Turn 80% of Your Marketing Into Repeatable AI Recipes

Last Thursday, Barclays said it will cut about 5,000 jobs this year, roughly 5% of its staff. Its CEO, CS Venkatakrishnan, was careful about how much of that he credits to AI. He called it "the icing on the cake" and said "it's not the cake," because "you need to do the infrastructure build to get the benefits of AI."

Earlier this month, HubSpot cut about 660 jobs. Its CEO told staff the cuts were "not driven by AI-related efficiencies."

Marketing teams are already feeling it. In Spencer Stuart's survey of CMOs, 17% said they had cut staff because of AI in the past year, and 36% expected to cut in the next two years.

So two years in, what has AI really changed inside a marketing team, and how much of the work can it take on?

I asked someone who should know. Utpal Bhatt is CMO of Luminary, an AI company backed by NVIDIA's venture arm. Its models predict how a car or an aircraft will behave before anyone builds it. Before that he ran marketing at Tigera, Qubole and Neo4j. If any CMO was going to swap his team for AI, it would be him.

Utpal's target is that "almost 80% of what we do should just be programmatic." He means work with a fixed process that can be written down as a recipe. His examples are running a webinar, compiling a win-loss report, building dashboards and reports, and even weekly status reports. His team turns that work into skills and agents.

The point is what the other 20% is for. AI frees his people for the work it can't do, such as talking to prospects about what they like and don't like about the product, and getting a deeper understanding of how the product works.

Start with this audit. Set aside an hour with each team lead.

A finished sheet for one person can be this short. Copy it and change the lines.

Here is what a written recipe looks like. You can paste each one into ChatGPT or Claude as it is, or give it to an agent.

Inputs: a CRM export of every opportunity created in the last 7 days (name, amount, stage, lead source, campaign, region, partner sourced, created date) and last week's report.

Steps: add up new pipeline by lead source, including partner-sourced deals. Compare each source with last week and with the four-week average. Flag any source that moved more than 20% either way. List the five largest new opportunities. Finish with three plain sentences: what went up, what went down and what to watch next week.

Rules: use only numbers from the export. If a field is missing or a number looks wrong, say so instead of guessing.

Good looks like: a VP can read it in two minutes, and every number matches the CRM.

Two more recipes from the same audit, ready to copy.

Inputs: the attendee list (name, company, role, attended live or registered only, questions asked in the chat), the webinar title and its three key points, and the recording link.

Steps: split the list into three groups: attended and asked a question, attended without a question, and registered but didn't attend. For people who asked a question, answer it in two or three sentences and link to the recording at the right moment. For people who attended, recap the three key points and offer a 20-minute call. For people who didn't attend, send the recording with the single most useful point.

Rules: under 120 words per email. Use the person's first name and company. Never invent a question someone didn't ask. Nothing sends until a person approves it.

Good looks like: each email reads as if the host wrote it that afternoon.

Inputs: every opportunity closed won or lost this quarter (amount, segment, loss reason, sales notes) and last quarter's summary.

Steps: count wins and losses by segment and by loss reason. Group the sales notes into recurring themes, and quote two or three notes for each theme word for word. Compare the top three loss reasons with last quarter's. List any deal where the loss reason and the sales notes disagree, so a person can check it.

Rules: quote notes exactly. Never turn a paraphrase into a quote.

Good looks like: product marketing can pick next quarter's priorities from it without opening the CRM.

Automating the recipes didn't empty Utpal's team. "I don't think that the roles themselves have disappeared," he told me. He hasn't found an AI product marketer that can replace the ones on his team, or an AI field marketer either. "You still need one of everybody," he said: product marketing, demand generation, field marketing.

What changed is the next hire. "You don't necessarily need your next hire, the second hire doing the same thing right away," he said. Each person now gets much more done, so the second product marketer or the second field marketer can wait much longer than before. His teams have got smaller since AI arrived, and "the kind of output that they're putting out is as much or even greater than when I had larger teams."

For your 2027 plan, ask how long you can wait before hiring the second person into each role.

Utpal was driving his son to a football match when his phone rang. A sales rep was on the line, and for a couple of minutes he thought he was talking to a person. Then he noticed signs that it was an AI. He put the call on mute, told his son, and started asking questions the AI wasn't scripted for. "They kept repeating the same answer," he said.

His worry is the brand. Every conversation with a prospect represents the company, he said, so if he hands that to an AI, "it better be as good as who I have trained on the team." On cold calls, where the AI knows nothing about the person, "I'm not a believer yet."

He is far more confident when the AI already knows something about the person. When someone visits your website or is already in your database, the AI can follow up well because "you feed it the context, you feed it history." That makes personal messages possible at a scale no team could manage by hand.

The rule is simple. Use AI to follow up with people you already know something about. Keep people on conversations with strangers until the AI can hold one.

Last week Semrush published a study of 400 US commercial search terms in banking, software and ecommerce. Google showed an AI Overview for 88% of them in August, up from 46% a year earlier. Your buyer's first answer about your category now often comes from an AI summary.

Utpal's team has changed how it writes now because of this. "We write in blocks," he said, so that each section answers one question and an AI can lift it cleanly. They also changed the way they expose their docs, and they provide MCP servers, a standard way for AI tools to read a company's information directly.

To start this week, take your five most visited product pages and rewrite each one as short sections that each answer one question a buyer would ask. Then ask ChatGPT and Google's AI Mode the same questions and see whether your answer appears.

Walmart recently launched tools that let AI agents shop on its site for customers. Shoppers can set spending limits for each agent or approve every purchase.

B2B isn't there yet. Utpal's buyers are still people going through a traditional sales process, and he has yet to see agents make and complete the buying decision. "I'm sure it's around the corner," he said.

When it comes, he thinks brand changes shape. Brand would focus on "the proof engine where you're actually getting the customers to advocate for you," with enough content out there backing you, and MCP servers that make your information easy for agents to read. Customer reviews, case studies and references become the brand, because they are what an agent can check.

The CMO's job is changing, we all know that but the role is not going away, because somebody still has to build the brand, run demand generation and report to the board.

The question is whether the person in the seat can "evolve and expand their skill set," including GTM engineering. If not, "you should cede things that you're not good at to other groups in the organization." That is what Utpal thinks.

His answer to what makes a CMO last is to be "a systems thinker." That means seeing how the whole function fits together, then building a system, now assisted by AI and agents, that runs it. He also warns against carrying one playbook from job to job. He has seen companies grow from about 40 people to about 350, and each size needed a different structure and a different mix of campaigns.

He traces much of this to his years in product and general management before he ran marketing. Watching customers use the product within their natural habitat taught him how to write for them. Running a P&L taught him respect for numbers like acquisition cost. His advice is to continuously question what you were taught, learn how product and sales work, and don't put on blinders by calling yourself only an events person or a media person.

Here is what Utpal’s approach could look like two weeks after a marketing team hands over the first three recipes from its audit.

For the first month, a person checks every run. In week two, the pipeline report missed two partner-sourced deals. The head of marketing ops fixed the recipe, not just that week’s report, so every run after it picks those deals up. Once a recipe runs four weeks without an edit, it can move to spot checks.

The second check is whether the freed hours turn into pipeline. Compare the accounts that got the extra prospect calls with a similar group that didn’t. That tells you roughly whether the calls help, but not how much pipeline they caused. To measure that, you need incrementality testing with a proper statistical design, statistically derived multi-touch attribution or marketing mix modelling.

Your hiring plan depends on the system underneath: written recipes, clean data, pages an AI can read and a team spending its time with buyers. Build that, and the question in your 2027 plan stops being how many people to cut. It becomes how long each role can run with one person, and what that person does with the time they get back.

Forward this to your marketing ops lead, and run the task audit with one team before your 2027 headcount plan is final. Watch my full conversation with Utpal Bhatt.

P.S. Agents can only follow a recipe if they can see the data behind it. RevSure connects your GTM data into one record, so a weekly pipeline report or a win-loss summary can run without someone pulling numbers by hand. See how it works.