← Back to blog
NewsAbout 6 min read

Adobe Stopped Asking Marketers to Open Adobe

Published Nov 7, 2026
Adobe Stopped Asking Marketers to Open Adobe

On November 4, Adobe announced that its CX Enterprise Coworker can now be reached from inside ChatGPT as a plugin, and also from Claude, Databricks Genie, Gemini for business and Microsoft Copilot. Amazon Quick, IBM and Snowflake are listed as coming soon.

The pricing was not disclosed. The distribution change was.

For a decade, enterprise software competed on becoming the place where work happens. Own the interface, own the workflow, own the data. Adobe built one of the most successful versions of that strategy: a creative and marketing suite so embedded that leaving it meant retraining a department.

The new announcement inverts it. If your marketing team already chats with an AI all day, Adobe would rather appear in that window than make them open another one.

What the Coworker does, and what it sits on

Adobe describes CX Enterprise Coworker as an agent that runs with human oversight and is built on open standards, specifically MCP and Agent2Agent. New connectors cover Amazon Ads, Demandbase and Snowflake Cortex AI.

The technical choice is the interesting part. MCP is the same protocol Anthropic published for tool access, and Agent2Agent is the inter-agent protocol Google introduced. Adobe is building a plug that fits sockets other companies defined, rather than a proprietary doorway.

That makes the agent portable, and portable agents are a different commercial object from a suite. If the customer can reach your agent from their chat tool of choice, your advantage comes from the quality of the connector and the data behind it, not from the fact that your interface is the only one that exists.

The open question, which Adobe did not answer, is where permissions are enforced. When a marketing agent operative is callable from five platforms, the approval rules have to live somewhere that all five respect. That is the same problem CrowdStrike and AWS spent the same week attacking from the security side.

Agents landing in systems of record

Smartsheet took the opposite approach at its ENGAGE conference on November 4, and it is the more conservative bet. Instead of exporting the agent to other chat surfaces, it put agents inside the project tool.

The announced pieces: agentic workflows that turn a plain-language request into a recurring task, a Creation Agent that builds projects complete with dashboards and automations, a Data Quality Agent that flags issues by severity, and an Integration Agent in early access. Alongside them, near-real-time security event streaming to Splunk and Microsoft Sentinel, plus data classification labels.

That last addition is the one worth noticing. An agent that can create projects and edit data is a privileged actor inside a system of record, and the security team needs to see what it did. Shipping the streaming integration in the same release as the agents is the correct order of operations, and a lot of vendors do it the other way around.

The clearest business case came from a travel platform

On November 3, Klook launched K.ai, an agent built for the small merchants behind its experiences listings.

K.ai does back-office work rather than answering questions. It updates inventory, pricing and bookings directly inside Klook's merchant portal, and generates structured listings from whatever material a merchant already has, which for many operators means a phone full of photos and a WhatsApp history.

Klook's reported numbers: individual task updates complete 95% faster, batch checks 97% faster, and merchants using K.ai log 3.4 times the monthly bookings of non-users. The agent is rolling out globally ahead of the year-end travel season.

Treat the 3.4x carefully. Merchants who adopt a new tool are not a random sample; they are the ones already trying to grow. The defensible claim is the 95% time saving on a specific task, because that is measurable and hard to fake. The bookings number is a correlation with a plausible mechanism behind it.

Still, the mechanism matters. A tour operator in Da Nang with twelve listings and no back-office staff does not need an AI strategist. They need someone to fix the pricing on eleven products at once. That is a task with a start, an end and a correctness check, which is exactly the shape agent products work best on.

Systems integrators become the channel

One more signal from the same few days: Glean partnered with IT services firm Emergys to move joint customers from agent pilots into production, pairing Glean's agent platform with Emergys' cross-platform agents to unify search and workflows across SAP, BMC, ServiceNow, Freshworks and Salesforce.

That partnership exists because most large organisations cannot connect agents to fragmented legacy systems themselves. The integration layer is the bottleneck, and consultancies are becoming the default distribution path for agent platforms, the same way they were for ERP.

Put the three moves together and a pattern appears. Adobe goes wide, exporting its agent to whatever surface the customer already uses. Smartsheet goes deep, embedding agents where the work and the audit trail already live. Klook goes narrow, solving one merchant workflow completely and measuring it.

All three are responses to the same realisation. The agent itself is becoming a commodity. What is scarce is the connection to a specific workflow with a specific data model, and whoever owns that connection does not need to own the chat window.

The standard question

All three announcements point at a quieter shift: agents are being written against shared protocols rather than vendor SDKs.

Adobe's agent speaks MCP and Agent2Agent. That means a customer's own orchestration layer can call it, and it means a competitor's agent could sit in the same conversation. Standards lower the cost of integration, and they also lower the cost of replacement.

For buyers, that is bargaining power. A marketing agent that can be reached from five chat platforms and speaks two open protocols is a marketing agent that can be swapped. For vendors, it is a bet that the workflow data and the connectors are the durable asset, and that the interface was never the moat.

The counter-argument is that standards fragment as they scale. MCP and Agent2Agent overlap, security models differ, and the version people deploy in two years may not be the one shipping today. Anyone writing procurement criteria now should ask which protocol revision the vendor supports, and what happens when the two protocols disagree about permissions.

What to ask vendors now

Three questions separate a real deployment from a press release.

Where are permissions enforced when the agent is called from a third-party surface? If the answer is "inside our product," then the agent is only as secure as the weakest platform it appears on.

What happens to the audit trail when the call originates in someone else's chat window? Security teams will ask this, and "we log our side" is not a complete answer.

And what does the agent do when the task fails halfway? A booking agent that updates eleven prices and fails on the twelfth has left the catalogue inconsistent. Rollback semantics are boring, and they are the difference between a tool a merchant trusts and one they check manually afterwards.

Adobe's move is the visible one. The unglamorous version of the same strategy is being built in project tools and merchant portals, one workflow at a time, with the logging shipped before the demo.

Related articles