If you want an early read on what clients will buy next quarter, read job postings. They are the first public signal of what a company has decided it can no longer operate without. Right now the signal is unusually specific: employers are hiring people who can ship working automations, not people who write clever prompts.
What the Postings Actually Ask For
MarTech reviewed 60 live marketing job descriptions published on employers’ own careers pages in late September. In 45 of the 60, designing or shipping AI agents and automated workflows is a listed requirement. Only two call it a nice-to-have. One posting says it plainly: AI is the way the role will be carried out, and the employer wants a builder’s mindset — “workflows and agents, not just prompts.”
The titles have not caught up. Only one of the 60 includes the term “AI-native” in its heading, and 33 mention AI nowhere in the title at all. The requirement is sitting inside ordinary marketing operations roles whose job descriptions quietly changed underneath the same name.
Wider labor data agrees. Indeed’s Hiring Lab found the share of marketing job ads referencing AI rose from 8.4% to 14.9% during 2025, with AI-mentioning postings growing more than 130% while overall hiring stayed flat. PwC’s 2026 AI Jobs Barometer, drawn from more than a billion job ads, puts a 62% wage premium on AI skills. The read for an agency is blunt: the internal capability clients are now staffing for is the deliverable you should be selling.
The Ladder: Understand, Describe, Automate, Ship, Prove
The most useful part of the research is that it names the order of operations, and most teams start on the wrong rung — buying an agent-building course before they can write down how one of their own processes works. An agent only performs the process you hand it. Hand it an undocumented process and you get output that looks polished, reads confidently, and is wrong in ways nobody notices for a quarter.
- Understand it. A model predicts the next piece of text from the context you give it. An agent is configured software that performs one job on a schedule with a specific set of tools, writing to a specific place. A skill is the written instruction set that makes it work your way. MCP is how it reaches your own systems instead of guessing from training data.
- Describe it. Write the process as numbered steps someone else could execute. This is where agency work usually stalls: voice, product truth, and claims rules live in a 40-page brand PDF with mood boards, and an agent cannot read a mood board. Voice becomes real example passages with a note on why they work. Product truth becomes one page per product with a named owner. Claims split three ways — say it, say it with a link, never say it.
- Automate what repeats. A good prompt becomes a reusable file, the file becomes a skill, the skill gets a trigger and becomes an agent.
- Ship one small thing. Take one manual task end to end: build it with a coding assistant, put it in a repository, deploy it so it has a URL other people use. Publishing is what forces scope discipline and surfaces the requests you did not anticipate.
- Prove it still works. Define the business outcome, then check at 30, 60, and 90 days. Automations decay: an agent that worked in March is not evidence it works in September.
The Stack Clients Will Name
Employers have stopped asking whether candidates have “used AI.” Thirty-six of the 60 postings name at least one specific product. Claude appears in 26, OpenAI or ChatGPT in 19 — a surprising tilt if you assumed the market centers on ChatGPT. One orchestration platform appears repeatedly (n8n, Zapier, or Make — the concepts transfer between them in an afternoon, so standardize on whatever the team already pays for). MCP shows up in eight. The build chain is consistently three parts: a coding assistant, a repository, and a host.
Treat that list as perishable. Roughly half of it will be different within two years, which is an argument for the ladder rather than for tool loyalty.
Why Finance Is Now in the Room
Selling agent work is getting easier and harder at the same time. Salesforce’s CFO Priorities Report, a survey of 865 finance leaders across three continents, found that nearly three-quarters say their role expanded in the past year — and managing AI use and expansion is the most-cited reason, ahead of revenue complexity. Half now describe themselves as primary decision makers in their company’s AI strategy.
The good news: among finance leaders using AI, about 90% report positive ROI, and agent users report gains in time savings, productivity, cost savings, and forecast accuracy. The friction sits in security and governance concerns, integration with existing systems, and a shortage of internal subject-matter expertise — the exact gaps an agency can productize around.
Practical consequence: your economic buyer may no longer be the CMO. Whoever signs will ask what controls exist, what happens when the agent is wrong, who approves a correction, and how you will demonstrate the thing still works in 90 days. Answer those four questions in the proposal and you skip a procurement round.
The Competitive Clock: Vertical Agents Are Shipping
Agentic software is no longer a demo category. In September, Vertafore released four AI agents into independent insurance agencies under its Agentic Agency model: document research, an SOP-grounded answer agent living inside the agency management system, policy intake, and policy checking. Early-adopter testing reported document review time down up to 90%, time spent hunting for answers down up to 88%, and policy intake up to 87% faster.
The design choice matters more than the percentages. The agents extract, map, and propose — then a person reviews, edits, or rejects before anything is written. Human judgment stays on the approval step; the agent takes the assembly work. That pattern is directly portable to most agency service lines and it is the one clients will eventually ask for by name.
What to Do This Week
- Pick one process a client runs every week. Write it down as numbered steps with its inputs and decisions, and charge for that document — it is the cheapest audit you will ever sell, and it reveals the approvals nobody can justify.
- Turn the most-repeated step into a skill file with a trigger attached.
- Write ten input-output pairs and put a monthly rerun on the calendar.
- Get brand voice, product truth, and claims rules into machine-readable form before your next agent build.
Prompt tips are free and everywhere. A documented process, a running agent, and monthly proof that it still works are what employers — and clients — are now paying for.