The Day Office Moved Inside Copilot── A new base for work, an audit with no rules, agents in clinical trials
Download the video (MP4, 11 MB)
Tomorrow, the first screen you open to draft a document may differ from today's. Writing and spreadsheets are starting to be called up from inside a chat with an AI. When the place where a task begins moves, the place where its history is kept moves too. Today we carry one question. As AI moves to the centre of our jobs, what can we still verify?
The story runs from how a task starts to how its trail is kept. It then reaches drug development, and how we test the guardrails built into AI. Pharma has always relied on checking things after the fact. That assumption is now shifting.
01Office Runs Inside Copilot
Source Ars Technica / SiliconANGLE / 조선일보
When we prepare a document, we first open a familiar program. Where will that very first click happen from now on? That is where we begin.
| Phrase used at launch | Outlet | What it implies |
|---|---|---|
| The new OS for work | Ars Technica | Copilot becomes the base that hosts other apps |
| Office runs inside Copilot | Ars Technica | Copilot becomes the entry to Word, Excel and PowerPoint |
| Multi-player era | 조선일보 | Framing for the Copilot Workspaces launch |
| Autonomous engine of enterprise work | SiliconANGLE | Copilot recast from helper to the one doing the work |
All within the headlines reporting the Ignite 2026 announcements
A large software maker has put its AI chat screen at the start of every task. Documents, spreadsheets and slide decks are summoned from that screen. You tell the AI what you need before you pick a program. The outlets chose different words, but they point the same way. The AI is moving from assistant to the one driving the job forward.
One starting point is clearly convenient. But the history of what you showed the AI, and what it produced, will pool in that single place too. To trace events later, you will open that screen's history first.
For a drugmaker, this is no small matter. Internal files hold interim study results and unreleased information. If the AI screen is where the day begins, decide in advance where such information flows and stays. Staff need a clear line on what they may type into the AI.
Once one company holds the starting point, whose products will sit on that screen? Next, we look at the outside firms joining it.
02Microsoft Not Tied to One Partner
Source Seeking Alpha / Breakingthenews.net / Axios / The Futurum Group
The firm that built this starting screen is not filling it with its own products alone. I think that is the real aim behind the move.
| What is added | Where it goes | What it implies |
|---|---|---|
| Salesforce (Copilotforce) | Microsoft Copilot | Another firm's business software usable inside Copilot |
| Databricks Genie | Copilot, Teams, Excel | Data analysis tools also reachable from Copilot |
| DeepSeek inference and Baseten | Microsoft Foundry | Choice of other firms' models, including Chinese ones |
| 3 Meta models | Microsoft Foundry | More model options beyond OpenAI |
Axios: Microsoft also aims at cheaper, specialized models to rival OpenAI and Anthropic
One vendor's customer software, and another's way of reading data, can now be called up from the same screen. For the engines that generate answers, customers can pick from many sources, including new overseas developers. The firm is plainly stepping away from relying on its long-time partner.
Why widen the choice so far? If you own the screen users open first, you keep the customer, whatever engine runs underneath. That is how I read it. Competition seems to be shifting from raw engine performance to who owns that first screen.
For users, wider choice is welcome. But engines from different developers may swap in the middle of one task. In pharma, we must be able to explain later which answer came from which source. More parts make that harder. So set in advance which programs are allowed for each type of task.
More parts raise a hard question about keeping a trail. Yet the rules for that trail still have almost nothing in them.
03An Audit Standard for AI Agents
Source Investing News Network / Tech Times
If AI works without waiting for people, we need a trail to look back on. How far have the rules for that trail come?
The same firm has begun, with many industries, an effort to make AI more visible. It also launched a shared benchmark for checking what AI does. Yet nobody has said what must be written down, in what form, or by when. The substance is still to come.
An AI that runs its own steps moves on without a person checking each one. So what it touched can only be traced through files kept outside it. If every company keeps those files differently, nobody can compare or verify them.
Without a yardstick, on the day something goes wrong, you cannot replay events in order. Reports do not say why the rules are slow. Still, checking seems to lag behind adoption.
Pharma has always kept trust through documentation. Study procedures survived review because people logged who did what, and when. Handing tasks to AI should demand the same. Each company can already decide which items it will keep.
Even before these rules settle, money is already starting to flow into AI for making medicines.
04What Big Pharma Pays Outside Firms
Source TradingView / www.marketscreener.com / grafa.com
So far we have looked at those selling AI. What about the buyers? Where are the big drugmakers starting to spend?
| Partners | What was reported | What it implies |
|---|---|---|
| AstraZeneca × Tempus AI | Oncology deal extended; Tempus AI expected to receive $200M | Amount to be received; may not be the total deal value |
| Novo Nordisk × nference | AI partnership launched, focused on heart and metabolic diseases | No amount given; scale unknown |
| DataMeds AI × DiaGen | Partnership in cancer drug discovery | Terms not in the headline |
All within headlines. Big pharma money is going to oncology data and analytics firms
The news was about large drugmakers teaming up with specialist partners. In cancer, a long-running contract was renewed, and the expected payment to the partner was reported. Another big firm signed a new tie-up on illnesses of the heart and metabolism. Each one borrows a partner's patient files and analytic skill.
But the figure in the headline is what the partner expects to get. It cannot tell us the value of the whole contract. The other tie-ups disclosed no money at all. The bigger a number looks, the more we must check what it measures.
For a drugmaker, partnering also means handing over patient histories and study results. Before the price, the contract must set what is shared and how far the partner may use it. A renewal is a chance to revisit those terms.
The flow of money and patient files now reaches beyond research, into testing medicines in people.
05In Pharma Operations
Source Unite.AI
Testing medicines in people is the slowest and costliest stage. What changes when AI steps in?
What was announced
Oracle announced AI agents that coordinate clinical trial operations. The target is the work of those who run trials.
What is not reported
The headline gives no results by trial phase and no figures from use. It is at the stage of a tool supporting trial operations.
Where it sits for pharma
Not a tool for making drug candidates. Oracle supplies a trial-operations platform that drugmakers could adopt.
A large software firm has unveiled AI that handles the logistics for those who run studies. Its target is not the medicine itself, but the effort of keeping a study moving. No figures on results from use appear in the reporting. It covers back-office tasks, not the tasks facing doctors and hospitals.
At a study site, many hospitals and many people follow set procedures. One late step can push back the whole timeline. If AI handles this coordination, studies might get shorter. That is still only a hope. But patients waiting for the medicine would gain.
Study management is where documentation under national rules carries the most weight. Who checked what can always be examined later. If AI handles logistics, what the AI did must be kept with equal weight. Ask a vendor how that trail is kept before choosing a product.
If we hand logistics to AI, we also need to check that the guardrails inside it really work.
06Safeguards Broken in Undercover Tests
Source Fox News / California State Portal | CA.gov
AI developers say they build in mechanisms that refuse dangerous requests. So how well do those mechanisms hold when outsiders test them?
| Event | What was reported | What it implies |
|---|---|---|
| US government auditors' undercover probe | Safeguards of ChatGPT, Claude, Gemini, Grok bypassed for terror plans | Makers' safeguards can break when tested from outside |
| California announcement | Next steps on mandatory AI safety standards and oversight | Moving away from leaving safeguards to makers' discretion |
The auditors' case is a Fox News report. Methods and company responses are not known from the sources
Federal investigators hid their identities and tried several chatbots. According to the report, they slipped past the guardrails and drew out scenarios for attacks on people. It was not one firm's failure. Several systems gave way in the same manner. On the same day, one state announced a further step toward enforcing its rules.
Why do developers' guardrails fail? They are built around the risky requests the developer can imagine. Outsiders ask in ways beyond those assumptions. So in-house testing alone cannot find every weakness. Note that a single outlet carried this story.
The AI that drugmakers bring in carries the developer's guardrails too. But they may not hold in your own setting. If you handle patient information or unreleased study results, test against your own scenarios before rollout. Notes of those tests also support you in a later inspection.
Do not take a developer's word on trust. Test it yourself. That stance ran through every scene today.
Whether Microsoft's agent audit standard will set a log format and a deadline.
Open the full transcript
Intro
Tomorrow, the first screen you open to draft a document may differ from today's. Writing and spreadsheets are starting to be called up from inside a chat with an AI. When the place where a task begins moves, the place where its history is kept moves too. Today we carry one question. As AI moves to the centre of our jobs, what can we still verify?
The story runs from how a task starts to how its trail is kept. It then reaches drug development, and how we test the guardrails built into AI. Pharma has always relied on checking things after the fact. That assumption is now shifting.
CH 01 Office Runs Inside Copilot
When we prepare a document, we first open a familiar program. Where will that very first click happen from now on?That is where we begin.A large software maker has put its AI chat screen at the start of every task. Documents, spreadsheets and slide decks are summoned from that screen. You tell the AI what you need before you pick a program. The outlets chose different words, but they point the same way. The AI is moving from assistant to the one driving the job forward.One starting point is clearly convenient. But the history of what you showed the AI, and what it produced, will pool in that single place too. To trace events later, you will open that screen's history first.For a drugmaker, this is no small matter. Internal files hold interim study results and unreleased information. If the AI screen is where the day begins, decide in advance where such information flows and stays. Staff need a clear line on what they may type into the AI. Once one company holds the starting point, whose products will sit on that screen?
Next, we look at the outside firms joining it.
CH 02 Microsoft Not Tied to One Partner
The firm that built this starting screen is not filling it with its own products alone. I think that is the real aim behind the move.One vendor's customer software, and another's way of reading data, can now be called up from the same screen. For the engines that generate answers, customers can pick from many sources, including new overseas developers. The firm is plainly stepping away from relying on its long-time partner.Why widen the choice so far?
If you own the screen users open first, you keep the customer, whatever engine runs underneath. That is how I read it. Competition seems to be shifting from raw engine performance to who owns that first screen.For users, wider choice is welcome. But engines from different developers may swap in the middle of one task. In pharma, we must be able to explain later which answer came from which source. More parts make that harder. So set in advance which programs are allowed for each type of task. More parts raise a hard question about keeping a trail. Yet the rules for that trail still have almost nothing in them.
CH 03 An Audit Standard for AI Agents
If AI works without waiting for people, we need a trail to look back on. How far have the rules for that trail come?The same firm has begun, with many industries, an effort to make AI more visible. It also launched a shared benchmark for checking what AI does. Yet nobody has said what must be written down, in what form, or by when. The substance is still to come.An AI that runs its own steps moves on without a person checking each one. So what it touched can only be traced through files kept outside it. If every company keeps those files differently, nobody can compare or verify them.Without a yardstick, on the day something goes wrong, you cannot replay events in order. Reports do not say why the rules are slow. Still, checking seems to lag behind adoption.Pharma has always kept trust through documentation. Study procedures survived review because people logged who did what, and when. Handing tasks to AI should demand the same. Each company can already decide which items it will keep. Even before these rules settle, money is already starting to flow into AI for making medicines.
CH 04 What Big Pharma Pays Outside Firms
So far we have looked at those selling AI. What about the buyers?Where are the big drugmakers starting to spend?The news was about large drugmakers teaming up with specialist partners. In cancer, a long-running contract was renewed, and the expected payment to the partner was reported. Another big firm signed a new tie-up on illnesses of the heart and metabolism. Each one borrows a partner's patient files and analytic skill.But the figure in the headline is what the partner expects to get. It cannot tell us the value of the whole contract. The other tie-ups disclosed no money at all. The bigger a number looks, the more we must check what it measures.For a drugmaker, partnering also means handing over patient histories and study results. Before the price, the contract must set what is shared and how far the partner may use it. A renewal is a chance to revisit those terms. The flow of money and patient files now reaches beyond research, into testing medicines in people.
CH 05 In Pharma Operations
Testing medicines in people is the slowest and costliest stage. What changes when AI steps in?A large software firm has unveiled AI that handles the logistics for those who run studies. Its target is not the medicine itself, but the effort of keeping a study moving. No figures on results from use appear in the reporting. It covers back-office tasks, not the tasks facing doctors and hospitals.At a study site, many hospitals and many people follow set procedures. One late step can push back the whole timeline. If AI handles this coordination, studies might get shorter. That is still only a hope. But patients waiting for the medicine would gain.Study management is where documentation under national rules carries the most weight. Who checked what can always be examined later. If AI handles logistics, what the AI did must be kept with equal weight. Ask a vendor how that trail is kept before choosing a product. If we hand logistics to AI, we also need to check that the guardrails inside it really work.
CH 06 Safeguards Broken in Undercover Tests
AI developers say they build in mechanisms that refuse dangerous requests. So how well do those mechanisms hold when outsiders test them?Federal investigators hid their identities and tried several chatbots. According to the report, they slipped past the guardrails and drew out scenarios for attacks on people. It was not one firm's failure. Several systems gave way in the same manner. On the same day, one state announced a further step toward enforcing its rules.Why do developers' guardrails fail?
They are built around the risky requests the developer can imagine. Outsiders ask in ways beyond those assumptions. So in-house testing alone cannot find every weakness. Note that a single outlet carried this story.The AI that drugmakers bring in carries the developer's guardrails too. But they may not hold in your own setting. If you handle patient information or unreleased study results, test against your own scenarios before rollout. Notes of those tests also support you in a later inspection. Do not take a developer's word on trust. Test it yourself. That stance ran through every scene today.
Wrap-up
Today's answer is simple. AI has entered the start of our tasks, begun to act on its own, and drawn money into medicine-making. Yet the rules to check what it did, and the means to test its guardrails, have not caught up. So users must decide first what to keep and how to check. Tomorrow, we see whether those rules gain substance.
- CH 01Ars Technica「Copilot is "the new OS for work," and now Office runs inside of it」 arstechnica.com
- CH 01SiliconANGLE「At Ignite 2026, Microsoft reinvents Copilot as the autonomous engine of enterprise work」 siliconangle.com
- CH 01조선일보「Microsoft Unveils Copilot Workspace, Declares Multi-Player Era」 chosun.com
- CH 02Seeking Alpha「Microsoft Foundry adds Baseten, DeepSeek AI inference (MSFT:NASDAQ)」 seekingalpha.com
- CH 02Breakingthenews.net「Microsoft to add three Meta AI models to Foundry」 breakingthenews.net
- CH 02Axios「Microsoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI models」 axios.com
- CH 02The Futurum Group「Copilotforce Partnership Puts Salesforce Inside Microsoft Copilot」 futurumgroup.com
- CH 03Investing News Network「Microsoft and organizations across industries launch AI Transparency Initiative to strengthen enterprise trust in AI models and agents」 investingnews.com
- CH 03Tech Times「Microsoft Unveils AI Agent Audit Standards Initiative With No Spec, Log Format or Deadline」 techtimes.com
- CH 04TradingView「AstraZenecaがオンコロジー分野の契約を延長、Tempus AIは2億ドルを受け取る見通し」 tradingview.com
- CH 04www.marketscreener.com「Novo Nordiskがnferenceと、心疾患・代謝疾患に焦点を当てたAI提携を開始」 marketscreener.com
- CH 04grafa.com「DataMeds AIがDiaGenと、がん創薬で提携」 grafa.com
- CH 05Unite.AI「Oracle Introduces AI Agents to Coordinate Clinical Trial Operations」 unite.ai
- CH 06Fox News「Government investigators find AI platforms produce terror plot plans in undercover op」 foxnews.com
- CH 06California State Portal | CA.gov「Governor Newsom’s nation-leading AI safety framework is getting even stronger as California releases next steps on mandatory safety standards and oversight」 gov.ca.gov
Articles used
- AI Daily News 2026-11-18 Evening
- AI Daily News 2026-11-19
- AI and the Pharmaceutical Industry 2026-11-19
Today's related reports
- AI Daily News November 19, 2026
- AI & Economy News November 19, 2026
- AI & Finance News November 19, 2026
The narration is synthetic speech. The content draws only on this site's articles from the same day. Each edition passes seven plain-language checks before publication. That means known obstacles to comprehension are held below threshold — it is not a guarantee of comprehension.