How to Judge an AI Teammate That Claims to Work Everywhere

Key Takeaways

• Two big vendors shipped "an AI teammate that follows you" in the same week. Salesforce announced Agentforce Coworker on 2026-08-04 and Alibaba announced the QwenWork beta on 2026-08-03, both on their own blogs.
• Surface count is the weakest thing to compare. Being present in five apps says nothing about whether the agent can finish a job in any of them.
• Ask where the context comes from. Agentforce Coworker draws its context from Salesforce Data 360, which is a real advantage if your business already lives in Salesforce and a hard ceiling if it does not.
• Ask who does the reaching when work leaves the home system. A connector project you have to staff is a different product from an agent that already has read and write access to the tool.
• Ask what happens at the boundary of the suite. QwenWork's beta is a China-only public beta today, and its deepest integration is DingTalk, per Alibaba's own announcement.
• Judge it on one finished deliverable, not a demo. Give the same real request to each candidate and compare what comes back in the thread.

Our head of ops read two launch posts in one week and asked a fair question: if Salesforce and Alibaba both now sell an AI teammate that follows you across apps, what is left to compare? She was not being cynical. She runs a 40-person company, and both announcements described her Tuesday accurately: renewal proposals stuck behind three months of case notes, a standup that needs a pipeline summary nobody has written yet.

The honest answer is that "everywhere" is the least useful part of any of these announcements, including ours. A chat window in five places is still a chat window. What decides whether an AI teammate earns a seat is narrower and duller: where its context comes from, who does the reaching when work leaves its home system, and what it hands back when it is done.

What did Salesforce and Alibaba actually announce?

Two versions of the same idea, each anchored to the vendor's own home turf.

Salesforce's Agentforce Coworker post, published 2026-08-04, describes an autonomous AI teammate built into Salesforce that is powered by Data 360, orchestrates other agents such as a Tableau analytics agent and a Sales Coach agent, calls CRM actions, Flows and third-party APIs from one conversation, and ships with built-in observability. Salesforce describes it as built headless-first so the same teammate appears across Salesforce, Slack, Microsoft Teams, ChatGPT and Claude. Their availability line is more specific than the headline: it is in beta for Salesforce today, with web, Microsoft Teams, ChatGPT, Claude and a desktop app coming later this year.

Alibaba's QwenWork announcement, dated 2026-08-03, describes a workplace agent platform in public beta in China, reachable through a web interface or a desktop client with direct access to local computers, with the Qwen3.8 Max model available on it from 3 August. Alibaba says it will soon be embedded in DingTalk desktop and mobile, and that DingTalk serves more than 20 million enterprises and organizations. The same post lists agent work alongside web app generation and multimodal image, video and audio generation, plus saving a workflow as a reusable skill.

Microsoft made a quieter move in the same week: the GitHub Copilot Agent in Microsoft Agent Framework went stable for .NET and Python on 2026-08-04, with shell execution, file operations, URL fetching and MCP tools, each gated by a permission handler. That one is aimed at developers building agents, not at an ops lead picking one.

Three announcements, three different centers of gravity. None of them is a small player, and none of them is lying about what it does. They are just each bound to something.

Why is "it works everywhere" the wrong thing to compare?

Because presence and capability are different products, and only one of them shows up in a demo.

An agent can appear in your Slack sidebar and still be unable to do the job you need there. If its context comes from one system of record, a request that depends on data outside that system either fails or turns into a connector project. If it can only read, the answer is a summary and the work is still yours. Every AI teammate on the market is bound to something. The three common bindings:

• System-of-record binding. The agent's context is one platform's data model. Inside it, the answers are excellent, because the platform owns the objects. Outside it, someone builds a pipe.
• Suite binding. The agent lives inside one collaboration suite and is best where that suite already is. If your company is not on that suite, the deepest version of the product is not available to you.
• Tool-graph binding. The agent's context is whatever accounts you connect, and its home is the chat where the team already talks. The tradeoff is that it knows nothing about your business until you connect things.

Naming the binding out loud is most of the evaluation. It also tells you where each option honestly wins.

Which three questions actually separate them?

Ask about context, reach and boundaries. In that order, because each one only matters if the previous one holds.

Where does its context come from on the first prompt?

Salesforce is refreshingly direct here: Agentforce Coworker knows your opportunities, forecasts, accounts, cases and contracts because Data 360 already holds them. That is the strongest version of this answer if you are a Salesforce shop with an admin team. If your revenue truth lives in Stripe, your projects in Linear and your client history in a shared drive, the same answer becomes a scoping question.

The version worth asking of any vendor: on day one, before any integration work, what does it know about my company, and where did that come from?

Who does the reaching when the work leaves its home system?

This is where evaluations usually go wrong, because both answers sound the same in a meeting. "It connects to third-party APIs" can mean the agent has an account and credentials, or it can mean your team will wire it up. The test is a single request that crosses two systems neither of which is the vendor's.

@Viktor check which of last month's paid signups have not booked onboarding yet, cross-reference Stripe with HubSpot, and put the list plus a suggested email in this thread for review.

If the answer to that is a project plan, you are buying a platform. That can be the right purchase. Just budget the build too, not only the seat.

What happens at the edge of the suite?

Every one of these products is best inside one boundary. Alibaba states its boundary plainly: public beta in China, DingTalk as the deep integration, with a standalone mobile app and an international edition still ahead. Salesforce's boundary is the platform and its rollout schedule for other surfaces. Ours is Slack and Microsoft Teams, which is where Viktor works, so a team that runs mostly on email and spreadsheets gets less from him than a team that talks in channels.

Ask where the boundary is and what degrades when you cross it. A vendor who cannot answer that has not thought about your setup.

How does Viktor fit this comparison?

Viktor is the tool-graph version: an AI employee you @mention in Slack or Microsoft Teams, with real read and write access to 3,200+ integrations, who does the work and posts the result in the thread for review before anything is sent.

That means the answer to question one is honest and unflattering at first: on day one, Viktor knows nothing about your business. You connect Stripe, HubSpot, Google Ads, Linear or Notion, and from then on the context is whatever your tools hold rather than what one platform's data model can represent. Corrections stick, so the second version of a recurring report starts from the corrected format.

The answer to question two is where the difference shows up on a Tuesday. Viktor holds the credentials and does the reaching himself, so a cross-system request is a message rather than a ticket. He also has a persistent Linux sandbox, which is why the output can be a PDF, a spreadsheet, a deck or an internal app served at its own URL instead of text in a chat window. Review-first is the default, and Viktor maintains SOC 2 Type I.

Where the others win, plainly: if your company runs on Salesforce and has admins who already govern it, Agentforce Coworker starts with context Viktor has to be given. If your company runs on DingTalk, Alibaba's product will sit closer to your daily work than anything we can offer. If the work is engineering inside a repo, a coding harness beats a generalist.

For deeper reading on the mechanics behind these questions, see which integrations your AI employee actually needs, how to verify your AI employee's work, and the platform comparison in Viktor vs Salesforce Agentforce.

How should you actually run the bake-off?

Pick one real request that crosses two systems and is annoying enough that someone avoids it. Give the same request, word for word, to every candidate. Then compare four things: what it asked you for before starting, how long it took, whether the artifact was usable without rework, and whether the second run needed the same explanation as the first.

Skip the pilot that only touches the vendor's own data. Every product wins that one.

Frequently Asked Questions

What is an AI teammate?

An AI teammate is software that works inside the tools your team already uses, takes a request in plain language, and completes the task with real access rather than returning instructions. The distinction from a chatbot is access and output: it acts in your systems and hands back finished work.

Is Agentforce Coworker available right now?

Per Salesforce's 2026-08-04 announcement, it is in beta for Salesforce, with web, Microsoft Teams, ChatGPT, Claude and a desktop app described as coming later this year. Check their page for the current state before planning around a date.

Can I use QwenWork outside China?

Not in the beta announced on 2026-08-03. Alibaba's post says the public beta runs in China and that an international edition plus a standalone mobile app are planned.

Does an AI teammate need a data platform to be useful?

No, but it needs access to something. Either the vendor's platform already holds your data, or you connect the tools that do. What you cannot skip is access, because an agent without access can only advise.

How do I check that the work is right?

Keep a review step on anything that leaves the company, and check the first few runs against the source system yourself. Once a recurring report has been right repeatedly, thin the review rather than dropping it.

What should I ask a vendor about permissions?

Ask which accounts the agent uses, what it can write, who approved that access, and where the log of its actions lives. If the answer is vague, treat it as a security review rather than a product question.

Which one should a 40-person company pick?

Whichever one can finish your crossing-two-systems request without a build. If your business runs inside one platform, the platform's own teammate has a real head start. If it is spread across a dozen tools and a chat app, an AI employee in that chat app has less to explain.