Your Marketing Agency Sells AI. Where's the Proof?
Part 2 of my research into 350 selected marketing agencies: the people and dated AI projects agencies can show prospects in their proposals.
Kristin Tynski's team at Fractl put roughly 600 AI-generated articles on a marketing blog. She described it in an October 2019 interview, three years before ChatGPT.
Her team installed Grover, a publicly released language model, on their own servers. They pulled the titles of widely shared marketing articles, fed them in as prompts and bulk-loaded the output into WordPress. Tynski said the point was to start a conversation about what machine-written content would do to search and publishing.
In the same interview she described a second experiment: Salesforce's CTRL model on a GPU virtual machine, fine-tuned on speech transcripts.
Put that level of detail in your AI proposal: which model your team used, what they changed and what it produced.
I found the Fractl account while researching 350 selected marketing agencies. 217 of them publicly advertise AI services or tools. I then focused the background research on 82 of those agencies. Their histories turned out to be far more interesting than their job titles: developers, product leaders, automation specialists and a researcher.
What did Fractl actually do in 2019?
The interview, dated October 4, 2019, walks through the work: collect popular marketing titles, use them as prompts, get the released model running, load the generated articles into a site. It also records the CTRL fine-tuning experiment while that work was current, not as a memory reconstructed years later.
Two different skills sit inside that story. Running an existing model means getting its code and environment to work, which Tynski said was "a little difficult." Fine-tuning changes how a model behaves by training it on material you choose. Fractl did the setup and the experimenting. Grover and CTRL came from other researchers.
Give prospects the same detail about your team's relevant projects: why you chose that model, why you used that material and what happened when you ran it. Tynski's public notebooks give you more to inspect; my review stopped at the published material.
Fractl's 2019 account documents that experiment. Your proposal needs the same specificity about the people who will do the work now, with examples relevant to the engagement.
Who wrote the code behind a 2022 AI holiday app?
Park & Battery's write-up of its Holid-A.I. Gift Factory is dated December 1, 2022. The agency credits Matt Dignan with the technology and the code.
The stack was a Vue.js front end, Stable Diffusion, an NLP Cloud library and custom middleware built from scratch to connect them. Dignan spearheaded the technology work.
That's the part I want a buyer to ask about. The model was one component. Someone had to wire the pieces into something people could actually use, and the agency put a name on that person.
Dignan is now Head of Website Development. His bio lists more than fifteen years in web development, UX and design, including MRM//McCann, Number Six and Clearlink. Those fifteen years cover his whole technical career, with a documented AI integration in 2022 inside it.
The agency presented the app as a promotional project. I'd still use it to open the conversation. Walk me through the middleware decisions, then show me the work closest to my project.
Does a 2026 AI title mean someone is new to AI?
Patrick Kidwell's professional record lists him at Primacy as an AI Solutions Associate starting August 2026. Stop there and you'd figure he's new to this. Then read his personal site.
His retrospective on Margot, an AI assistant prototype, says the project began in November 2024 with FastAPI, OpenAI function calling and about five Google Analytics tools. Later he added memory, routing and browser automation. His writing index dates the retrospective February 23, 2026.
That's three dates answering three different questions: when the agency job started, when he says the project started and when he wrote it up. Mash them together and you get the wrong answer.
Margot is his project, not Primacy's, and the November 2024 start comes from his own retrospective. Read it on those terms. It still makes the August 2026 title a lousy shortcut for judging when he started doing relevant work.
If someone on your team built a relevant assistant outside work, include it and credit it as their project. Explain which parts they built and how that experience applies to the engagement.
What does an AI leader's title tell you?
Clayton Kjos at Aletheia is a leadership example. His timeline says he joined Pluralytics as Head of Product in September 2023, leading development of its language-AI and content-intelligence software. Aletheia's bio backs up the 2023 arrival and the product-development role.
After Aletheia acquired Pluralytics, his timeline moves him into Aletheia, and he dates his COO/CTO role to March 2026. Before that came software and product leadership at Fabric Interactive, Lola Tech Systems and Rakuten Link.
That history supports a conversation about running product development. It doesn't say what code he personally wrote, and it doesn't have to. Your proposal should explain whether that leader will direct the work, build part of it or advise the people who do. It's the same test I'd use when hiring a Chief AI Officer. What did you ship, and which part was yours?
What do course dates tell you about someone's AI experience?
Course completion dates establish when someone finished that course. They don't establish when that person started working with AI, and a certificate is not a prerequisite for the experience documented here.
StudioNorth's Caroline DeVore has recent AI coursework alongside a much longer technical career. Her personal bio lists a master's in Electrical and Computer Engineering and more than twenty-five years of development and leadership. Those are her biographical claims; I didn't resolve the degree dates or institutional confirmation. The twenty-five years describe her broader career, not twenty-five years working in AI.
Credential IDs on her professional profile match Coursera records under her fuller name: AI For Everyone, completed June 6, 2025, and Agentic AI and AI Agents: A Primer for Leaders, completed March 29, 2025. Neither date tells us when she began applying AI.
StudioNorth also publishes applied work: a GPT-powered brand-messaging and sales-enablement project for an unnamed technology client. The August 2025 version quotes DeVore as Executive Director, AI Practice. It describes team work and doesn't credit her with individual code.
Brandience's Ryan Grant has a different set of claims. His profile lists AI for Marketing from SmarterX and Claude Code 101 from Anthropic in September 2026, plus eight Anthropic courses in August. Brandience's company feed celebrates the marketing course. Those are his and his employer's claims; I didn't check issuer records.
Include relevant training if it helps explain the team you're proposing. Give prospects dated examples of the work and each person's contribution so they can judge the experience behind it.
What does "already serving clients" tell you?
In its 2025 Innovator of the Year post, Big Leap credits Aubrey Shepherd with an LLM dashboard already serving clients.
That's a client-facing claim, and it hands a buyer an easy follow-up. Which part did Shepherd build? What data does the dashboard pull from? Who runs it now?
The announcement doesn't supply those implementation details.
Which AI work fits the job you're hiring for?
CTI Digital's Saiful Thoombath explains an internal HR assistant in the agency's July 2026 write-up. The team indexed HR documents in a vector store, retrieved the relevant material for each question and connected it to Slack. CTI says the model runs privately on locally hosted infrastructure.
That tells you where the answers come from and how a question becomes one. His title on the case is Digital Innovation Specialist; his professional headline adds AI Automation. CTI describes the build as team work. Where the model runs matters too. If it's hosted by someone else, ask where your data actually goes.
If you're proposing an internal knowledge assistant, show how your team handled source documents, access and answers. For an AI-search visibility engagement, include relevant work in that field. A prospect should be able to see why you chose the example.
Lynton's Benjamin Agyekum brings research. With Fabio Santos, he co-authored a study of what happens to security when an LLM repeatedly repairs Infrastructure-as-Code. Across 5,968 scenario timelines, their strict detection method found security regressions in 3.3% of scenarios, and many of the broader flags turned out to be measurement artifacts. The lesson for a buyer: a scary warning isn't enough. Show how you checked it.
The paper went up on arXiv in August 2026 and was published in the ESEM proceedings in October. A professional post connects Agyekum to the paper and to Lynton.
An internal assistant, a client dashboard and a research paper are three different kinds of evidence. Any one of them tells a buyer more than a slide that says "AI experts."
What should an agency show before charging for AI expertise?
217 of the 350 agencies I reviewed publicly advertise AI services or tools. If your agency sells AI work, put the people and relevant projects in the proposal.
Include these specifics before the prospect has to ask:
- The names of the people who will work on the account.
- What each of them contributed to relevant projects.
- What each of them will own on this engagement.
- An implementation write-up, demo or repo the prospect can inspect.
If someone else wrote the middleware, name them. Explain what the senior person will review and label demos clearly. Give the prospect enough detail to judge whether your team's experience fits the work. Put it in the proposal before a question about who built what becomes an objection you have to answer on a sales call.
The people in this piece made that easier because someone published the work. Others may have work like this that I didn't find. If so, put it in the pitch instead of making the buyer guess. I made a similar argument about contract terms in Marketing Agency Predictions for 2027: Show Me the Contract, built on my earlier review of 50 agency brands.
This is Part 2. Part 1 is here. The full study is coming, with a deeper look at the data, public AI offers, the people behind them, their backgrounds and the projects I could trace.
If you're charging for AI expertise, show the work in the proposal.
Questions to answer in your proposal
What counts as relevant AI experience at a marketing agency?
Show a documented project and explain your team's contribution to it. That might be installing and running a model, writing an integration, directing product development or conducting research. Choose work relevant to the engagement you're proposing. Credit personal projects separately from work delivered by the agency.
What do AI job titles and certificates establish?
A job title describes a role. A verified certificate documents completion of a particular course; it is not a requirement for relevant AI experience. Neither a title nor a course completion date establishes when someone first worked with AI. Include dated examples and name each person's contribution so a prospect can judge the experience behind the title.
What should agencies include in an AI proposal?
Name the people assigned to the account, explain their relevant contributions and state what each will own. Include a write-up, demo or repository the prospect can inspect. Explain whether it shows an experiment, an internal tool or client work, and what it does not establish. Offer that evidence with the proposal rather than waiting for the prospect to challenge the claim.
Research note
The roster contains 350 selected marketing-agency brand/domain records. It is a historical selection, with observation depth varying by agency. It should not be read as an industry census or as 350 independent legal companies. The October 18, 2026 offer and staffing data connect role-linked public signals to 82 of the 217 agency records advertising AI services or tools. That is agency-record coverage; it does not count engineers or establish the staffing of the remaining agencies.
The October 19 background pass screened 187 original named candidate records across that 82-agency group, with two supplemental people recorded separately. At least one priority person at each agency received bounded source checks covering identity and current role, professional history or education, and dated AI work. Other named candidates were screened, but not all received complete career investigations; source depth varied. The initial records contain named current-role claims at 78 agencies, with four agency-person links unresolved. Personal and employer accounts are attributed throughout. I did not independently audit code performance, delivery or payroll. Project, publication, job-start and course dates describe different events; the earliest documented event does not establish when someone began learning AI.