Ryan Bozeman @RyanOpsAlien

I’d love to stop giving AI the same brand rundown every time I start a task. Here’s the audience. Here’s the offer. Here’s how we sound. And please stop using that phrase we’ve already changed twice. Meta’s new Muse Wonder models put a lot of emphasis on remembering context across conversations. That’s something I’d want to test for marketing work. Can it remember approved guidance? Pick up feedback from the last draft? Use what we learned from an earlier campaign? I’d also want an easy way to update what it remembers. Prices change. Offers end. The team settles on a better way to explain the service. Old information needs to get replaced. For the AI systems I build, I’d test this over several real assignments. See whether the same corrections keep coming back and how much time the reviewer spends fixing things. Remembering useful feedback could save a fair bit of back-and-forth.

Google is testing a way to set up ads by describing what you want in plain English. AI then suggests campaign goals and assets. I can see the appeal. I’d still want a decent brief before touching it. “Get us more leads” gives it a lot of room to guess. Who are we trying to reach? What are we selling? Where do we operate? What’s the budget? What makes a lead worth following up with? When I build marketing automation, I want those answers written down. Along with the offers we can use, the claims we can make and where enquiries should go. Then I’d check what AI suggests against that brief before launching anything. The Google setup is still an experiment. But getting the brief sorted is something a team can do now—and use whichever tool they end up working with.

Movable Ink’s new AI tools got me thinking about how much time goes into running an email campaign. Pull the results. Figure out what happened. Come up with the next test. Write another version. Get it approved. They’ve announced assistants for analysis, brand guidelines and subject lines, plus an early-access connection to tools like Claude and ChatGPT. I’d want to try that on one campaign we run regularly. Have AI look at the results, explain what changed and suggest a test. Then draft the next version using the brand guidance. Someone checks the analysis and approves the changes. That’s the kind of workflow I like building for companies. After a few rounds, I’d want to know: did it save time? How much did we have to fix? Did we learn something useful from the test? Those answers would help me decide what to automate next.

@atishayhyperke This is about as old as cold email itself haha

OpenAI’s latest ad announcement got me thinking about what happens after the click. It’s expanding measurement integrations and planning to test a new visual ad format later this month. Interesting channel. Plenty to test. But before I’d recommend putting much budget into it, I’d want the basics working: conversion tracking, clean CRM records, lead routing and timely follow-up. Because getting someone’s attention is only one step. If their enquiry disappears into a spreadsheet or nobody follows up, the campaign has a problem. That’s where I spend a lot of my time building AI marketing systems: connecting those steps so the team can see what happened and act on it. I’d start with a small test and a clear definition of success. Then check whether it brought in additional business. A new channel is a lot more useful when the process behind it works.

An ad campaign slows down, and everyone starts looking at the creative. Meanwhile, the product might be running out of stock. Or a checkout update might have broken conversion tracking. That’s why Snowflake’s new advertising agents caught my attention. They’re in preview, connecting Meta and TikTok campaign signals with company data. For me, that’s a useful direction for marketing automation. Before AI recommends changing a budget, I want it to check what’s happening across the business. Then show its reasoning and flag anything it couldn’t verify. I’d start with recommendations that a person reviews. Give the system more control as it proves useful. When I build these workflows, getting the right information into the system is a big part of the job. You can automate a decision pretty easily. Making it a good decision takes more work.

Claude working inside Google Docs, Sheets and Slides feels like a useful update for marketing teams. It’s still in beta, but it could make some everyday work a lot easier. I’m thinking about the weekly reporting routine: pull the numbers, clean up a spreadsheet, write a summary, then put everything into a deck. When I build AI marketing systems, that’s the kind of process I want to simplify. Start with a consistent template and clear instructions. Let AI put together the first draft, analyse what changed and suggest next steps. Have someone check the numbers, review the AI’s analysis and decide what to do next. Then measure whether it actually helped. Did the report take less time? Was it accurate? Did anyone use it to make a decision? There’s plenty of value in making the stuff your team does every week less of a headache.

Anthropic’s October 9 report about unintended agent actions caught my attention this week. It covers earlier incidents during testing and internal use. Anthropic described the reported impact as minimal and said it was removing live internet access from internal evaluations while it checks its controls. For me, there’s a practical takeaway for companies using AI in marketing: define what the agent is allowed to do to finish its job. “Research this account” needs a scope. “Update the CRM” needs rules about which information can change. “Improve this campaign” needs limits around budgets, targeting, and publishing. These are the details I think about when designing AI marketing systems and operations. Imagine an agent spots a drop in campaign performance and recommends changing the budget. Before anyone approves it, I’d want the system to show the evidence, the exact proposed change, and the account it applies to. Then I’d want a record of what happened and a way to reverse the change. If information is missing or the instructions are unclear, the workflow should flag the issue and ask for help. That makes it easier for the team to understand what they’re delegating and where they still need to be involved. I’m interested in systems that can take useful work off people’s plates. Getting there means making the process clear enough that someone can check it, manage it, and step in when something looks off. That’s a big part of what AI operations means to me. Which marketing actions would you always want someone to review first?

Anthropic’s Haiku 5.5 launch this week has me looking again at the small marketing tasks companies repeat all day. The model is designed for high-volume work, with lower pricing than its predecessor. That creates some interesting possibilities. Sorting inbound requests. Tagging customer feedback. Checking campaign names. Preparing a handoff before something reaches sales. Those jobs might seem small individually, but they’re exactly the kinds of things I’d look at when designing AI marketing systems and operations. There’s one number I’d want to understand before calling any of it a success: What does it cost to get a useful result? The model call is part of that. So are retries, connected tools, monitoring, and the time someone spends checking the output. If an AI workflow sorts an enquiry in two seconds, but someone then spends five minutes fixing the category and rewriting the handoff, those five minutes belong in the calculation. I’d also test different models against the actual work. A routine classification task might run well on a smaller model. A complicated account summary might justify something more capable. An unclear request might need a person. That gives you a way to build around the needs of the business instead of putting every task through the same setup. Lower prices give us more room to experiment. I’d use that room to test a specific workflow and track accuracy, review time, and cost per completed task. The result I’m looking for is pretty simple: useful work getting done with less effort from the team. Which repetitive task in your company would be worth another look?

Google’s Gemini agent announcement this week got me thinking about all the work that happens around an AI tool. Google describes an agent that can plan work, connect to business systems, and bring back something finished. That sounds useful. But what does “finished” actually mean for a marketing team? A campaign report sitting in a document? A recommendation someone can act on? A task assigned to the right person, with the information they need? When I think about building AI marketing systems for a company, that’s where my attention goes. Take a weekly campaign review. Pulling the numbers and writing a summary is one part of the job. The system also needs to understand what the business considers a qualified lead, which campaigns are still being tested, and whether sales followed up. Then the result needs to reach someone who can do something with it. Otherwise, everyone gets another report to read on Monday. My starting point would be one recurring workflow. Define where it starts, which information it uses, who checks the result, and what happens next. Then measure the whole process, including the time people spend reviewing, correcting, and chasing things down. That’s what interests me about this next wave of AI agents. There’s an opportunity to make the day-to-day work inside a company run more smoothly. A useful system should give the team fewer loose ends to deal with. What’s one recurring marketing task you’d happily hand off?

Google opened SynthID Detector to everyone. It checks for a specific AI watermark. No watermark detected doesn’t prove human authorship. My marketing takeaway: keep asset sources, tool history, and permissions alongside the creative.

WARC forecasts $1.34T in global ad spend this year, up 11.9%. Before raising your budget, ask: are campaigns creating additional sales and customers who stay? Industry growth provides context. Your next dollar needs its own evidence.

ChatGPT’s new Intelligent UI can turn answers into interactive comparisons and tools. My marketing takeaway: audit your pricing, product differences, and proof. If AI helps buyers evaluate you, vague positioning becomes a liability.

ServiceTitan’s September 24 survey release puts improving cash flow among the top three goals for 40% of commercial contractors. Acquiring new customers was a top-three goal for 29%. The survey covered 1,020 US commercial owners, executives, and general managers, mainly in mechanical, electrical, and plumbing. Those percentages shouldn’t be applied to every service business. For firms serving that audience, though, it’s a reason to check whether the content plan addresses anything beyond finding new customers. I’d make room for explanations existing customers can use: what a service agreement covers, what sits outside the scope, and how to prepare for the next visit or project. Start with questions the team actually receives. Ask the person doing the work to explain the answer, then check it against the agreement before publishing. A new prospect may find those answers useful too. But their first job is to help a customer understand what they’ve bought and what happens next. servicetitan.com/guides/commerc…

Forrester’s September 29 research summary says 50% of marketing decision-makers report that AI often gives incorrect information with unjustified confidence. That’s a finding about the surveyed marketers. It doesn’t tell us how often a particular firm’s drafts are wrong. My takeaway for a B2B service firm: include the checking work when you assess an AI content process. A draft produced quickly can still take a long time to review. Someone has to find the source, check what it supports, and catch wording that promises more than the firm can deliver. For the next few posts, record how long drafting and review take separately. Keep a note of what the reviewer had to repair. A missing source calls for a different fix than an inaccurate description of the service. Use those notes to decide what the process needs next. It might be better source material, clearer offer notes, or a different task for AI. The useful measure is how much work it takes to reach a post you’re comfortable publishing. forrester.com/blogs/as-ai-ad…

HubSpot announced on September 30 that you can build and edit landing pages in Claude using HubSpot templates, then publish them to HubSpot. Marketing emails can move from Claude into HubSpot as drafts. That’s a useful distinction to keep in the setup: a page can go live, while an email arrives for review. For a service firm trying this, I’d start with one email draft. Check that the copy arrives correctly, the reviewer knows where to find it, and the final version says what the firm actually offers. Then look at the landing-page path. Decide who can publish and when they need someone else to check the page. Being able to send content straight from a conversation makes that decision more pressing. The connection could remove a tedious handoff. I’d judge it by whether the person reviewing the content gets a usable draft with the right context, and whether everyone knows when it becomes public. community.hubspot.com/t/write-in-cla…

A shorter version of an approved post needs its own review. Here’s a hypothetical example. A consultant writes: “For firms with a stable offer and someone available to review drafts, this is a reasonable place to start.” During adaptation, that becomes: “This is the best place to start.” The shorter line loses the conditions and makes the recommendation stronger. It needs another look, even if the original post was approved. For a service firm, I’d put channel adaptation before final sign-off. Let the hook, examples, and length change to fit the channel. Then check that each version still says what the expert meant and only claims what the firm can support. Keep the exact approved copy with its source notes. If someone changes a claim after approval, return that version to review before scheduling it. Give one person responsibility for the handoff so the scheduler receives the version that was cleared. That makes approval easier to trace when you later need to explain what was published, where the claim came from, or why a correction is needed.

A review correction can help with more than the post in front of you. Suppose an AI draft describes a service your firm doesn’t actually deliver. You fix the sentence and approve the post. If the next assignment starts with the same old context, the same mistake can come back. I’d keep a small record of corrections that change how future drafts should be written. Save the original line, the approved wording, and the reason for the change. For example: “We can advise on this, but we don’t deliver it.” That explains a boundary a writer can apply to the next piece. Different edits deserve different treatment. A factual correction belongs with the evidence. A change to the offer belongs in the service notes. A preference for a less formal phrase belongs with voice examples. Some edits are specific to one post. Keep them there rather than turning every reviewer preference into a permanent rule. Start with a few corrections that would actually change the next draft. Give someone responsibility for keeping those notes current and making sure the drafting process uses them. Otherwise, the reviewer gets to explain the same thing all over again.

“What should we post this week?” asks a busy expert to do a lot of work before they’ve even answered. They have to choose a topic, decide what the audience cares about, and start turning it into content. I’d start with a question closer to the work: “What did a prospect misunderstand on a recent call?” For a service firm, that could be a question about scope, a price comparison that missed something, or a request the team had to push back on. Use an actual conversation, with client details removed. Then follow the explanation a little further. What did you tell them? Why? When would that advice change? That last question is worth keeping. A consultant’s qualification or exception can be the most useful part of the answer. It shows the reader how to make a decision in their own situation. AI can organize the notes and draft a post. The expert should check whether the draft kept the reasoning intact, including the part that made the advice conditional. I’d rather start with one explanation the team can stand behind than ask them to fill a calendar from scratch.

Before you fill the time an AI content workflow saves with more posts, give some of it to the harder work. Talk to customers. Check the claims. Get the founder’s explanation right. Measure what you made room for, alongside how fast you produced it.