Make the AI earn its place - AI Field Notes 001

Welcome

Welcome to the first Clarus AI Field Notes. Every other week, we will filter the flood of AI announcements down to the few changes worth your attention: what is useful now, what is coming next, and where the caveats live.

Read it your way. Take 90 seconds for the signal, spend five minutes trying something useful, or go deeper into the evidence and the more interesting edge of what is becoming possible.

The interesting change is not another writing shortcut. Microsoft is turning Copilot into a place where different models, reusable company skills, connected business systems, and event-triggered work can operate together. That is much more useful, and much more consequential, than a better first draft.

Get the pattern in the 90-Second Brief , turn a PDF into an auditable decision brief in the 5-Minute Win , or Go Deeper into model choice, event-triggered work, company skills, Work IQ, hopeful science, and the approval boundary.

The One Thing

Measure the work before you automate it

Microsoft Digital says its biggest early measurement mistake was failing to instrument the slow business processes before deploying Copilot. Its corrective framework is practical: talk to the people doing the work, identify three to five recurring pain points by role, measure the current process, train by role, deploy in cohorts, measure again, and decide where the reclaimed capacity will go.

MIT researchers reached a compatible conclusion from another direction. Their map of thousands of AI tools shows today's systems clustering around information-heavy work such as writing, research, interpretation, and communication. The map can help locate possibilities; the measured workflow determines whether the business actually improved.

Sources:Microsoft Inside Track - Six steps to proving Copilot and agent value MIT - AI and the Future of Work ontology

The 90-Second Brief

The useful version, before the detail

Where to aim
Find three to five repeated pain points by role, record the present time and outcome, and choose one workflow whose finished result a person can inspect.
What changed
Copilot is becoming a multi-model work surface with event triggers, reusable skills, plugins, and Work IQWork IQ in one glanceWork IQ supplies permission-aware workplace context from Microsoft 365 to Copilot and agents.Open Microsoft Learn in a new tab grounding, not merely a collection of writing aids.
What to protect
Before an agent sends, changes, buys, deletes, or acts in production, require a named human approval.

Sources:Microsoft Inside Track - Proving value from Copilot and agents Microsoft 365 Copilot release notes NIST - Identity and authority for software agents

The 5-Minute Win

Turn one PDF into a decision brief you can audit

Use a non-sensitive proposal, policy, or report whose facts you can verify. The point is not another summary. The point is a working brief that makes missing information and source checking visible.

Read this PDF and return: (1) a five-sentence executive summary; (2) every decision requested; (3) action items with owner and due date, writing "not stated" when either is absent; (4) risks and open questions; and (5) a page reference for every decision, action, and risk. Do not infer missing commitments. Finish with the three statements I should verify most carefully in the source.

Check the result: Open every cited page. Confirm that each decision, owner, date, and risk is supported. If the source does not name an owner or date, the brief must say "not stated." Availability can vary by Microsoft 365 tenant, and this candidate still requires the planned Clarus hands-on check before approval.

Go Deeper

Worth your time, not just more links

These are the pieces we would actually hand to someone who wants the fuller business, product, or technical picture.

Business playbookSix steps to prove Copilot and agent value

Microsoft Digital explains the measurement mistake it made, then gives a role-by-role framework for finding pain, establishing a baseline, deploying, and reclaiming the value.

Open article Product tourThe July Copilot changes, with screenshots

The monthly Microsoft roundup covers model choice, event triggers, Cowork customization, brand assets, connectors, dashboards, and governance in one visual article.

Open article Platform deep diveWhat Work IQ actually provides to agents

The Microsoft Learn overview separates the workplace context layer from the APIs and agent surfaces that use it, including the permission and compliance foundation.

Open article Capability catalogThe current Cowork plugin lineup

Review the Microsoft and partner plugins, what systems they connect, and which capabilities deserve a closer business-fit and security check.

Open article

Microsoft Pulse

Copilot is becoming a multi-model work surface

Microsoft 365 Copilot now offers current OpenAI and Anthropic models inside the same Microsoft work surface. Microsoft's July roundup says GPT-5.6 is available across Copilot experiences and Claude Sonnet 5 is available in Word, PowerPoint, and Cowork. Cowork also lists specialized Anthropic choices for longer or more demanding work.

The useful story is not a model horse race. A person can choose a model for the kind of work in front of them while Copilot remains grounded in the files, meetings, chats, and business data they are permitted to use. Some model options have different preview, enablement, billing, or data-retention terms, so the model label is part of the business decision.

Use a capable daily-driver model for routine extraction, comparison, drafting, and formatting. Deliberately step up to the strongest model when the work is long, ambiguous, multi-stage, or expensive to get wrong. The best model strategy is not “maximum” everywhere; it is enough intelligence for the task, with evidence and human review where consequence rises.

Clarus take: Choose the model for the work, but keep the work inside a governed system. Confirm availability and data terms before promising the experience.

Sources:Microsoft - What's New in Microsoft 365 Copilot, July 2026 Microsoft Learn - Models and data retention in Cowork

The Trigger

Cowork can begin when work happens, not only when you prompt it

Event-triggered tasks let Cowork watch for a matching email, a Teams message or @mention, a keyword, a topic, or a named event. The user describes the trigger in plain language, reviews the proposed automation, and the task appears beside scheduled work.

That can turn recurring monitoring into useful preparation: assemble the background when a priority client writes, prepare a response packet when a project is mentioned, or collect the decisions and open questions after a named meeting. Microsoft described the July rollout as a Frontier capability, so it belongs on the watch list rather than in a casual five-minute experiment.

Clarus take: This is genuinely useful. Start with preparation and a review step; do not begin with unattended sending, deletion, purchasing, or production changes.

Sources:Microsoft Learn - Event-driven tasks in Copilot Cowork Microsoft July Copilot roundup - Event-triggered task example

Skills & Plugins

Package the way your company works, then connect the systems it uses

A skill is reusable guidance: the structure of a client brief, the questions required before a proposal, the organization's writing standard, or the exact handoff checklist for a finished project. Cowork can create skills through a guided conversation or import them from Markdown, ZIP, or skill files, then share them with specific people while the owner continues to update the source.

A plugin adds systems and tools. Five examples worth understanding are Dynamics 365 Sales for pipeline context, Dynamics 365 Customer Service for cases and knowledge, Fabric IQ for governed Power BI and Fabric data, Miro for visual collaboration, and monday.com for team work management. They are examples, not blanket endorsements; the connected data, actions, identity, licensing, and approval path still decide whether a plugin belongs in the business.

The practical opportunity is larger than a prompt library. A company can preserve small pieces of operating knowledge: how it prepares meetings, qualifies opportunities, reviews changes, formats deliverables, or closes work. Then it can make that knowledge consistently available inside the tools employees already use.

Clarus take: Build skills from proven operating practices. Add plugins only when the business system, owner, data boundary, and measurable outcome are clear.

Sources:Microsoft Learn - Customize Cowork with skills and plugins Microsoft Learn - Available plugins for Copilot Cowork

The Foundation

Work IQ is what turns a generic answer into work-aware assistance

Work IQ is Microsoft's workplace-intelligence layer for agents. It combines Microsoft 365 content with context about people, relationships, communications, and ongoing work so an agent can reason over the material the current user is allowed to access.

This is the platform argument that matters for a business. Useful context, permissions, compliance controls, connectors, and auditability can sit under multiple Copilot and Foundry experiences instead of being rebuilt around every isolated AI tool. It does not repair overshared files or replace governance; it makes the quality of that foundation more important.

Clarus take: The value is not that AI knows everything. The value is that it can use the right company context through an accountable platform boundary.

Sources:Microsoft Learn - Work IQ overview Microsoft Foundry - Connect agents to Microsoft 365 with Work IQ

The Spark

Old genomic data produced new rare-disease diagnoses

Genomic knowledge changes after a patient receives an initial test. Talos, an open-source system developed through an international collaboration that includes Microsoft, repeatedly reanalyzes stored genomic data against newer evidence.

In a prospective cohort of 4,735 undiagnosed people, the researchers reported 241 new diagnoses. The system surfaced a small set of changed, high-confidence candidates for expert review, and accredited laboratories confirmed the reported variants.

Clarus take: Good automation does not remove the expert. It helps the expert spend time where new evidence can change an outcome.

Source:Microsoft Research - Talos rare-disease genomic reanalysis

The Future, Shipping Now

Tomorrow called. It's in procurement.

Stage: Field-tested, procured, and filed

Alexander Wissner-Gross calls Lonestar Space the first node of a future Dyson swarm. That is his deliberately enormous frame, not the project's official name, and he discloses a financial interest in 021T Capital, which backs Lonestar. The underlying activity is still worth watching. What is a Dyson swarm? What is a Dyson swarm? Artist's rendering of many independently orbiting collector panels surrounding a star A swarm uses many independent orbiting collectors, not one solid shell. Artwork by Kevin M. Gill, CC BY 2.0.A Dyson swarm is a hypothetical fleet of independently orbiting collectors around a star, built to capture a meaningful share of its energy. It is not a solid shell.Open Space.com in a new tab

Lonestar announced StarVault as a commercial space-based data-storage service in April and said it had procured a second orbital payload from Sidus Space. In June, Lonestar and Space Solar signed a letter of intent to explore hosting those storage modules on orbital power platforms. Meanwhile, the FCC has accepted SpaceX's much larger orbital-data-center application for filing. Accepted for filing means regulators will consider it. It does not mean approved, launched, or economically sensible.

Clarus take: Humanity did not finish a Dyson swarm between lunch and the afternoon status meeting. But orbital storage, power, and compute have crossed from speculative slides into payloads, procurement, and regulatory filings. Watch the stage changes, not the adjectives.

Sources:Space.com - What a Dyson swarm is and how one might be built Physica Scripta - Review and viability of a Dyson swarm Kevin M. Gill - Dyson swarm artwork (CC BY 2.0)The Innermost Loop - The First Dyson Swarm Node Lonestar - StarVault and second orbital payload announcement Space Solar - Letter of intent with Lonestar FCC - SpaceX orbital data center application accepted for filing

The Perimeter

Preparation and permission are different things

An AI agent can gather information, compare options, prepare a draft, or recommend a next step. None of those activities automatically gives it authority to send, post, purchase, delete, change access, or act in production.

Identify the owner, limit what the agent can reach, preserve the evidence, and put a named approval before the consequential action.

Clarus take: AI prepares. A person approves. The organization remains accountable.

Sources:OWASP Top 10 for Agentic Applications NIST - Identity and authority for software agents

Sources

Check our work

Corrections

Change history

No corrections as of 17 Aug 2026. If this issue changes, this section will say what changed, why, and whether the Clarus verdict changed.