SHEYU.AI舍予基业
AI Daily Brief

Microsoft is putting AI agents to work in the lab; what that means for smaller firms

Microsoft, Anthropic and Amazon are embedding AI into continuous research workflows, a signal for smaller firms to hand over their own repetitive processes.

·4 min read
Microsoft is putting AI agents to work in the lab; what that means for smaller firms
AI-generated illustration, not a news photograph

Microsoft is turning AI from a question-and-answer tool into something that can take over a stretch of continuous work. For smaller companies, the lesson is not about drug discovery but about which of their own processes to hand over first.

The big platforms are buying a place in the workflow

A report by TMTPost notes that in June 2026 Microsoft Discovery opened to enterprises, allowing researchers to dispatch specialised AI agents that connect scientific software, experimental data and different tasks. Microsoft is also working with Ginkgo Bioworks to explore the link from data analysis and hypothesis generation to automated experiment execution. The same report says Anthropic disclosed partnerships with six institutions in October 2026, and that Amazon reached a strategic partnership with Novo Nordisk.

These names are far from most businesses. But the structure of the news is close to home: the big platforms are rebuilding AI from a question-answering tool into a role that can take over a continuous workflow.

What the giants are really competing for

One line in the report is worth underlining: Microsoft wants AI to do more than answer isolated questions, and to take part in a continuous research process. Nvidia sells computing power and model tools, Amazon provides cloud and data infrastructure, and Microsoft and OpenAI are competing over research platforms and workflows.

The three take different routes but land in the same place. Whoever embeds AI first into a process that others run every day, and get wrong every day, can charge for it over the long term.

What does this mean for smaller companies? You do not need to make drugs, but you certainly have continuous processes: a quotation that travels from Excel to contract, an order that travels from WeChat to dispatch, application materials collected from several departments and assembled into one document. Today these processes depend on people watching them. Tomorrow AI will watch them.

As the big platforms push AI deeper into workflows, the opportunity for smaller firms is not to use the most expensive model, but to hand over the stretch of work that is repetitive and error-prone.

The counter-argument: AI cannot even make drugs, so why hurry

The most common objection is that drug discovery is far from practical use, that AI has not put a candidate drug into human trials, and that smaller companies have no reason to join in.

That is half right. The report is clear that Isomorphic Labs' own candidate drugs and the early-stage projects Anthropic acquired are a long way from becoming usable medicines, and that the final judgment will rest on clinical trial results.

But the objection confuses technical maturity with the timing of process change. The willingness of big platforms to build in a high-barrier field like research shows one thing: AI's ability to take on continuous tasks is moving from demonstration to deliverable. Wait until it is fully mature and you will not even be able to say which part of your own process should be handed over first.

When does this judgment not hold? If your business has no high-frequency, repetitive process with a clear standard of right and wrong, there is indeed no rush. But in reality that is exactly where most smaller companies feel pain: not a lack of process, but processes that live entirely in the head of an experienced hand.

Find the thing that repeats every day and goes wrong every day

Shu Yi, executive director of Sheyu Jiye, has a saying: when going digital, first find the thing that repeats every day and goes wrong every day, and start there. Another: the value of industry software lies in turning the experience in an experienced hand's head into a process anyone can follow.

Put those two lines beside this news and the logic holds. What the big platforms are doing inside research workflows is essentially what you should be doing inside your own business: turning experience into a process that can be checked.

Take applications for technology programmes. This is a textbook continuous process: tens to hundreds of pages of material, every figure traceable to its source text and page number, every conclusion traceable back to the original guidelines. What SHEYU ZHISHEN (舍予AI智审) can do is turn application materials into a checkable evidence table, with every extracted value carrying a page number and source text, cross-checked by two engines, with four-value results not passed by default, and data kept inside the internal network. What it cannot do is make the application decision for you, or replace the reviewers.

That is the boundary. The tool puts the numbers you need to see in front of you; the decision is still yours to make.

One thing you can do now

There is no need to rush out and buy a tool. Spend half an hour writing down the task your team repeated most this week and had to redo most often. Note where it starts, whose hands it passes through, and at which step it most often goes wrong.

If you cannot write it down, you are not ready to take on AI. If you can, you know what kind of tool to look for.

The big platforms are laying the foundation. Whether you can build on it depends on how clearly you understand your own processes.

閱讀繁體中文版 →