The new pitch for marketing automation is that rules should decide what is allowed and a model should decide what is best. It is a sensible split. The trouble is that in a five-year-old instance, the two kinds of logic are wired into the same campaigns.
On 23 September Jon Miller, a co-founder of Marketo, announced Phave, a marketing automation platform that he and former Marketo product head Nick Bonfiglio have been building for two years. Its engine, called Maestro, takes campaigns the marketing team has built and approved and computes an ordered sequence of touches for each person, weighing everything queued before anything is sent. Accounts and buying groups are treated as records in their own right, with separate scores and lifecycle stages. Pricing starts at $36,000 a year and is based on monthly recipients reached rather than database size.
The launch lands in a month when every incumbent has made a similar move without asking customers to migrate. Salesforce introduced a campaign agent at Dreamforce. HubSpot rebuilt its Breeze assistant and added campaign and content agents at its UNBOUND event. Adobe has Marketo AI agents in open beta and an MCP server exposing more than a hundred operations. Demandbase released an agent that builds audiences and runs campaigns across the major ad and automation platforms.
Phave also published a comparison scoring itself at 2.97 against HubSpot at 2.46 and Marketo at 2.38 across 481 requirements, with the weighting done by an AI model and the scoring done by Phave. It is a vendor exercise and should be read as one. The more durable contribution is a line from Miller's announcement.
"Rules are good at what is allowed, but bad at deciding what is best. Consent, frequency and quiet hours should be rules. Choosing which of two good emails a person should get was never a job for a rule."
Two kinds of logic, one set of campaigns
That sentence describes a split every marketing operations team should be able to draw in its own instance, whichever platform or agent it ends up using. Call the first kind permission logic: consent status, regional compliance, frequency caps, quiet hours, suppression lists, sales-owned exclusions. It decides what may happen, and it should behave identically every time. Call the second kind choice logic: which asset, in which order, after how long, at what score someone is handed to sales. It decides what should happen, and it is where a model can plausibly do better than a rule written three years ago.
The pitch from Phave and the incumbents alike is that models take over choice and leave permission alone. In principle that works. In practice, a mature instance was not built with the split in mind.
Figure 1
Logic | Kind | Where it often lives in an older instance | Risk if a model takes over the campaign |
|---|---|---|---|
Regional consent check | Permission | A filter inside individual nurture programmes | Removed along with the programme it sat in |
Frequency cap | Permission | Wait steps and "not sent in last 7 days" filters spread across flows | No longer enforced once flows stop being the unit |
Sales-owned exclusions | Permission | A manually maintained static list referenced by some campaigns | Silently ignored by anything that doesn't reference it |
Asset sequence | Choice | Hard-coded nurture streams | The intended target for replacement |
MQL threshold | Choice, with a contract attached | Scoring model agreed with sales in an SLA | Handover volume changes without the SLA changing |
Permission logic scattered through campaigns is the main migration hazard. The examples are common patterns in long-running instances, not findings from an audit of any particular platform.
Hand choice to a model in an instance like that, whether by migrating or by switching on an agent layer, and you risk removing permission logic nobody knew was there. The consent filter that lived inside the EMEA nurture stream goes when the stream is replaced by a computed sequence. The frequency cap implemented as a wait step disappears with the step.
Two further things that change
The reporting unit. Programme reporting assumes the programme is a fixed object: this nurture, these emails, this conversion rate. When sequences are computed per person, the nurture no longer exists as something you can report on. Performance attaches to the engine's decisions, and "which campaign worked" becomes a harder question. Teams should decide what they will report on before the change, not reconstruct it for the first quarterly review afterwards.
The handover. Account and buying-group scores are a better fit for how B2B purchases happen than individual lead scores. But the MQL threshold is usually written into an agreement with sales about volume and response times. If a model starts making that call, the number of handovers will move, and the agreement needs to move with it.
An audit worth doing before evaluating anything
Inventory permission logic.Find every consent, frequency, suppression and exclusion rule, wherever it sits, including inside individual campaigns.
Pull it into one layer.Centralise permission rules so they apply to everything sent, regardless of which campaign, agent or engine triggered it. Get legal sign-off on that layer specifically.
Tag choice logic separately.Mark what could be handed to a model and what the team wants to keep deterministic.
Agree the reporting unit.Decide with finance and sales what will be measured once sequences are no longer fixed.
Reopen the SLA.Any change to how handover is decided should come with a revised agreement on volume and follow-up.
Miller has a name for the reluctance to switch platforms: marketing automation Stockholm syndrome, the fear that moving will break lead flow or compliance. That fear is usually treated as inertia. Often it is an accurate reading of an instance whose permission rules nobody could list with confidence. The audit above is worth doing whether a team buys a new platform, switches on its incumbent's agents or changes nothing at all.
Sources and notes. Phave announcement of 23 September 2026 (general availability August 2026): Maestro engine computing per-person sequences from marketer-approved campaigns; accounts and buying groups as records with independent scores and lifecycle stages; pricing from $36,000 a year based on monthly active recipients; a self-published comparison across 481 requirements with weights assigned by an AI model and scores assigned by Phave (Phave 2.97, HubSpot 2.46, Marketo 2.38). The quotation is from Phave's release. Incumbent announcements (Salesforce campaign agent at Dreamforce; HubSpot Breeze Assistant and Marketing Studio at UNBOUND; Marketo AI agents in open beta and MCP server; Demandbase Mojo) as reported in trade coverage, including CMSWire and Solutions Review. Product details have not been independently tested and should be confirmed with each vendor. MarketingHubMedia has no commercial relationship with any vendor named. Journalism, not procurement advice. Corrections welcome.
