Docusign has opened its connector to the AI assistants people already use. Claude, ChatGPT, Gemini, Copilot, Slack's assistant and Perplexity can now send documents for signature, check who has signed, and search contracts for things like the governing law or the expiry date. Some of this already worked during the beta, which began in February with Claude. What changed is that it is a supported product, and that it adds two things: a way to upload up to 10 local files for signing from inside Claude, and tools for contract lifecycle workflows that large companies run.
Chasing is the work that gets skipped
Sending a contract takes a few minutes and everyone does it. Chasing is different. Nobody has it as a job, so it happens when someone remembers. A sales operations manager could ask the assistant, every Monday at 9 a.m., for each contract she sent that has waited on a signer for more than three days, with the signer's name and the date it went out. That is a list of phone calls, and it arrives without anyone opening a dashboard. A legal operations lead could do the same for agreements approaching an automatic renewal, so the cancellation notice goes out before the price renews for another year.
Because the assistant works as the person who signed in, it can only see what that person can see. The same holds for a procurement clerk who asks, before sending a new NDA, whether the company already has one with that vendor and when it ends. An account administrator can switch the connection off for the whole account.
Why asking first is the design that works
Docusign's own account of how it got this to production is more useful to a reader than the feature list. A model left to work out every step will sometimes skip one, stop halfway or confuse two similar words, and the mix-up of an envelope with an agreement happened often enough that Docusign built around it. Two fixes did most of the work. When a request is ambiguous, the assistant has to ask the person to choose. And for the steps where a mistake costs something, such as who signs and in what order, the chat shows a short form with the same four steps every time.
The model still does what it is good at around that form. After sending, you can ask what happened to the contract, and it reads the status and tells you who is holding it up. Ask it to look for renewals in the next 30 days and it will find them and suggest what to do. The principle is to be strict where an error is expensive and let the model use judgment where it only has to read and summarize.
I expect that split to become the normal shape of every assistant that touches a company's records. Invoices, payroll and customer files will get the same pairing: free searching and summarizing, a question when the request is unclear, and a form for the part that moves money or names a person. If you build your own small tools with AI, copy it before you add features.
What to check before you rely on it
The tools you get depend on your Docusign plan, and the connector works in English only, with other languages left to the assistant. It is not yet available for Docusign's government accounts. Claude, ChatGPT and the other assistants ask for your approval before they act, but a scheduled job or a custom agent does not, so whoever builds it has to add that step.
There is one risk to own. Free-text fields in a contract can contain anything a counterparty typed, including a sentence written to look like an instruction to the assistant. Docusign marks those fields as untrusted so the assistant treats them as data. Even so, a scheduled task that only reads and reports is a safer first job than one that sends, and it is the one that finds the contract you forgot.