Dotdigital moved two AI agents to general availability on November 4, a little over three weeks before Black Friday. Segment agent writes the audience. Content agent writes the words. Read the help pages side by side and one design choice jumps out for anyone running a peak calendar on the platform: neither agent can see what the other made, and the one that builds your audience cannot tell you how many people are in it.
That is not a reason to leave them switched off. It is a reason to put the count, and the brief, back in a human’s hands before anything either agent touches goes anywhere near a send.
This is a docs-based review. I have not opened a Dotdigital account, typed a prompt into either agent, or watched a segment refresh. What follows comes from Dotdigital’s November 4 product blog, the 2026.4.3 release notes from the same day, the Help Centre articles for each agent and for the MCP server, and the company’s full year results filed on November 3. Independent coverage is thin. A Kalkine Media investor preview on November 2 noted the agents had launched after the financial year end, and that is about it. Everything below is Dotdigital describing Dotdigital, so I read it as a set of promises to test, not results.
What it is
Segment agent lives in Audience, then Segments, behind an AI icon that opens a side panel. You describe the conditions you want. It opens the normal segment builder, adds rules, names the segment, and waits. You review the rules, keep chatting or edit by hand, then press Generate segment. The release notes add two things for this release: it can now build a complex segment from one prompt, and it can open any segment, including ones built by colleagues, to explain the rules in plain language or change them.
Content agent sits inside EasyEditor and the SMS and MMS editors. It drafts, rewrites, shortens and personalises copy. In email it can also restyle text and buttons, insert contact data fields and write Liquid inside text blocks. In SMS it works with message segments and character limits and hunts down characters that push a message into Unicode. It can generate and edit images for email and MMS. Both agents pull from a brand voice profile you set up once.
The same release put an Analytics agent into private preview and promoted more MCP server tools out of beta. Dotdigital’s own pages disagree on SmartBlock agent: the blog calls it public preview, while the Get started with Dotdigital AI article lists it under private preview with a waitlist. Ask your success manager which one applies to you.
Scale matters for reading this launch. The results statement says the agents programme ran with around 250 customers, that 75 accounts used one or both agents between July and September, and that around 7% of the customer base was queued for full release. Most teams getting the button this month have never used it.
What changed
Until last week this was a preview for a few hundred accounts. Now the blog says Segment agent and Content agent are available to all Dotdigital Marketing customers, with Content agent covering email, SMS and MMS rather than just one editor.
The quieter change happened on October 21. That maintenance release added 55 MCP server tools, so an outside assistant such as Claude, ChatGPT or Copilot can now send and schedule email campaigns, create and send SMS, manage segments, enrol contacts in programs and send marketing and transactional email. Tools now follow each user’s permissions, and the sending tools only work when sending is turned on for the account. The connection article still labels the MCP server a public preview beta, even as the November notes say more tools have left beta.
What works
The segment agent works inside the builder, not around it. Its output is the same visible, editable rules a person would write, and when a prompt matches more than one list or segment it asks you to choose instead of guessing. That is the right shape for audience work. You can audit a rule. You cannot audit a score.
Summaries of inherited segments may be the most useful thing in the release. Every mature CRM account has segments named like a ransom note, built by someone who left. Being able to open one and ask what it actually selects, before you reuse it for a peak send, saves real hours and real mistakes.
On the content side, the SMS behaviour targets a real cost. A stray curly quote or emoji can switch a message to Unicode and change how it splits into billable segments, and Content agent is documented to find those characters and replace them. It also will not apply a subject line unless you ask it to, and generated images you do not apply are not saved to your library.
Dotdigital is also specific about its own models. The AI FAQ says generative features run on Microsoft Azure, that customer data is never used to train those models, and that prompts and responses are stored at rest in your Dotdigital region.
What breaks or is undocumented
No count. The Segment agent overview lists what it does not do, and two of the lines are about size: it does not show segment counts while building or editing, and it does not get the contact count or reporting for segments. The general segments article adds that a saved count is not live anyway; you refresh to recalculate it. So an agent can build or rewrite an audience and you can carry it into a schedule without ever seeing a number. That same article explains why that hurts. Exclusion rules run after include rules, so excluding a big segment can barely change the total. A prompt that misplaces one exclusion produces a segment that looks right in the rule list and is wrong by thousands. The count is the only quick tell.
Vague prompts. Marketing copy and help docs pull in different directions. The blog says that even if your request is vague, the agent suggests matching options from your data, and the agents landing page offers “lapsed buyers in the UK who visited the site in the last 90 days” as the kind of brief it will just get on with. The help article says the agent needs defined rules, cannot interpret broad ideas, and does not brainstorm segments. My reading: it will match names and fields when you are ambiguous about which list you mean, but it will not decide what lapsed means for your business. That is a definition, and someone on your team has to own it.
Edits to live audiences. The overview says the agent updates the segment directly from your prompts, and it can now open segments other people built. The segments article says a program refreshes its own copy of a segment as contacts reach each node. Put those together and an agent edit to a segment feeding a live program changes who flows through it next. I found no mention of segment version history or an undo in the articles I read.
Copy written blind. Content agent’s own list of limits says it cannot access audience data, segments or campaign performance, and it cannot optimise content using results. The blog’s line about telling it “who it’s for” is accurate in the narrowest sense: who it is for is whatever you type. It also does not fully validate Liquid or personalisation logic, so a fallback that breaks for half the list is still your problem.
A shared meter. Dotdigital asks for fair use at a maximum of 200 AI requests per account every 24 hours. Image generation runs on 100 credits per month, shared by every user for email and MMS, reset on the first of the month with no rollover. Black Friday is November 27 and Cyber Monday is November 30, so any credits burned on exploration this month are gone during the week that matters, and the refill lands on December 1.
Two retention sentences. Each agent article opens by saying agents store what you type in conversations to improve their performance over time. The FAQ says customer data is never used to train the models. Those can both be true, but your privacy lead should read them together before someone pastes a customer complaint into a chat panel.
The other door. The MCP tool list includes sending a campaign, sending batches of marketing email, deleting a campaign along with its reporting data, and emptying Insight data collections, which Dotdigital says cannot be undone. Dotdigital’s guidance is that many MCP clients ask you to approve a tool before it runs. So the agents page promise that nothing goes live until you have reviewed it describes the in-app agents. On the MCP side the review step belongs to whichever assistant you connected, and the FAQ is explicit that Dotdigital’s own AI safeguards do not apply to those providers.
Who it’s for
Mid-market CRM and lifecycle teams already on Dotdigital who own large, inherited segment libraries and want to understand them before peak. SMS-heavy retail programs that keep paying for Unicode surprises. Agencies running many client accounts, as long as they know each MCP connection reaches one account only and a parent connection does not reach its child accounts.
It is a weaker fit if you wanted an agent that proposes audiences, if your peak plan depends on knowing audience size at build time, or if you expected copy that learns from last week’s results. Analytics agent is the piece that reads results, it is still in private preview, and nothing in the docs connects it to Content agent.
Pricing and limits
Dotdigital publishes no separate price for the agents. The Get started article says availability of AI features depends on your account package. The MCP server is listed for the Accelerate, Amplify, Email and SMS Marketing, and CXDP packages. Documented limits: 200 AI requests per account per 24 hours, 100 shared image credits per month, up to 1,000 segments per account with up to 30 rules each, and engagement data counted back 10 years. Brand voice setup needs the account owner or the Can manage account permission. For context only, the results statement puts average monthly recurring revenue per Dotdigital Marketing customer at £1,943, priced by modules and contact and message volumes.
What to do this week
Write the count rule today. No agent-built or agent-edited segment goes into a schedule until someone refreshes it, writes the count next to the hand-built version it replaces, and explains any gap. Tick the option to refresh segments before sending on every peak campaign.
Freeze live audiences. From November 20 through December 1, nobody uses Segment agent on a segment that feeds an active program. Copy the segment, edit the copy, count it, then swap it in on purpose.
Write your definitions before you prompt. Lapsed, VIP, engaged, Black Friday buyer: each one as rules in a shared doc, pasted into the prompt word for word, so the agent matches fields instead of guessing meaning.
Build the bridge by hand. Ask Segment agent to summarise the final audience, then paste that summary, the offer and the exclusions into Content agent. It cannot see the segment, so give it the brief you would give a freelancer.
Ration the meter. Agree who gets the 200 daily requests during Cyber Week and hold image credits for real campaigns, since nothing comes back until December 1. Set brand voice first so the requests you spend come back closer to usable.
Check the MCP switches. Confirm whether sending is enabled for the account, list which users hold MCP server access, and get your privacy lead’s sign-off on each assistant provider before it reads contact data.
Then send your success manager three questions in writing: is there version history or undo after an agent edits a segment, does the 200 request cap count each chat turn, and when will counts appear inside the agent.
