The word agent has stopped carrying information. On a pricing page it now covers a chat box that writes subject lines and a system that routes live customers down different journey paths without asking anyone. Both are sold as agents, and only one of them can lose you a customer before anybody on your team notices.
Behaviour is the only test a buyer can run. How many model calls are chained, and whether there is a planner in the loop, is a question for the vendor’s engineers. Two questions decide the purchase: can this thing change something without me, and can I take that permission away.
Table of Contents
What separates an agent from an assistant
An assistant answers a prompt. You ask, it produces something, you decide what happens next, and the world is unchanged when you close the tab.
An agent pursues a goal across several steps and changes something at the end of them. The change is the load-bearing part. Generating a campaign is not agentic; sending it is. The same line runs between drafting a CRM note and writing it to the record. A system that plans a dozen steps and then shows you a preview is an assistant with a long attention span.
Most teams should buy the assistant first, because its worst day produces a draft you delete instead of a badly segmented send to your whole list.
| Question | Assistant | Agent |
|---|---|---|
| What starts the work | You, one prompt at a time | A trigger you configured once, firing without you |
| What it produces | A draft inside the tool | A change to a live surface |
| Who presses send | You | The system, inside limits you set |
| What you can revoke | Nothing, there is nothing running | Permissions, per surface, ideally without disabling the feature |
| Typical metering | Seats, words, messages | Runs, actions, or completed outcomes |
| Failure looks like | A bad draft you notice and delete | Something happened, or silently did not, and you need a log |
Run that against the last demo you sat through. If every row lands in the left column, you were shown an assistant, whatever the slide said.
Four questions to ask about any agent on a pricing page
1. What exactly can it change
Ask which surfaces it is allowed to write to. A capability list will not answer that, because a draft, a live web page, a customer’s inbox, a CRM field and an ad budget are five different risk levels and vendors describe all five with the same verb.
Braze draws the sharpest boundary here. Its Agent Console agents generate copy, route users down Canvas paths and enrich catalog fields, and the docs state that agents fire only from a Canvas step or a catalog field, and do not autonomously initiate actions outside those contexts. That tells you the blast radius before you buy.
Klaviyo sits at the other end. It calls Composer an AI agent offering strategies and assistance for marketing specialists, then says content is prepared for your review before reaching customers and that you are always in control, free to approve, modify or reject the draft. Composer does assemble the pieces (segments, campaign content, channel recommendations, timing), and a human ships every one of them.
2. Who approves, and can approval be revoked per surface
Three answers are possible: approval is mandatory and enforced, configurable per action, or missing from the documentation.
Salesforce puts the gate in writing for Campaign Agent: marketers review, refine and approve content and guardrails before it launches and sends. Iterable’s Nova Agent is described as handling execution within guardrails marketers define, but no public Iterable page says whether a generated campaign deploys automatically or waits for a human, and Copy.ai and Tofu are equally silent about approval on their own product pages. Get that answer in writing.
Then ask the question people skip: revocation granularity. If a content agent starts publishing thin pages, can you revoke publishing and keep drafting, or is your only lever switching it off? A single global toggle makes every incident an outage.
3. What is the metering unit, and what does one unit buy
Four units are in use across this category: seats, credits per action, credits per completed outcome, and credits per run. Which one a vendor chose tells you what it is prepared to stand behind. Then get the second half of the question answered, because an allowance is not a rate. Klaviyo’s 10,000 Composer credits and Copy.ai’s monthly workflow credits both arrive with no published per-action cost, so neither one converts into a budget you can defend. Where the rate is published you can price a job before you run it: Clay charges one action per AI prompt for a standard Claygent and one action per record for an Account Agent, so the size of the list you point it at decides the bill, and you can work that out in advance.
4. What happens when it fails
Agents fail three ways: silently doing nothing, doing the wrong thing at volume, or half-finishing and leaving inconsistent state. Ask which of the three the vendor has instrumented, and who pays for a failed run.
Braze publishes an answer. Rate-limit errors, model unavailability and daily-limit failures do not consume credits; timeouts do, because a timeout means work happened. Braze also publishes a default ceiling of 250,000 agent runs per day, a maximum of 1,000,000, and a per-agent daily credit limit set by the model’s per-invocation credit ratio. Ceilings decide whether a misconfigured agent costs you a morning or a quarter.
Metering is the tell
To learn what a vendor privately believes about its own agents, look at the unit it charges for. Per seat promises nothing about the work. Per completed outcome bets the vendor’s revenue on the work being good enough that you come back and ask for it again.
HubSpot made that bet most explicitly, moving its agents to outcome-based credits on 14 April 2026: 50 credits per resolved conversation for Customer Agent, about $0.50, 100 credits (about $1.00) per lead Prospecting Agent recommends, and 10 credits per smart properties run for Data Agent.
| Vendor | Unit | What one unit buys | Published rate |
|---|---|---|---|
| HubSpot | Credits, per outcome | 50 per resolved conversation, 100 per recommended lead, 10 per smart properties run | About $0.50 and $1.00 for the first two; top-ups $9.00 per 1,000 credits |
| Salesforce Agentforce | Flex Credits, per action | 20 per action, 30 per voice action | $500 per 100,000 Flex Credits, or a per-conversation model at $2 each |
| Braze | Message or Action Credits, per run | Varies with the model’s per-invocation credit ratio | Not published |
| Clay | Actions plus data credits | One action per AI prompt for a standard Claygent, one per record for Account Agents | Free tier: 500 actions, 100 data credits |
| Jasper | Credits, per action | One action is, for example, a single GEO Hub report run | Rate card not public; seats $69 monthly, $59 billed annually |
| Copy.ai | Workflow credits | Varies with workflow complexity; chat words uncredited | Allowances published per tier, per-workflow cost is not |
| Klaviyo | Composer credits | Not published | Free plan and new-customer trial include 10,000 credits; the trial expires at 90 days or when credits run out |
Read that last column for its billing basis as well as its number. Copy.ai’s Chat tier is $29 monthly or $24 billed annually, while Growth, Expansion and Scale are quoted monthly but billed annually at $12,000, $24,000 and $36,000, so the monthly figures on those three describe an annual contract.
Two price sets are odd enough to flag. HubSpot Marketing Starter lists $7 per seat per month billed monthly and $20 per seat on an annual commitment, so the annual figure is the higher one; do not read that as a discount either way. Clay’s Launch tier lists $167 monthly alongside $54 billed annually, a gap that does not reconcile with the saving claimed on the same page. Braze, Iterable, Adobe and Tofu publish no agent pricing at all.
Where the category actually is today
Several of the flagship agents here are gated, partly built, or not shipped. Salesforce Campaign Agent, which selects audiences, generates variations, picks channel and send time and suppresses competing sends, reaches general availability in October 2026 in Marketing Cloud Next Advanced, and Salesforce’s own wording adds “with some limitations”. Additional agentic campaigns cost extra, at a price it does not publish.
Adobe’s Journey Agent builds a journey from a text prompt, then stays on the recommending side, and its own verbs say so: it flags audience overlap, suggests timing adjustments, highlights underperforming journeys, and will soon offer proactive recommendations to fix errors. Adobe says human oversight remains a central cornerstone of its AI tools, and its Experience League documentation adds that the agent’s plans are shown to you so you can follow the process and intervene. Its separate Brand Visibility product is the exception, intercepting AI agent requests at the CDN layer and serving optimized content in minutes, with instant rollback.
Copy.ai does not state on its homepage, pricing page or GTM AI page whether workflows write to a CRM or send messages, so whether anything leaves the platform is unverifiable from outside. Tofu’s Launch Agent claims it sequences emails, stages leads, fires campaigns and updates CRM records in HubSpot, Marketo, Salesforce and Outreach, which would put it among the most autonomous products here, and there is no pricing page and no approval documentation behind the claim.
Two well-known agents deliberately do less than the label implies. Clay’s Claygent researches and enriches, writing structured output, and does not send outbound. Jasper’s agents run multi-step workflows while humans set intent and approve outputs, preparing assets for activation. Both are regularly described in listicles as doing things they do not do.
The gap: nobody is measuring whether it worked
Assume the optimistic version of all of this: agents draft your pages, personalize your journeys and ship at a volume a human team could not match. Four weeks later a question arrives that none of these tools can answer. Did any of it make the brand more visible when a buyer asks ChatGPT, Perplexity or Google AI Mode for a recommendation?
Campaign platforms measure sends, opens, clicks and pipeline. They do not watch AI answers. As discovery moves into those answers, and as agentic commerce routes purchase decisions through models instead of result pages, content produced at agent speed with no visibility feedback loop is content produced blind.
LLM Pulse sits in that slot. It tracks how often a brand appears in AI answers across ChatGPT, Perplexity, Gemini, Google AI Mode and Google AI Overviews, which sources those answers cite, share of voice against named competitors, and sentiment. Two of its surfaces sit on opposite sides of the line drawn above. The LLM Pulse Agent acts inside the platform when asked, answering questions about your own data and then creating prompts in bulk, adding competitors, tagging or launching a Recommendations run, confirming first on anything that spends quota. GEO Writer is the assistant of the two: it produces briefs and articles grounded in your own prompt and citation data, and a human decides what gets published.
Where it stops matters as much. LLM Pulse runs no campaigns: no email, push or SMS, no ad campaigns, nothing written to a CRM, no CMS write-back, and no auto-publish anywhere. If you want something that executes marketing end to end, this is not it. If you want the measurement layer the executing agents lack, plans start at 49 euros a month. Either way, collect the baseline before you point an agent at your content, because once production has scaled you can no longer tell which part of the movement the agent caused.
FAQ
What is agentic marketing, in one sentence?
Marketing work carried out by software that pursues a goal across multiple steps and changes something real at the end: sending a message, routing a customer, updating a record, publishing a page, inside limits a human set beforehand.
Is it just marketing automation with new branding?
Partly. Classic automation follows rules you wrote; an agent decides how to reach a goal you stated, so it can produce actions you never explicitly authorized. You see the difference when something goes wrong.
Which marketing agents actually change things today?
By each vendor’s own documentation: Braze’s Agent Console agents route real users and write real field values, HubSpot’s Customer Agent resolves live conversations while Nurture and Revenue Agents send, and Adobe Brand Visibility deploys site changes at the CDN edge. Tofu claims execution into your live stack with nothing published to corroborate it, and Salesforce Campaign Agent is not generally available until October 2026.
How much does it cost?
The units differ too much for one answer. HubSpot publishes outcome pricing, roughly $0.50 per resolved conversation and $1.00 per recommended lead, and charges for the platform separately: Marketing Starter is $7 per seat per month billed monthly, or $20 per seat on an annual commitment, while Enterprise starts at $3,600 a month for five seats plus a one-time $7,000 onboarding fee. Salesforce publishes $500 per 100,000 Flex Credits at 20 credits per action. Jasper’s Pro seat is $69 monthly or $59 billed annually.
Do I have to approve everything an agent does?
No, and at volume you would not want to, since approving every send removes the reason you bought it. The arrangement that works: approval on new surfaces and new audiences, automatic execution inside proven ones, a hard ceiling on runs per day, and revocation you can apply to one surface without disabling the agent.
How do I tell whether agent-generated content helped?
Measure the outcome the content was for, which increasingly means presence in AI answers rather than rankings alone: how often your brand appears in answers to the questions your buyers ask, which sources those answers cite, and how your share moves against competitors.
