Adobe is about to let a language model edit your lead records in the middle of a Smart Campaign. Callable agents, announced on October 15 and listed in the Marketo Engage November release notes for a rollout that starts November 6, run as a flow step. A lead reaches the step, the agent reads the record against instructions you wrote in plain English, and it writes an answer back to a field. A clean job title. A normalized company name. A flag saying this click was probably a security scanner and not a buyer.
That’s useful plumbing. It’s also the first time most Marketo shops will have a probabilistic judgment sitting inline with scoring and routing, firing on every qualifying lead, with nobody watching each call. Decide how you want that to work before the feature lands in your instance.
This is a docs-based review. I haven’t run callable agents on a live instance, and I haven’t found a customer who has published results. Everything below comes from Adobe’s own posts and documentation, plus two independent write-ups.
What it is
Callable agents belong to what Adobe now calls CX Enterprise Coworker for Marketo Engage, the agent layer formerly branded Marketo AI. Most of Coworker is conversational. You ask it to import a list or validate a program, it does the work, you approve the result. Callable agents are a different animal. They live inside a Smart Campaign as a step, next to Change Score and Change Data Value, and they run whenever a lead gets there.
Adobe’s October 15 post ships five templates:
- Job title standardization, so “VP MKTG” and “VP Marketing” stop scoring as two different people.
- Company name normalization, for routing, segmentation and reporting.
- Phone number formatting.
- Automated click activity flagging, to “identify clicks that may have been generated by security scanners before those clicks affect engagement-based decisions.”
- Suspicious lead record detection: junk names, disposable email addresses and other patterns typical of bad form fills.
You can also write your own. Adobe says teams with existing normalization logic “can deploy it as a custom callable agent with their own parameters and rules,” and that setup takes “just minutes.”
The most useful sentence about how this works sits in an earlier August 25 post: when you describe a custom agent, “Marketo AI builds the agent and configures the webhook, just as it does with the built-in templates.” So the agent gets called through Marketo’s webhook machinery. That tells you a lot about how it will behave under load, which I’ll come back to.
What changed
Until now, cleaning data inside Marketo meant one of three things. Smart list logic plus Change Data Value steps, a webhook to an enrichment or validation service you pay for, or a script someone on the team maintains. All three are deterministic. “VP MKTG” becomes “VP Marketing” because somebody wrote that exact rule.
Callable agents swap the rule for an instruction. You describe what a clean job title looks like, and a model decides lead by lead. Adobe’s data information sheet says Coworker reasons with Azure OpenAI GPT-4.1 and Claude on AWS Bedrock, and that Adobe doesn’t use your data to train those models.
The release status is messier than the launch posts make it sound. The August post described five templates as “available and ready to use in Marketo AI.” The October post says the feature is “now available for early access” and tells you to contact your account team. The November release notes, last updated October 21, file Callable Agents under “Coming soon,” describe agents that “act on a lead record based on instructions you write in plain language,” and say the release rolls out in phases from November 6, subject to change. Andrea Veggiani’s MarTech Watch for week 43 records the same November 6 date and notes that Marketo’s October release dropped off the schedule entirely. Read it this way: early access if you ask, wider rollout starting next week, exact timing for your instance unknown.
What works
The templates go after real, boring problems, which is exactly where an agent like this should start. Rich Ord at WebProNews, who runs a B2B email company, called them janitorial work and argued that “right now that may make them the most valuable agents in the entire marketing stack.” Hard to disagree. Title and company values feed every scoring model, segment and routing rule you own, and most teams are running on a mapping spreadsheet that went stale two reorgs ago.
Placement is the second strength. A lead who submits a form hits your trigger campaign immediately. If the title is junk, that lead gets scored and routed on junk before any nightly cleanup runs. An agent at the top of the flow fixes the value before the next step reads it. Adobe’s August post makes this argument directly, and it’s the best case for doing the work inside Marketo instead of in a weekly batch.
Plain language also lowers the bar for who can maintain it. An instruction like “if the email domain is a consumer provider and the company field is blank, set Lead Type to Consumer” is something an ops manager can write, read back and hand to a colleague. That beats a service nobody has opened since the contractor left.
Governance looks reasonable on paper. The data sheet says Coworker follows each user’s existing Marketo permissions, won’t surface fields or partitions that user can’t already see, and keeps outputs “reviewable through conversation and audit history.” Adobe’s community office hours from August 27 added that Marketo AI runs under user permissions and doesn’t consume your API quota, unlike the Marketo MCP server, which shares quota with your other integrations.
What breaks
Start with the webhook. Marketo’s webhook documentation is blunt. Use the Call Webhook step only in trigger campaigns. Marketo waits up to 30 seconds for a response before it gives up. Response mappings write back to the lead only when the service returns a 2xx code. And “even if a service takes only 50ms to execute, 100,000 executions take 1.5 hours.” A model call takes a lot longer than 50 milliseconds. Adobe hasn’t said whether its own agent endpoint gets special treatment, so assume the normal rules until someone tells you otherwise. In practice that means an agent at the top of a high-volume flow can slow everything queued behind it, and a timeout writes nothing while the lead carries on to the next step.
Second, nobody watches each call. The August post promises that “none of these agentic skills run unattended. Your team reviews and approves everything before it goes live inside Marketo.” That’s true of the chat skills it was describing. A flow step works the other way round. You approve the instructions once, then it runs on every lead that qualifies, at 2 a.m., with nobody looking. The data sheet’s audit history is framed around conversations. I couldn’t find documentation of a per-lead log showing what the agent saw, what it decided and why. Without that, you can’t explain a routing decision when sales asks where a lead went.
Third, the bot template overlaps with something Marketo already does. Native bot activity filtering matches opens and clicks against the IAB bot list and a proximity pattern (two or more activities from the same lead on the same email within a window you set, up to three seconds), then stamps each activity with a Bot Activity true or false attribute you can use in smart lists and triggers. That check is deterministic and lives on the activity. The agent writes its verdict to a field on the lead. Those are different shapes. A lead-level flag can’t tell you which of five clicks came from the scanner, and if you use it as a gate, one bad click can wipe out a real buyer’s later engagement. Run both and you need a rule for when they disagree.
Fourth, consistency. A rule maps “VP MKTG” the same way every time. A model probably will too, and probably is doing a lot of work in that sentence. If a field drifts between two values for the same input, your segments drift with it. Spell out the allowed output values in the instructions and check what actually comes back.
Ord also makes a fair case for cleaning upstream, at the form. Treat the agent as a catch net, not a reason to stop validating there.
Who it’s for
Marketo ops teams with messy title and company fields, a scoring model that leans on email clicks, and no appetite for another enrichment contract. I’d expect mid-market teams to feel it most, since they’re the ones normalizing from a spreadsheet.
If you already pay for a validation or enrichment service through a webhook, look at this as a possible replacement and test the two side by side on the same leads before you switch anything. If you work in healthcare, financial services, government or defense, the data sheet says the current rollout carries no vertical-specific capabilities for regulated industries, and the initial release excludes mainland China.
Pricing and limits
Adobe hasn’t published a price. The Coworker overview says the Coworker tile is available to all subscriptions once you accept the Core Gen-AI terms and the supplemental terms, and the October post routes callable agent access through your account team. Nothing I found says whether an agent run is metered. Adobe’s broader CX Enterprise Coworker trial counts chat usage in AI Credits, at an introductory 25 credits per input, so metering is clearly the pattern elsewhere in the stack. A flow step that fires on every new lead is a very different volume from a person typing prompts. Get the answer in writing before you hang one off a busy trigger.
The known limits, pulled together: trigger campaigns only and a 30 second ceiling per call if standard webhook rules apply, write-back only on a successful response, a phased rollout from November 6 that Adobe says is subject to change, and no regulated-industry tailoring yet.
What to do this week
- Ask your Adobe account team three questions in writing. Is callable agent usage metered, and at what rate? Do the standard webhook rules on trigger-only use and the 30 second timeout apply? Where can you see a per-lead log of inputs, outputs and reasoning?
- Get a bot baseline. If native bot activity filtering is off, switch it on in log mode and pull a week of flagged clicks. That’s how you’ll judge whether the agent template adds anything.
- Pick one field for the pilot. Job title is the obvious one. Write down the allowed output values first, then the instructions, and have the agent write to a new field so the original stays intact for comparison and rollback.
- Start on a quiet trigger. A single webinar form, not your main inbound flow. Compare campaign processing times before and after.
- Keep it out of scoring and routing until you’ve read a few hundred decisions. When you do wire it in, let a flag lower a score or send a lead to review, never delete or suppress it.
- Name an owner for each verdict. Decide whether the native filter or the agent has the final say on bots before both start writing to your database.
Adobe has the right idea here. Data problems should be fixed before scoring reads them, inside the flow where the damage happens. What’s missing is the documentation you’d need to put an agent on your main lead flow on day one. Spend this week collecting the answers and the baseline, so when your instance flips you’re running a pilot instead of a cleanup.
