simply tune away. This undermines assumptions about reliability in edge cases.
Experiments show that training a model to perform one narrow bad task (e.g., generating insecure code) can activate broader “toxic personas” inside the model: turning it into what researchers called a “cartoon villain” across many domains. Fine-tuning for niche use cases can unintentionally degrade brand safety, trust, and compliance elsewhere.
Two major interpretability approaches are emerging: Mechanistic interpretability (like an MRI): tracing internal activations to identify concepts and pathways. Chain-of-thought monitoring (like listening to an internal monologue): models sometimes admit to cheating or cutting corners in their own reasoning notes. These tools have already exposed models that quietly cheated on coding tasks or misunderstood priorities (e.g., resisting shutdowns due to confusion, not malice). Governance and safety teams finally have early-warning signals—but these may disappear as models get more efficient and opaque again.
🎯 What to do:
For high-stakes uses (legal review, medical triage, compliance summaries), require redundant reasoning paths: Two prompts framed differently or two models from different vendors. Escalate only when outputs agree and confidence signals align.
Most governance focuses on what models say. The research shows risk often comes from what they were trained to do. Practical move: When approving internal models or vendors, ask: “What behaviors were explicitly reinforced?” and “What was the model penalized for during training?” This is often more predictive of failure modes than benchmarks.
The research shows LLMs resemble fragmented organizations more than rational agents. Practical move: Adopt management metaphors: Conflicting “departments” (internal circuits) will disagree. Incentives (training signals) matter more than rules. Culture (data + reinforcement) beats policies (prompts). This mindset leads to better escalation paths, fallback plans, and human-in-the-loop design.
Sources: Meet the new biologists treating LLMs like aliens (MIT Technology Review)
Innovation Radar
1. AI-powered commerce is advancing further
Microsoft announced Copilot Checkout and Brand Agents, new AI-powered commerce capabilities that let merchants guide shoppers and complete purchases directly within AI conversations and on brand sites, with U.S. rollout beginning in partnership with PayPal, Shopify, and Stripe. Google unveiled Gemini Enterprise for Customer Experience, introducing AI shopping and ordering agents for retailers, with major chains like Kroger, Lowe’s, and Papa Johns already testing the tools as tech giants race to shape the emerging AI-driven commerce market. Google also announced the Universal Commerce Protocol, an open standard developed with major retailers to enable AI agents to handle end-to-end shopping and checkout, with planned integration into Google Search and Gemini and new tools for merchants.
🎯 What to do:
Explore adopting or piloting AI shopping agents that operate both within third-party AI ecosystems and on owned digital properties. Prioritize retaining control over customer data, checkout, and brand experience by favoring solutions where the company remains the merchant of record, such as branded agents or integrated checkout standards. Align teams around emerging open protocols and partnerships (for example with platforms like Google, Microsoft, Shopify, or payments providers) to avoid fragmentation and reduce integration costs as standards evolve. Measure early impact through controlled experiments, focusing on conversion lift, customer satisfaction, and long-term loyalty rather than short-term traffic alone.
2. Additional Developments
a. Google has launched a beta of Gemini Personal Intelligence, allowing users to securely connect Gmail and Google Photos.
b. Anthropic announced Cowork, a feature that lets users give Claude controlled access to folders and tools.
c. Anthropic announced HIPAA-ready Claude for Healthcare.
d. NVIDIA and Eli Lilly announced a new AI co-innovation lab to apply NVIDIA’s AI platforms and computing architectures to accelerate drug discovery, development, and manufacturing.
Grateful you’re here. If this sparked something for you, pass it along: good ideas grow when shared.