Featuring:

Adobe is making its tools callable inside AI assistants and assigning work to agents. When a workflow crosses products and companies, accountability can disappear with the interface.

Who owns the failure?

When a user asks ChatGPT to create a campaign package, Adobe tools perform the edits, a Workfront agent advances the project, and Marketo executes a campaign operation, who takes responsibility when the outcome is wrong or incomplete?

During nine years at Adobe, I led support work across multiple products like Marketo - making up Adobe’s agentic customer-experience offering.

One pattern repeated: difficult customer problems rarely respected product boundaries.

A campaign problem might begin in one product, turn into a data or identity question, cross an integration boundary, and surface somewhere else entirely. Inside the company, each handoff could make perfect organizational sense. To the customer, it could feel as if nobody owned the outcome.

AI orchestration makes that problem harder because the user may not know which product was selected—or even which company’s technology might perform a particular step.

Ownership should follow orchestration. If a system accepts an end-to-end request, it should own the first response when the workflow fails. It should preserve the trace, identify the unsuccessful step, and route the issue without forcing the customer to reconstruct the underlying product map.

The organization using the workflow also needs a named human owner. That person doesn’t need to perform every step, but someone must remain accountable for the design, approval points, exceptions, and final result.

On August 6, Adobe replaced the separate Photoshop, Express, and Acrobat experiences it had launched in ChatGPT in December with one unified @Adobe plugin. Incredibly exciting!

That consolidation captures the strategy in miniature. Adobe has stopped asking users to know which product owns which function.

The new plugin can select from more than 70 tools across Photoshop, Lightroom, Firefly, Premiere, Acrobat, Express, Illustrator, InDesign, Adobe Stock, and elsewhere. A user describes the outcome; Adobe and ChatGPT determine which supported capabilities to invoke.

The entry point is unusually accessible. People can begin in guest mode without a Creative Cloud subscription. Signing in with a free or paid Adobe account unlocks additional tools, generative capabilities, Creative Cloud storage, and continuity between sessions.

On August 13, Adobe made Workfront AI Collaborators generally available, allowing organizations to assign certain kinds of work to an agent much as they would assign a task to a person.

Adobe is also expanding access to a Marketo Engage MCP server with more than 100 operations across programs, emails, forms, leads, smart campaigns, lists, folders, and other marketing work.

These are not simply three product announcements. They point towards software where the conversational platform accepts the request, Adobe supplies callable capabilities, an orchestration system coordinates the work, and people intervene at defined decision points.

Adobe is distributing different capabilities across several AI environments (note that this chart changes often!)

“Integration” does not mean the same thing in every row.

A creative plugin, an MCP server, a Microsoft 365 agent, and an AI collaborator assigned through Workfront have different capabilities, permissions, context, and approval mechanisms. Treating them as interchangeable because they all begin with a conversation obscures both control and ownership.

Adobe’s Acrobat experience in Microsoft Copilot offers an important complication to the “disappearing interface” story.

The request begins in Copilot. A user describes the PDF task - edit this document, combine these files, extract information, or compress the result. Acrobat processes the request and opens the document in an interactive Acrobat editor. Quick changes can be made in the preview, while more precise work moves into the full editor. The person reviews and confirms the changes before saving.

The conversation accepts the request. The application handles precision. The person retains approval.

Adobe is not always eliminating the interface; it is deferring it until the interface becomes useful. Users do not need to navigate the application merely to begin, but Adobe restores direct control when the work becomes consequential or exact.

That may be the stronger design principle: hide product selection and routine mechanics, then surface the appropriate interface, provenance, and approval controls at the moment they matter.

It also creates a cleaner allocation of responsibility. The conversational layer owns interpretation and routing, the application owns the operation it performs, and the person owns final approval.

Imagine an entrepreneur preparing to launch a physical product with two employees and no dedicated creative-production team.

The founder has product photographs, a short demonstration video, brand guidance, and a spreadsheet containing approved features, prices, and URLs. The request might be:

The workflow has three stages.

First, validate the inputs. The system identifies missing information, conflicting prices, low-resolution images, and unsupported claims. A person confirms what the product is, what the business is promising, and which source is authoritative.

Second, produce the drafts. Adobe capabilities normalize the photographs, develop coordinated social designs, reformat the video, and populate a product sheet from structured data. A person chooses the creative direction and inspects the product, packaging, labels, colors, faces, hands, captions, and brand details.

Third, verify and hand off. Every price, URL, claim, disclaimer, caption, and image association is checked before export. Structured automation can reproduce one small mistake with impressive efficiency. Anything requiring precise retouching, complex layouts, layers, masks, or production finishing moves to the appropriate professional in the native Adobe application.

The result is not “AI created our campaign.” A small team removed repetitive production work while retaining control of the decisions.

Adobe’s creative plugin illustrates how capabilities can be invoked without choosing applications. Workfront shows how agents can participate in organized, visible work.

Current Adobe documentation describes two available AI Collaborator types. A Reviewer evaluates assets against configured brand guidance. A Task Collaborator connects an agent created in Claude, Microsoft Copilot Studio, or Writer to a Workfront task.

A Task Collaborator can receive the task title, description, comments, and attached custom-form information. When the task becomes ready-including after required predecessor work is complete, and the agent can perform its assigned function and return the output to Workfront.

The result stays attached to the work instead of disappearing into somebody’s private chat history. That is more than an administrative convenience: it gives the outcome, the evidence, and the first response to failure a home.

Workfront MCP operates in the other direction, allowing people to find, create, update, and manage permitted Workfront items from an external assistant. Read tools are enabled by default; write tools require administrative enablement and operate through the user’s existing permissions.

Marketo adds an execution layer for marketing operations. Its Adobe-hosted MCP server exposes more than 100 operations, while Adobe says destructive actions are blocked. Smart List and Smart Campaign creation and updating are currently targeted for September.

The product’s public availability status is not perfectly synchronized. Adobe’s August 3 blog describes open beta, Niranjan Kumbi announced general availability on LinkedIn this week, and the current Experience League documentation does not label it either way. The safest conclusion is that Marketo MCP is transitioning into broader availability, but individual customers should verify access and enabled capabilities before designing around it.

The implementation details reinforce that point. Adobe documents Claude Desktop, Claude Code, Cursor, Codex, VS Code with GitHub Copilot, and Glean as Marketo MCP clients, but not every web version of those AI platforms. Marketo instances must also be enabled for MCP access, and authentication depends on configured credentials and headers.

“Generally available” does not necessarily mean “enabled for you on Monday.”

The building blocks point toward a campaign workflow in which:

This is not yet a single push-button workflow. The important development is the operating model: assign work, supply context, return the result, preserve the record, and require a person at consequential decision points.

Most enterprise support organizations rely on an important assumption: the customer can identify which product has failed. Hiding this creates significant tension.

That assumption determines the intake form, entitlement check, tier-one queue, specialist routing, telemetry, escalation path, and service-level target. But a customer using an agentic workflow may know only that “the campaign package did not complete.” They may not know which Adobe tool was invoked, whether the failure began in ChatGPT, or whether a permission, policy, model, connector, or source file caused it.

Product-aligned queues cannot resolve workflow-level failures unless somebody can reconstruct the chain.

Support will therefore need more workflow-level intake systems, traceability across tool calls and handoffs, and escalation agreements that cross product, and often company, boundaries. The first team receiving the problem should be able to see what the orchestrator attempted rather than asking the customer to reproduce a workflow they never directly operated. Everyone in support will tell you just how difficult this can be to realistically achieve.

Otherwise, AI will remove friction from completing the work only to recreate it when the work fails. Please be patient if you need to call on support 🙂

There is a second tension in Adobe’s strategy.

Adobe is making the production process less visible at exactly the moment parts of the creative economy are demanding greater visibility into how work was made.

Independent publishing provides a useful example. In author-organized promotions hosted through BookFunnel, I regularly see participation rules stating that books containing generative-AI writing, artwork, or even AI-generated covers will be removed from the promotion.

This is not a BookFunnel-wide rule. Individual organizers establish the requirements. Small, independent entrepreneurs are creating standards for how creative work should be produced and represented, often partly to support human writers, illustrators, cover designers, editors, photographers, and narrators.

Professional designers and photographers have raised parallel concerns about training data, creative employment, client expectations, loss of control, inconsistent output, and the assumption that thoughtful work can be reduced to a prompt.

It also matters that “using Adobe through AI” is not synonymous with “generating an image.” Correcting exposure, resizing a video, extracting a PDF table, applying an approved template, and inventing a cover illustration are materially different activities.

Adobe’s trust opportunity is to make those differences visible. People should be able to determine which steps used conventional automation, which introduced generative content, what source material was used, and where a person reviewed or changed the result.

Adobe’s Firefly is trained on licensed content, including Adobe Stock, and public-domain material rather than customer Creative Cloud files. Content Credentials can record information about how an asset was created. Those credentials, and comparable workflow records, must remain attached as work moves through external assistants, Adobe tools, project systems, and final exports.

Human review does not convert prohibited generative work into acceptable work. Review is a quality and accountability mechanism; it is not retroactive permission. Communities, customers, and creative professionals must still be able to decide whether generative technology belongs in a particular workflow at all.

That decision, and the evidence supporting it, must be designed into the workflow rather than reconstructed afterward.

You do not need an enterprise agent architecture to apply the lesson.

Choose one bounded workflow. Start with a repetitive, reviewable process such as resizing approved campaign assets, preparing social variants, checking a campaign brief, or generating a PDF from verified data.

Draw the human boundaries first. Identify which source is authoritative, which decisions require approval, and which actions the system must never take.

Record how each step was performed. Distinguish conventional automation from generative creation, retain the source material, and preserve relevant Content Credentials or activity records.

Name the outcome owner. One person should own the workflow even when five tools and two vendors participate in it. Define where failures are recorded and who investigates them.

Pilot with the least authority necessary. Start in guest mode, a sandbox, or with read-only connections. Add write access only after the workflow is repeatable and its failure modes are understood.

Adobe’s biggest opportunity is not to eliminate its applications or the professionals who use them. It is to become the trusted execution layer that people and agents call when work must be editable, traceable, governed, and production-ready.

Hide the interface if it makes the work easier.

Never hide who is responsible for the result.

Sources: Adobe for ChatGPT, Adobe’s December 2025 ChatGPT launch, Adobe for creativity in Claude, Marketo MCP documentation, Niranjan Kumbi’s Marketo status announcement, Workfront AI Collaborators, Workfront Task Collaborators, Acrobat for Microsoft Copilot, editing PDFs through Acrobat for Copilot, and Adobe Marketing Agent for Microsoft 365 Copilot.

Affiliate disclosure: As an Amazon Associate I earn from qualifying purchases.

Meta AI can now connect to your Facebook and Instagram analytics and Meta ad campaigns. It can analyze performance, audit campaigns and create recurring reports. The capabilities are available through Meta AI on the web, mobile and its new Mac app. Meta announced the features this week.

That creates a useful workflow for a creator or small business—not because you need AI to produce even more content, but because you probably do not have time to study what happened to the content you already published.

Try this prompt:

That last sentence matters. First decide whether its analysis is sound. Check that the recommendations reflect your actual goals—not simply reach—and look for conclusions based on one unusually successful post.

Once you trust the report, ask Meta AI to turn it into a Monday routine:

One caution: Meta AI can also connect to Google Workspace. According to Axios’s reporting, information shared from connected business accounts may be used for future AI training and ad targeting under Meta’s privacy policy. I would begin with Facebook and Instagram analytics and review the data terms before connecting email, documents or calendars.

The workflow is not “let AI run my account.” It is: shorten the distance between publishing, learning and choosing the next experiment.

That is a small workflow worth trying on Monday.change.

📝 75% of this newsletter is written personally by our team.

🤖 The remaining 25% is supported by AI—for things like topic research and grammar checks.

🦾 Each piece of content takes a lot of effort from our team, so your feedback means a lot to us. Let us know what you think!

To make sure you never miss an update, please move this email from the Promotions or Spam tab to your Primary Inbox. This helps ensure you receive the latest AI news, tips, and tutorials—right where you’ll see them first.

Reply to this email with any specific feedback or interesting insights!

Thanks for reading!