Marketo's co-founder launches marketing automation that reasons instead of following rules

On Sept 23, Marketo co-founder Jon Miller and former Marketo product EVP Nick Bonfiglio launched Phave, a marketing automation platform built around a reasoning engine called Maestro. Instead of running a flowchart, Maestro builds a "Playlist" for each contact: the order and timing of touches most likely to reach the objective the marketer describes in plain language, weighed across every campaign queued for that person. Accounts and buying groups are records of their own, with their own scores and stages. Pricing starts at $36,000 a year and is based on monthly active recipients rather than database size. The launch names HubSpot, Marketo, Pardot and Eloqua as the platforms it's built to replace.

Miller draws the line himself: "Rules are good at what is allowed, but bad at deciding what is best." Consent, frequency caps and quiet hours stay as rules, and the decision about what to send next moves to the engine. For marketing ops, the build work shifts from branching logic to writing objectives and guardrails. For product marketing, most of what the system reasons from becomes the messaging inside each campaign, plus the description of who it's for and what it should accomplish.

A useful check before evaluating anything in this category: take one nurture program the team already runs and write down its audience, its objective and the three things it has to land, without looking at the flowchart. If PMM and ops write different answers, the engine will be reasoning from both. Treating buying groups as records of their own also means someone has to define which roles make up a buying group in each segment. That job tends to fall between PMM, ops and sales.

Microsoft gives Copilot agents their own email addresses, and partner outreach is one of their jobs

On Sept 25, Microsoft folded chat, full versions of Word, Excel and PowerPoint, a coding tool and Autopilot into one Copilot app. Autopilot agents are always on and have their own identity, memory and email address. You can tag them in Teams or Outlook like a colleague. Fortune reports their jobs include "project updates, tracking work decisions and deadlines," "reaching out to partners" and "reminding colleagues when commitments are due." Seats cover everyday chat and Office, while Autopilot, Code and the newest models can be billed by usage. Autopilot's private preview expands at the end of the month.

For partner, alliance and channel teams on Microsoft 365, the follow-up to a partner may soon come from an agent. It will draw on whatever partner messaging, co-sell material and program terms it finds in the tenant. Two things are worth checking now: which version of the partner pitch lives in SharePoint and Teams, and whether anyone would sign off on an agent quoting it. For teams that sell agents, the billing split shows how enterprise buyers are being taught to budget: seats for assistance, usage for autonomous work.

OpenAI and Anthropic cut model prices on the same day

On Sept 22, OpenAI released GPT-6 Sol and Luna at roughly half the price of the 5.6 series and says the prices are permanent. Luna costs $0.10 input and $0.50 output per million tokens and is aimed at high-volume summarization and extraction. Sol costs $2 and $10 and is aimed at recurring coding and agent work. The same day, Anthropic released Claude Opus 5.5 and cut output pricing from $25 to $20 per million tokens.

The tasks GTM teams hand to models most often got cheaper at both labs on the same day: call summaries, pulling fields out of documents, first drafts. Teams that shelved a research or content agent because of run cost should redo the math. As per-token cost falls, more of the remaining effort in these workflows goes into preparing and maintaining the material the model works from.

SOCi passes 412,000 marketing agents in production

On Sept 22, SOCi said it has deployed more than 412,000 of its Genius Agents for localized marketing, up 37% in a quarter. The agents manage local search listings, local social content and review responses in brand voice. CEO Afif Khoury: "Enterprises added more than 100,000 agents in a single quarter. That is not experimentation."

The figures are vendor-reported, but they describe agents publishing to customers, not pilots. SOCi's headline quality metric is a 98.7% publish-ready acceptance rate, which measures how well the agents match the brand guidance they were given.

How it connects: all four items point the same way. Phave reasons from the objectives a team writes down, Autopilot works from what's in the tenant, SOCi's agents answer in brand voice, and the labs made running all of them cheaper. In every case, the output depends on the written material the system is handed.

My take

I once knew a founder who saw where human-AI collaboration was headed well before the rest of us, and his conviction raised a round just shy of unicorn. When he wasn't in the meeting, his team brought the story down to what the software did that quarter, because that was the version they could actually say out loud. Nobody butchered anything; under pressure, they reached for the version that fit.

A reasoning engine is going to do exactly what that team did. Phave, Autopilot and every agent that gets cheaper to run this fall will work from the version of your story that's written down and findable. At a lot of companies, that means the feature-level version, a few messaging docs that disagree with each other, and a partner deck from two launches ago. That's GTM debt, and it's been easy to live with because a person in the loop was quietly smoothing it over before anything went out.

The fix is the same one it's always been, and it's worth a lot more now: write down why you win, who you win with, and what you won't claim. Write it clearly enough that someone (or something) who never met your founder can carry it into the room for you.