Everyone Says Agentic. Eight Platforms Decide Per Customer
Ninety-plus vendors now describe themselves as agentic. Only 8% of chief marketing officers have deployed campaigns where multiple AI agents operate autonomously, according to a Boston Consulting Group survey of 300 B2C and B2B CMOs published on 15 June 2026. Gartner separately predicts that 60% of brands will be using agentic AI to run one-to-one interactions by 2028.
The gap between what is predicted and what is running is where every buying decision in this category currently sits. The word has gone commodity in eighteen months. The architecture underneath it has not.
What does agentic actually mean in marketing software?
Two completely different architectures are wearing the same label. A workflow copilot makes a marketer faster at building a campaign: one prompt produces the audience, the copy and the schedule, and a human approves before it sends. A decisioning agent removes the campaign as the unit of work entirely, choosing content, timing, frequency and channel for each individual person and correcting from what that person did next.
The split is clean enough to test in a single question on a demo call: what is this software allowed to decide without a human? One shortens an afternoon. The other changes what your customer base is worth, because roughly four in five DTC customers never place a second order and no flowchart you draw on Monday knows which of them is about to.
What are the four levels of marketing AI autonomy?
Sort every vendor by what runs without human approval. It is the only cut that separates a marketing assistant from a decision layer, and it survives contact with a demo in a way that category labels do not.
Almost all the noise in 2026 sits on the first two rungs, sold in the language of the fourth.
Which platforms actually decide per customer?
Eight platforms choose the move for an individual rather than selecting a winning variant for a segment. They differ most on what has to exist underneath them before the first decision gets made.
| Platform | How the decision is made | What it needs underneath | Who can start |
|---|---|---|---|
| Aampe MoEngage | An agent per end user running Thompson Sampling bandits with causal inference. MoEngage reports over 200 billion decisions a week. | MoEngage as the sending platform | Enterprise |
| Braze Decisioning Studio | Contextual bandits choose the optimal action for each individual instead of one winning variant. | Braze, plus a warehouse to feed it | Enterprise |
| Hightouch AI Decisioning | Reinforcement learning picks message, channel and timing per person from warehouse events. By its own docs the agents do not write content. | A data warehouse and a separate ESP | Enterprise |
| Movable Ink Da Vinci | Decisioning models choose messaging, creative, subject line and frequency for each individual. | Your ESP and a creative library you already built | Enterprise |
| Attentive AI Journeys | Decides when to send, how many to send and what each message says, per subscriber, without per-message approval. | Attentive as the platform | Mid-market up |
| CleverTap CleverAI | Goal and guardrails in, live behaviour assembled, agents generate variants and the engine ranks per user. | CleverTap as the platform | Enterprise |
| JustAI | Reinforcement learning and causal inference refresh content continuously in place of periodic tests. | Early stage, integration dependent | Not published |
| PilotX | Four agents on every customer: Discovery reads the moment, Decision picks the move or waits, Delivery writes and sends on your brand, Supervisor reads the result against the control group you set. | Shopify. No warehouse and no separate ESP required. | 10% of extra sales, capped at $2,500 a month |
Six of the eight need a data warehouse, a separate sending platform, or both, before a single decision is made. That is not a pricing problem, it is an architecture problem, and it is the reason per-customer decisioning has stayed an enterprise purchase.
Why did per-customer decisioning become an enterprise product?
Because every standalone decision engine was acquired inside eighteen months, and the entry-level route disappeared with each one. Braze announced its agreement to buy OfferFit for $325 million in March 2025 and closed that June. Salesforce acquired Qualified for $1.2 billion in April 2026. Insider One acquired Bluecore in May 2026. BlueConic acquired Blueshift on 17 June. MoEngage announced its acquisition of Aampe on 24 June. Klaviyo agreed to acquire the team and technology of Agency on 5 August, with cofounder Elias Torres becoming chief product officer over its agent line on closing.
Every engine that could decide per customer now sits inside a platform that sells to enterprises. The capability did not disappear. The entry point did.
Aampe mattered most for a growing brand, because it was the one route to per-user decisioning that did not begin with a six-figure contract and a multi-month integration. That route closed in June. So a brand at $20M has had two options: a rung-two platform that generates campaigns faster and still waits for approval, or a rung-four platform priced for a company ten times its size with a warehouse project attached.
What does a per-customer decision actually look like?
It looks like the same brand talking to eight different people in eight different ways on the same afternoon, with no campaign brief anywhere in the middle. Koa Botanicals is an invented brand, built out properly so the output can be judged, and no result is being claimed.
You are about four days from the end of the bottle
Your last two bottles lasted 47 and 45 days. This one landed 43 days ago, and a new one takes three days to reach Bristol.
No discount attached, because you have never needed one to reorder. If you have slowed down and there is plenty left, ignore this and the next one moves back on its own.
The email is the obvious surface. The harder decisions are the ones about channel and restraint.
Was the 100 ml too much oil to get through?
One question, and nothing is sold on the back of it. It changes what we send you next.
You are two orders from the refill tier
From your third order the 100 ml refill opens up, at £0.96 per ml against £1.27 in the 30 ml. Nothing to do now, it just appears when you get there.
Eight people, eight moves, one of which sells nothing, one of which asks a question instead of making a claim, and one of which is an ad deliberately hidden from most of the list. A campaign calendar has no slot for any of these, because each is true for one person and almost nobody else this week.
Why do point tools not solve this?
Because each one owns a surface and none of them owns the relationship. A brand at $20M is rarely starting from zero: expect a sending platform, a support agent, an analytics agent and two or three point tools already installed. Every one is good at its job.
| Tool | The one job it owns | Where it stops |
|---|---|---|
| Relo | Replenishment timing | Predicts when the jar runs out. It does not decide whether a reminder, a refill offer or silence is right for that person this week. |
| Rep AI | On-site conversational selling | Owns the session. Nothing happens once the visitor closes the tab. |
| Black Crow AI | Predictive scoring | Produces a score and hands it to your sending tool, where the decision is still a flow somebody built. |
| Gorgias, Siena, Yuma | Ticket resolution | Resolves the inbound question. Never initiates the outbound relationship. |
| Postscript, Attentive | The SMS channel | Excellent at the channel, and blind to what the same person is being sent on email that morning. |
| Triple Whale, Northbeam | Attribution and analysis | Tells you what happened. Acting on it is a person's job on Monday. |
Seven tools each optimising their own surface will still leave one customer getting four uncoordinated messages in a day and another getting nothing for eleven months. The judgement about whether this particular person should hear from you at all lives nowhere.
Does per-customer decisioning only make sense at enterprise scale?
No. It is usually described as an enterprise capability when it is really an enterprise price, which is a different problem. The architecture does not care how many customers you have. The meter does.
| Where you are | What breaks first | What changes |
|---|---|---|
| $1M to $10MFounder led | Nobody owns retention. Three flows were built once and have not been touched since. | The work happens without a hire. You set the goal and the limits, and the judgement stays yours. |
| $10M to $50MThe beachhead | Flow sprawl, segment maintenance, an agency retainer, and a contact bill rising faster than the list earns. | The flowchart stops being the unit of work. One goal in a sentence replaces the branching diagram nobody wants to touch. |
| $50M and upTeam in place | Real lifecycle headcount spending its week on QA, build and reporting rather than strategy. | The team keeps the judgement and stops doing the assembly. Depth per customer becomes possible at a size where it never was. |
| Beyond ecommerceApps, subscriptions | The same problem in a different container: millions of users, a handful of segments, one calendar. | The SDK drops into a website or an app, so the loop runs wherever the customer is. |
Three of the strongest platforms on this map require a data warehouse before the first decision is made, and one requires a separate sending platform on top. For a brand under $50M that is a quarter of engineering time before anything is learned. Reading Shopify directly is the difference between a project and an afternoon.
What should you ask a vendor claiming to be agentic?
Five questions, none of them about models or architecture. Each has an answer a rung-four platform gives immediately and a rung-two platform cannot give at all.
- Show me the last hundred decisions. Who each was for, what it chose, and what it chose not to send. A platform that holds campaigns rather than decisions cannot produce that list, because the list does not exist.
- What did it earn against a group you left out? Not attributed revenue and not last click. The difference between the customers it worked and a share you deliberately held back.
- What has to be true before the first decision? A warehouse, a separate ESP, an SDK release, a services engagement. Add the answer up in weeks, then ask what it costs to find out this was not for you.
- Can it decide to send nothing? Silence is the most common right answer in retention and the hardest to build, because it earns the vendor nothing on a contact-based meter.
- Who writes the message, and against what? Ask whether prices, stock and links resolve against your live catalogue at send time, or whether the model recalls them.
How does pricing on extra sales compare to contact-based pricing?
Contact pricing bills you for how many people are stored. Pricing on extra sales bills you for what the software added: a group of your customers is held back and gets no messages, and you pay a share of what everyone else spends above that group. Grow a list from 40,000 to 90,000 and a contact bill roughly doubles, whether or not those extra 50,000 people ever hear from you in a way that works.
The incentive that creates is perverse: prune the list to control the bill, which is the opposite of what a retention programme should be doing with dormant buyers.
| What you are billed for | Contact based | On extra sales |
|---|---|---|
| The meter | How many people are stored | What everyone spends above a held-back group |
| A dormant contact | Billed every month | Costs nothing |
| Doubling the list | Roughly doubles the bill | Changes nothing on its own |
| Choosing restraint | Costs you the same | Costs nothing |
| Seats | Priced per user | None |
When you pay on extra sales, a decision that this person should hear nothing this week costs nothing, and if the messages add nothing over the held-back group, the bill is nothing too. That is the right way round.
PilotX costs 10% of the extra sales it adds, and nothing if it adds nothing, capped at $2,500 a month, with no seats, no per-contact fees and no per-message fees. On a store with 150,000 customers the list size does not change that. Say PilotX adds $8,000 of extra sales in a month: the bill is $800. If it adds $40,000, the bill is $2,500, because it is capped at $2,500 a month. Modelled at full capacity against a control group you set, the return is up to 50% more revenue from the base you have already paid to acquire. Modelled, and never a guarantee.
Whatever ends up on your shortlist, check the claims attached to it before the demo. We ran 20 claims about the platforms on this page against each vendor's own primary source and eight of them failed, including one general availability date out by six months and one entry price understated by a factor of 2.4.
Where should a Shopify brand start?
With evidence from your own store rather than another vendor deck. Only 8% of CMOs are running autonomous multi-agent campaigns and 65.7% of marketers name data integration as their biggest martech challenge, which means almost nobody in this market has proof on their own data yet.
The free Revenue Leak Audit walks four mystery shoppers through your store the way real customers would: signing up, browsing signed in, filling a basket and leaving, starting checkout and leaving. We write down what came back and what did not, and put your own numbers against the silence. It takes nothing from you but the store URL, and the write-up is yours whether or not you ever buy anything. If you want the mechanics first, the Shopify solution page covers how the four agents connect and what runs in the first afternoon.