AI marketing agent platforms compared on price, approval and open source

Three homes for the same promise

Every platform on a 2026 shortlist does the same visible work. It reads the CRM, the email tool and the ad accounts, then drafts a campaign, picks an audience, or writes the Monday report. Salesforce sells Campaign Agent as "your digital campaign manager, doing the work for you within guardrails you set" (Salesforce). Deskferry sells AI employees that work inside the apps a team already uses and check with you before anything important (Deskferry). Set the product pages side by side and the promise is nearly the same sentence.

What separates these eight platforms sits one level down. Where do the instructions live, and who can change them before a campaign leaves the building?

Inside the tool you already pay for. Campaign Agent runs in Marketing Cloud Next, on the audiences and brand guardrails a Salesforce account already holds. HubSpot's Agent Hub works on HubSpot records, and Klaviyo on its own profiles and flows (HubSpot, Klaviyo). This shape gives the shortest path to a first send, because the customer data is already there. The rule sheet, the approval step and the record of what went out live in the vendor's settings.

Across the stack. Deskferry, Braze and Iterable sell reach across channels and apps. Deskferry connects 1,500+ apps and holds anything that sends, spends or changes a record until someone says yes. Braze prices by the customers who actively interact with your brand and wraps Agent Console inside an engagement platform. Iterable's Nova agents build audiences, journeys and campaigns on Iterable's own data layer (Braze, Iterable). A team with four tools and no engineer gets working agents fastest in this shape.

In a repo you own. Kortix is the open-source AI Operating System: your agents, their skills, your company memory, and every connector in one git repo you own, with the agents working on real cloud computers. The marketing rules are one of those files: "Your voice, your claims and your banned words live in the repo as a skill file every session reads. Marketing work lands as a draft in a change request, so the review is a diff" (Kortix). Nothing merges itself: work reaches main through a change request a person approves, unless an admin granted the merge capability.

The agent's reach is a short block of the project manifest, not a settings screen:

agents:  kortix:    file: agents/kortix.md    connectors: all    secrets: all    kortix_permissions: all    skills: all    workspace: { project_repository: true }

Those lines are the whole grant: connectors, secrets, permissions, skills and the workspace checkout. A change to them arrives as a diff, and so does a change to the tone file or to the list of words the team never uses. That is why the voice skill and the claims skill are one file family. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free.

That shape asks something of the team. Someone has to read a diff and approve it. Someone has to own the rules file, or the repo fills with claims nobody checked. A team with no engineer and no appetite for review moves faster on Deskferry. The Deskferry comparison names where the two part.

Eight platforms on the same three columns

Every cell comes from the vendor's own page: Deskferry's plans, Blueshift's, Braze's and Klaviyo's. Prices were checked in October 2026 and they move.

The approval column is the one that separates a writing tool from a production system. Braze, Klaviyo and HubSpot do not say on their pricing or product pages who approves a send. Deskferry answers in one sentence a buyer can hold it to, and so does Salesforce. Blueshift's guide gives governance a dimension of its own: "Does the agent require human approval before execution, or can it act autonomously?"

How to run the comparison in an afternoon

Ask each vendor three questions, in writing, and keep the answer beside the price.

Is the code open source, and can we run it ourselves? A good answer names what you can read and where you can run it. A vague answer about open APIs describes connectors. Source is a different question.

Who approves a send, and what does that person see? A good answer names the role, shows the exact message and its arguments, and says where the decision is recorded. A weak answer says "human in the loop" and has no screen to show.

What is the first month's bill, and what does the meter count? A good answer gives a number, the unit, and what happens when nothing is running. On a credit plan, ask that last part twice. What a stopped agent costs is where hosted and self-hosted plans separate.

Put the three answers in one row per vendor. The vendor that answers all three in writing has earned the trial. The vendor that answers two by email and one with a demo has shown where its documentation stops.

The two lists a buyer meets first

The top result for the term is an article on Arahi AI titled "AI Agents for Marketing: 10 Platforms Tested (2026)". It opens by naming Deskferry "the strongest all-rounder", and it states the arrangement in a line of its own: "Disclosure: This article is published by Deskferry. We include our own product alongside competitors for transparency" (the article). The fifth result is Blueshift's "Best AI Marketing Agent Platforms (2026): A Buyer's Guide", a comparison of seven platforms. It puts Blueshift's Launchpad first and says why: "Full disclosure: this is our platform" (Blueshift's guide).

Both pages do real work. Deskferry's article knows its own product cold, and Blueshift's guide sets out six evaluation dimensions that are worth borrowing. Both also put their own product first, which is what a vendor writing a list will do.

That is the shape of the field. The vendor with a product in the field has the strongest reason to write the comparison. It also has the most to say about one row of it. Read those pages for what they know about their own product. Read the rival's price page for everything else, and date what you read. One of the eight prices is billed annually, and one moves with your customer count.

Where the open-source row changes the answer

The open-source row changes the shape of the purchase more than the feature list. With Kortix, the tone, the claims and the approval rules are files in a repo the team owns. A rule change has an author and a review. A new hire can read the rules before writing a line of copy. Self-host is free, which takes the platform fee out of the monthly bill and leaves the model and compute the team chooses. What the team gives up is the setup a hosted product does for it. The app connections, the model keys and the first review all land on the people who own the repo.

A team that cannot staff that review should buy the hosted shape and keep the vendor's approval queue in front of every send. A team that can should open its repo on a Monday and see what changed since Friday.