Agentic Commerce Is the Mobile Moment. The Window Is Open. Are You Building?
Special Series: Agentic Shopping (5 of 5)
Key Takeaways:
⏵ The brands that own agentic commerce in 2028 are making infrastructure decisions right now — not in 2027.
⏵ The window to build ahead of the curve is open. It won't stay open.
⏵ Five posts. One argument: structure your data for machines before the agents arrive — or react to losses you can't explain.
This series started with a simple provocation.
Your next customer might not be human.
Over four posts we mapped what that actually means — not as a futurist thought experiment, but as a set of infrastructure decisions every brand needs to make right now.
The Arc, Compressed
Part 1: Is your data ready to be shopped by an agent?
Part 2: llms.txt — one file, two jobs. GEO citability and agentic transaction readiness. Build it once.
Part 3: The receiving stack — orchestrator, intake, retrieval, reasoning. A coordinated team serving every visiting agent.
Part 4: The audit — five layers, a scored gap map, a sequenced action plan.
Every post. Same conclusion.
The Strategic Imperative
The infrastructure is being standardized in 2026. The protocols are live. The agents are already being deployed by early movers.
This is the same window that existed with mobile in 2010. The brands that acted early built structural advantages that compounded. The ones that waited spent years catching up — and most never fully closed the gap.
Agentic commerce is that window. And it is open right now.
What the Winners Will Have Built
→ A machine-readable llms.txt serving both citability and transaction
→ Product catalog, pricing, and inventory structured for machine consumption
→ Data taxonomy governed consistently across every system an agent queries
→ A receiving architecture designed for visiting agents — not just human visitors
→ Agentic readiness audited before the agents arrived — not after the losses showed up
None of these are technology decisions. Every one is a leadership decision.
The agents are coming. The only question is whether your brand is ready to receive them — or scrambling to catch up when they arrive.
That's the series. Five posts. One argument.
Is agentic readiness on your radar? If you market/sell online - it should.
The Agentic Readiness Audit: The Five-Layer Framework Every Brand Needs Before the Agents Arrive
Special Series: Agentic Shopping (4 of 5)
Key Takeaways:
⏵ Agentic readiness can be audited today — before the agents arrive.
⏵ The gaps aren't where most brands expect them. They show up in taxonomy, accessibility, and transaction logic — not design or content.
⏵ A scored readiness framework gives you a gap map and a sequenced action plan. That's the starting point.
In Part 3 we mapped the receiving stack — the coordinated team of agents that serves a visiting agentic shopper.
Now the obvious question: how ready is your stack right now?
A New Kind of Audit
The marketing industry has SEO audits. GEO audits. UX audits. Conversion audits. None of them answer the question an agentic shopper is actually asking. Can I navigate this brand's data environment, evaluate options, and complete a transaction — without a human in the loop?
That's a different question. It requires a different audit. And it produces a different kind of score.
The Agentic Readiness Scoring Rubric
An agentic readiness audit evaluates five layers — each one a potential failure point in the receiving stack:
1. Discoverability
🔴 No machine-readable index exists → 🟢 llms.txt published, structured for citability and transaction
2. Data Structure
🔴 Fragmented across platforms and spreadsheets → 🟢 Structured and accessible without analyst intervention
3. Taxonomy Consistency
🔴 Inconsistent definitions across platforms → 🟢 Unified taxonomy governed across all data sources
4. Transaction Accessibility
🔴 Purchase pathway requires human navigation → 🟢 Transaction logic structured and agent-accessible end to end
5. Response Integrity
🔴 Data gaps produce incomplete agent responses → 🟢 Data quality governance ensures recommendations reflect reality
The Gap Map
Most brands auditing against this rubric will cluster at 🔴 or 🟡 across all five layers — not because of bad decisions, but because none of these systems were designed with a non-human visitor in mind.
That's the gap map. And it's more useful than a single score.
Each layer that scores 🔴 or 🟡 is a sequenced action item — prioritized by which failure point sits earliest in the receiving stack. Discoverability first. Transaction accessibility last. Fix the handshake before you fix the handoff.
An agentic readiness audit doesn't tell you how good your website is. It tells you how ready your data infrastructure is to serve a visitor who doesn't have eyes.
Where This Is Headed
Agentic readiness scoring is an emerging category. The rubric doesn't exist as a standard yet. The audit methodology is being defined right now — by brands and practitioners willing to ask the question before it becomes urgent.
The ones who audit now will have a gap map, an action plan, and a head start. The ones who wait will be reacting to agent traffic they can't explain.
Part 5 is the strategic imperative — why brands that build this infrastructure in 2026 own the channel by 2028.
The Visiting Agent Never Knew It Was Talking to Five AIs. It Just Got a Clean Answer.
Special Series: Agentic Shopping (3 of 5)
Key Takeaways:
⏵ When an agentic shopper arrives at your brand, a single AI isn’t handling the request. A coordinated team of agents is.
⏵ The brands that win agentic commerce aren’t just building better data — they’re building a receiving architecture designed to serve visiting agents well.
⏵ This infrastructure exists today. The question is whether your brand is designing for it — or waiting until it’s table stakes.
In Part 2 we talked about the handshake — the llms.txt file that tells a visiting agent what your brand offers and how to navigate it.
Now the agent is inside. What happens next?
It’s Not One AI. It’s a Team.
Most people imagine agentic commerce as a single AI having a conversation with your website. That’s not how it works.
What actually happens is closer to a well-run organization. Different agents handle different jobs. Each one specialized. Each one passing context to the next. The visiting agent never talks to your database directly — it talks to a receiving architecture your brand has designed to serve it.
The Receiving Stack
The Orchestrator is the front door — a Chief of Staff that greets the visiting agent, interprets intent, and decides who handles what. It doesn’t do the work. It directs it.
The Context-Intake Agent grounds the request against your brand’s data environment — translating intent into something your catalog, inventory, and pricing systems can respond to. This is where your llms.txt and data taxonomy do their heaviest lifting.
The Retrieval Agent surfaces the two or three closest matches from your structured data. Not everything you sell — a curated shortlist built from real data.
The Reasoning Agent evaluates fit — scoring options against budget, availability, specifications, and proximity — and synthesizes a recommendation.
The Orchestrator closes the loop — packaging the recommendation and handing it back to the visiting agent.
The visiting agent never knew it was talking to five agents. It just got a clean answer.
Why This Matters
Every layer depends on your underlying data. Fragmented, ungoverned data doesn’t fail loudly. It fails quietly — incomplete recommendations, lost transactions, visiting agents that don’t come back.
What’s missing isn’t technology. It’s organizational intent — designing the receiving architecture before the agents arrive.
One file. One stack. Built once. Ready for whatever agent shows up.
This is Part 3 of a 5-part series on agentic data readiness. Part 4 is the readiness audit — how to test your own site with AI today and find every gap before the agents do.
Have you started thinking about your receiving architecture yet?
The Difference Between Being Cited and Being Shoppable
Special Series: Agentic Shopping (2 of 5)
Key Takeaways:
⏵ Most marketers think llms.txt is a GEO tool. It's also the front door for agentic shoppers. Same file. Two very different stakes.
⏵ Two readiness layers: being cited by AI, and being transactable by AI. The infrastructure that serves both is the same.
⏵ Build a machine-readable interface now. Rebuilding under pressure later costs more.
In Part 1 - I asked whether your data was ready to be shopped by an agent.
Before an agent can shop you — it has to know you exist. And know what you can do. That's the handshake.
Two Problems. One File.
Most marketing leaders think of llms.txt as a GEO tool — the AI equivalent of SEO metadata. A way to help ChatGPT, Claude, or Perplexity cite your brand when someone asks a relevant question.
That's real. And it matters. AI-generated answers are already cannibalizing search clicks. If your brand isn't structured for LLM citability, you're losing visibility you can't measure yet.
But llms.txt is also something else entirely. It's the file an agentic shopper reads before it ever interacts with your site.
Two very different problems. Two very different stakes. One file — if you build it right.
The Difference Between Being Cited and Being Shoppable
GEO is a visibility problem. When someone asks "what's the best midsize SUV for a family of five," you want your brand in the answer.
Agentic readiness is a transaction problem. When an agent arrives at your site with configured intent — specific product, budget, availability window — it isn't reading your content. It's querying your data. It needs to know what's available, how it's structured, and whether it can transact on behalf of the human who sent it.
---
Being cited gets you into the consideration set. Being shoppable gets you the transaction.
---
Most brands aren't ready for either. The ones investing in GEO are solving half the problem.
The Markdown Principle
Humans organize content for humans. Machines need content organized for machines. The gap between those two structures is where AI fails — whether it's a knowledge base, a product catalog, or a brand's entire digital presence.
llms.txt is essentially a markdown file. Structured, consistent, machine-consumable. It tells a visiting agent what your brand offers, how it's organized, and what actions are available — whether that agent is retrieving a citation or executing a purchase.
The brands that win the agentic era aren't the ones with the most sophisticated AI. They're the ones whose data is structured for machine consumption before the agents arrive.
One file. Two jobs. Build it once.
Agentic Data Readiness: The Infrastructure Question Nobody is Asking Yet
Special Series: Agentic Shopping (1 of 5)
Key Takeaways:
⏵Agentic shoppers don't experience your website — they parse your data. If it isn't structured for machine consumption, they move on.
⏵The brands winning agentic commerce won't have the best AI. They'll have the best data infrastructure.
⏵The readiness question isn't "are we building agents?" It's "can agents shop us?"
Your next customer might not be human.
They won't browse your homepage. They won't respond to your hero banner. They won't notice your redesign.
They'll arrive with intent already formed, query your product data, evaluate fit, and either complete a transaction — or leave. In seconds. Without rendering a single pixel of your UI.
The Agentic Shopper Is Already Here
Mondelez is hiring a global lead for agentic commerce. Their own retail partners project 30% of site traffic will be agentic by 2028. McKinsey puts the global market at $3–5 trillion by 2030.
The infrastructure is being built right now. Google. Amazon. OpenAI. The question isn't whether agentic shoppers are coming to your brand's digital properties. The question is whether your data is ready to receive them.
The AI Readiness Self-Score: Agentic Data Readiness
🔴 Product catalog, pricing, and inventory data is fragmented — an agent couldn't navigate it reliably.
🟡 Some structured data exists but it's inconsistent across SKUs, categories, or regions — partial information at best.
🟠 Core product data is structured, but the connective tissue — policies, promotions, inventory logic — isn't machine-readable.
🟢 Data is structured, governed, and accessible enough that a non-human agent could navigate, evaluate, and transact reliably.
Most brands land at 🔴 or 🟡 — not because of bad decisions, but because no one has asked the question in these terms before.
The Uncomfortable Truth
Agentic commerce doesn't fail because of bad AI. It fails because the data is locked inside analyst queues, fragmented across platforms, and structured for human dashboards — not machine consumption. An agent hits the same walls your internal team hits every day. The difference: your team works around it. The agent doesn't.
The next wave of lost conversions won't show up in your bounce rate. Agents don't bounce. They just never come back.
This is Part 1 of a 5-part series on agentic shopping and data readiness. What's your gut reaction to the idea that agents will be shopping your brand before your team is ready for them?
Google's AI Max Doesn't Break Campaigns. It Exposes Them.
Key Takeaways:
► Google AI Max shifted search from keyword-based to intent-based bidding this week — your AI now interprets meaning, not just match types
► Intent-based systems expose data taxonomy gaps across teams and partners that keyword campaigns could quietly absorb
► Organizations without a governed taxonomy don't lose control gradually — they lose it at the moment the AI picks a winner
Google AI Max just went live. Your search campaigns are no longer keyword-based. They're intent-based.
That distinction matters more than most teams realize — and it has nothing to do with the bidding mechanics.
The Problem No One Is Naming
Google's AI now reads your landing pages, interprets your headlines, and infers the intent behind your audience's queries in real time. It connects signals across placements your team never explicitly mapped. By September, Dynamic Search Ads are gone. The system decides.
Here's the gap: intent-based execution is only as coherent as the data structure feeding it.
When your media team defines a conversion differently than your brand team — when your agency uses one attribution window and your internal analytics team uses another — keyword campaigns could absorb that ambiguity. They operated inside guardrails you set.
Intent-based AI doesn't split the difference. It resolves ambiguity by picking one interpretation. And it does that at scale, continuously, before anyone reviews the output.
The AI Readiness Self-Score:
🔴 No shared taxonomy — everyone defines KPIs differently.
🟡 Informally aligned — rough consistency, but nothing enforced.
🟠 Mostly governed — taxonomy exists, but legacy exceptions still exist.
🟢 Fully governed — one definition, enforced across teams, partners, and platforms.
Most organizations are 🟡. Many believe they're 🟠.
The Real Cost:
On The AI Readiness Maturity Spectrum, data taxonomy isn't a technical problem. It's a governance decision that predates your AI investment.
Google's AI Max will optimize against whatever signal you give it. If your conversion definitions conflict across your stack, you won't see a configuration error. You'll see a performance report that looks reasonable — until someone asks why the agency numbers don't match the internal dashboard.
That gap used to be a reporting problem. With intent-based AI running your campaigns, it becomes a budget allocation problem.
---
The diagnostic no platform runs before selling you automation is whether your organization is ready to be automated.
---
Source: Digiday, April 15, 2026 https://lnkd.in/gMNpRNAF
What Would Your AI Say About You?
What would your AI say about you?
Not what you say you can do with AI.
What it would actually say — based on how you work with it every day.
Most people can't answer that question cleanly. Neither can their managers.
That's the talent gap no one is talking about yet.
► The Problem No Performance Review Has Caught Up To ◄
Every organization is building AI into daily workflows. Tools are deployed. Training is checked off. Adoption metrics look fine.
But no one has defined what good looks like at the individual level.
You can see someone's outputs. You can measure their deliverables. You cannot see whether they're working with AI with discipline and judgment — or just accepting whatever the model hands back.
That invisible variable is about to become the most important one on your team.
What does your organization actually know about how well your team works with AI — not just whether they're using it?
🔴 No way to assess it — AI skill is assumed or inferred from tool usage, not evaluated
🟡 Managers have a general sense of who's good with AI but no framework to assess or develop it
🟠 Some informal recognition of AI capability differences across the team but nothing formal or repeatable
🟢 AI competency is defined, assessed, and incorporated into development goals and performance reviews
Most organizations are 🔴 or 🟡. Which means the team members doing the most sophisticated AI work are invisible to the talent system. And the ones producing mediocre AI outputs are invisible too — for a different reason.
What This Actually Costs You
You can't develop what you can't see.
If AI competency isn't defined, it can't be coached. It can't be rewarded. It can't be set as a goal. It can't be tracked year-over-year. And it can't be used to make better hiring decisions.
The organizations that get this right in the next 18 months will build compounding advantage — because their people will get measurably better at AI, not just more familiar with it. The ones that don't will wonder why their AI investments keep underdelivering despite widespread adoption.
---
The gap between using AI and using it well is real. And right now, there's no organizational system to measure it.
---
Agentic AI Doesn't Fix Your Data Problems. It Industrializes Them.
Key Takeaways
● The push to automate workflows and deploy agents is running directly into a data foundation nobody has audited.
● An agent doesn't pause when it hits bad data. It optimizes through it — at scale, without flagging the problem.
● Data quality monitoring is the reason 88% of agentic AI deployments never reach production.
There's a mandate moving through your organization right now.
Automate the workflows. Deploy the agents. Move faster.
Underneath that mandate — largely unspoken — is a data foundation nobody has fully audited.
The Setup
79% of enterprises have adopted AI agents. Only 11% are running them in production. That gap isn't about model capability or budget.
88% of AI agents fail to reach production. The organizations that succeed share one attribute more than any other: pre-deployment infrastructure investment. Most marketing organizations are skipping it.
The AI Readiness Self-Score:
🔴 Reactive — Quality issues surface when something looks wrong. No monitoring exists.
🟡 Informal — Someone usually catches issues. No formal process, no defined ownership.
🟠 Partial — Some monitoring exists. Coverage is incomplete across teams and partners.
🟢 Governed — Automated alerts, defined ownership by source, rapid resolution before issues reach decisions.
Why This Matters Now
In the human-in-the-loop era, bad data meant a slow decision or a wrong report. A human eventually caught it. In the agentic era, if a data pipeline drifts, an agent doesn't report the wrong number. It takes the wrong action. Confidently. At scale. That's not a technology problem. It's a data ownership problem the automation mandate just made urgent.
The Fix
I've designed a framework — the Blueprint Studio — specifically to address this. Before you automate any marketing process, answer three questions:
1) Who owns each data source? Not a team. A named individual accountable for freshness and resolution.
2) What does failure look like — and how fast will you know?
3) What happens to the workflow when a source fails?
If you can't answer all three in five minutes, you have an unmonitored single point of failure inside your automation. Agentic AI doesn't fix your data problems. It industrializes them.
#AIReadiness #DataStrategy #MarketingOperations
Stop Pitching the Demo. Start Shipping the Spec.
The gap between AI vision and enterprise deployment isn't technical. It's a packaging problem.
I help teams architect agentic workflows. We build a working prototype — often in a Claude Project — complete with a step map, design canvas, skills written in markdown, and a sharp system prompt. It feels real. It works.
But the real test comes at handover.
Every time, the engineering team asks the same question: "Is this a Claude thing… or can we actually ship this in production?"
That question is my signal. It means I haven't packaged the work clearly enough for them to run with it.
The reframe that changed my approach: Stop pitching the demo. Start shipping the spec. The prototype is just validation. The actual deliverable is documentation that speaks for itself when I'm no longer in the room.
Here's what effective packaging looks like:
1. Package the capability, not the platform. Deliver a clear step map, design canvas, defined skills in markdown, inputs, outputs, and success criteria. The implementation technology is engineering's decision — not yours.
2. Write the expectations brief. One concise document: here's what needs to be deployed, here's how we'll measure success, here's what I'm deliberately not prescribing. Then hand it over.
3. Step back. The moment you stay attached to the tool that helped you prototype, momentum dies.
The uncomfortable truth: most agentic workflows stall in enterprise not because of technical limitations or organizational politics — but because the visionary never fully translated the vision into something an engineering team could own.
Where are you in this right now — still in demo mode, or do you have a handover process that engineering actually trusts?
The Invisible Workflow Problem
Your best AI results aren't repeatable. They're dependent on whoever ran the workflow last.
Think about the last time your team got a genuinely great output from AI — a campaign diagnosis that was actually useful, a performance summary that led to a real decision, a creative brief that didn't need three rounds of revision.
Now ask: could anyone else on the team reproduce that result tomorrow? Could you reproduce it next week?
In most organizations, the answer is no. Not because the tool changed. Because the workflow lived in one person's head — their prompt approach, their context setup, their personal Claude account, their browser bookmark. Invisible to the organization. Impossible to scale. Gone when they're gone.
Quick self-score:
🔴 No documentation. Every workflow starts from scratch depending on who's running it.
🟡 A few people get consistently great results — but their system lives in their personal accounts and their own head. The org has no access to it.
🟠 Some team-level standards exist in some places. Nothing you'd call a system.
🟢 AI workflows are documented, version-controlled, and fully transferable — any team member can execute to the same standard.
Here's the uncomfortable math:
Jasper's 2026 State of AI in Marketing — 1,400 marketers surveyed — found that while 91% now use AI in their work, the share who can prove ROI actually dropped year over year. From 49% last year to 41% today.
Not because AI got worse. Because personal productivity isn't organizational capability. Leadership isn't seeing the value because the value isn't in the system. It's in the individual.
The gap between 🟡yellow and 🟢green isn't a better tool or a smarter prompt. It's a decision to treat workflow documentation as seriously as the output you're trying to scale.
Your AI strategy is only as durable as the least-documented workflow it depends on.
London Built It. Chicago Automated It. Singapore Is Testing It.
Your marketing organization is running AI experiments right now.
Someone in London built a workflow. Someone in Chicago automated a brief. Someone in Singapore is testing a content tool nobody approved.
None of it is connected. None of it is governed. And none of it is visible to the people responsible for what happens when something goes wrong.
That's not a technology problem. It's an operating model problem.
Quick self-score:
🔴 No process — AI tools are adopted informally, market by market, team by team, with no central visibility.
🟡 Some awareness — leadership knows AI is spreading but there's no intake process, no registry, no governance funnel.
🟠 Partial governance — approved tools exist but no structured process for evaluating new ideas before they get built.
🟢 Full funnel — every AI idea enters through a defined intake, gets evaluated against governance criteria, and earns deployment through human-owned decision gates.
Most enterprise marketing organizations are 🔴 or 🟡. Not because they don't care about governance. Because no one designed the system before the tools arrived.
Here's what a governed AI innovation funnel actually requires:
→ A universal intake — one mandatory entry point for every AI idea, every market, every team. If it isn't logged, it doesn't exist.
→ Structured evaluation — ideas scored against business impact, data compliance, scalability, risk, and execution readiness before anyone builds anything.
→ Human decision gates — AI does the processing. Humans own every approval. Nothing auto-promotes. Nothing auto-deploys.
→ A live registry — a running record of every idea, every evaluation, every deployment decision. Not a spreadsheet someone updates quarterly. A governance-grade audit trail.
The organizations that get this right aren't slowing innovation down.
They're making every idea better, faster, and more likely to actually deploy at scale — instead of dying in a market silo nobody else can find.
The uncomfortable reality:
If your AI governance policy lives in a legal document but your teams are building without a funnel, you don't have governance. You have liability with paperwork on top of it.
Nobody Owns the Gap
🔥The governance trap nobody is talking about.
Your organization didn't decide to move slowly on AI. It just never decided who owns the space between approved systems.
Here's the scenario playing out in marketing organizations right now:
The AI tool is approved. The data platform is approved. Legal signed off on both. IT provisioned both. And the two systems cannot talk to each other — because connecting them requires a decision that sits in no one's job description.
Who owns the MCP integrations? Who approves the data flow between a secured data platform and an approved AI layer? Who unblocks the connection between your media data warehouse and the AI skill that's supposed to analyze it?
Not who could answer that question. Who actually owns it — with the authority and accountability to move it forward?
In most organizations, that answer is silence.
Quick self-score:
🔴 No one owns it — approved tools sit unconnected. Capability exists on paper only.
🟡 IT and Marketing pass it back and forth. Nothing moves without an escalation.
🟠 Ownership is assumed but undocumented — progress depends on who pushes hardest.
🟢 A named function owns platform connectivity decisions with a defined process and SLA.
Here's the uncomfortable reality:
The governance trap isn't rogue AI. It's two fully approved, fully secured platforms sitting three feet apart with no one authorized to connect them.
Organizations aren't failing at AI adoption because they're reckless. They're failing because accountability for enablement — the unglamorous work of actually connecting approved capability to approved data — belongs to no one.
You don't need looser governance. You need someone whose job it is to own the gap.
70% Say It's a Priority. 17% Are Actually Doing It.
Here's a question most marketing leaders can't answer cleanly.
"If you could solve one AI or data challenge in the next 90 days — the one that would have the most meaningful impact on your marketing effectiveness — what would it be?"
Not a wishlist. One thing.
Most organizations can't answer that. Not because they don't care about AI. Because they've never been forced to prioritize.
Here's what the data says: 70% of marketing leaders say optimizing spend is a top priority. Only 17% are actually using AI to analyze and optimize campaigns. That gap — between what leaders say matters and what AI is actually doing — is the diagnostic. And the 90-day question is the forcing function that reveals it.
The exercise:
Ask this question in your next leadership meeting. Write down the answers before anyone speaks.
What you'll usually find:
→ Finance names a measurement problem
→ Operations names a data quality problem
→ The CMO names a tool problem
→ The agency names a brief problem
Four different answers. One shared budget. No shared priority.
That's not an AI readiness problem. That's a leadership alignment problem wearing an AI mask.
The organizations actually moving are the ones that have answered this question — and gotten the whole room to the same answer. Not because the problem is small. Because they made a choice.
What would your one thing be?
Drop it in the comments. Seriously. I read every one. And if you can't name it in one sentence — that's the answer.
Source: Supermetrics, Marketing Data Report 2026 — survey of 400+ marketers across the US, UK, Germany, Australia, and Singapore.
* 70% of marketing leaders cite optimizing spend as a short-term priority.
* Only 17% are actually using AI to analyze and optimize campaigns.
That gap is the whole conversation.
https://lnkd.in/g6d9Spka
We're Watching It' Is Not a Media Strategy
ChatGPT is now a paid media channel. Does your team have a buying thesis for it — or are you waiting to be asked?
On February 9, OpenAI launched ads inside ChatGPT. On March 2, Criteo became the first ad-tech partner. The Trade Desk is reportedly in talks to follow. Early data shows users referred through ChatGPT convert at 1.5x the rate of other channels. Target, Ford, Best Buy, and AT&T are already in.
This isn't a future planning item. It's live inventory.
Sources: Criteo press release, March 2, 2026; Digiday, "OpenAI is building the ad tech stack it's currently borrowing," March 10, 2026.
The question most marketing teams are quietly avoiding: do we have a framework for evaluating this — or are we going to end up reacting when a client asks?
Quick self-score:
🔴 No process. Channel decisions happen informally based on whoever raises their hand first.
🟡 We discuss new channels but there's no documented evaluation criteria or decision framework.
🟠 Some guidance exists but it isn't consistently applied — especially for AI-native channels with no measurement history yet.
🟢 Deliberate, documented framework governs every new channel decision: measurement approach, brand safety criteria, attribution model, minimum test budget threshold.
Here's the uncomfortable reality:
Most media organizations are excellent at executing on known channels. The process for evaluating genuinely new surfaces — especially ones with no historical benchmarks, unproven attribution models, and evolving ad formats — tends to be whatever a smart person argues convincingly in a room.
That's not a framework. That's a persuasion contest.
ChatGPT advertising isn't the point. The point is whether your team has the organizational muscle to evaluate it intentionally — with defined criteria — rather than reactively when budget pressure or client curiosity forces the question.
The organizations that will have an advantage here aren't the ones that jump in fastest. They're the ones that already have a decision architecture for situations exactly like this.
Your team is going to get asked about ChatGPT ads in the next 90 days. "We're watching it" is not a media strategy.
The Real AI Risk Isn't Rogue Agents. It's Ungoverned Skills.
POST #2 of 2 — Visibility & Governance Series
Anthropic recently told the enterprise world something important. Most marketing leaders missed it and we are now seeing momentum.
The framing was technical: don't build agents, build skills instead.
The implication for marketing organizations is anything but.
An AI agent is broad — a general-purpose system you point at a problem. A skill is different. Packaged instructions, domain knowledge, and procedural logic an AI loads on demand to perform a specific task, consistently, every time. Skills are how you take what your best analyst does intuitively — campaign post-mortems, pacing diagnosis, data reconciliation — and make it reproducible at scale. The promise is real. Package your best thinking into a skill and your entire team executes at that level.
Here's the risk nobody is talking about yet.
Custom AI without governance creates shadow AI — siloed, unmanageable, impossible to control. Skills deployed without oversight create the same problem. Just faster and at greater scale.
What ungoverned skill proliferation actually looks like:
→ An analyst builds a campaign diagnosis skill using a methodology leadership never approved
→ Another builds audience segmentation using KPI definitions that don't match the agency's
→ A third builds a reporting skill that's been running against stale data for six weeks
Each works. Each produces confident, fluent output. Each compounds your governance gap — now automated.
Anthropic built the answer into their platform: central admin controls that govern which skills are provisioned and enabled. The infrastructure exists. Most marketing organizations haven't built the muscle to use it.
Quick self-score:
🔴 Skills built by individuals. No inventory, no standards, no oversight.
🟡 Some workflows exist. Nobody has mapped what's running or who owns it.
🟠 Leadership has general awareness. No formal governance framework.
🟢 Skills are inventoried, owned, governed, and aligned to documented standards.
The organizations that win won't deploy the most skills. They'll govern them.
Ungoverned skills don't amplify your best thinking. They amplify whoever built them last.
Where does your organization land? Drop your color in the comments.
https://lnkd.in/gkUFJeXi (Anthropic, "Equipping agents for the real world with Agent Skills")
https://lnkd.in/gj7PuZ3j (Anthropic launches new push for enterprise agents with plug-ins for finance, engineering, and design)
You Can't Govern What You Can't See
POST 1 of 2 — Visibility & Governance Series
Your leadership team has no idea how AI is actually being used across your marketing organization. Not a rough sense. Not a partial picture. No idea.
Think about the last time someone asked your CMO to inventory every AI-enabled workflow running across media, analytics, content, and data teams. How long would it take to get a complete answer? Who would even own that answer? That silence — right there — is your governance gap.
Most marketing organizations have invested in tools, trained teams, and started workflows. What almost none have done is build organizational visibility into what's actually running, who owns it, and whether it's aligned to business priorities.
AI is being used across your org right now. Some of it is brilliant. Some of it is inconsistent. Some of it is producing outputs that are informing real decisions — audience strategy, budget allocation, campaign analysis — and nobody in leadership sanctioned the methodology, reviewed the output quality, or could reproduce the result tomorrow.
Quick self-score:
🔴 No visibility. AI is happening. Nobody knows where.
🟡 Leadership has a rough sense. No formal tracking. No governance.
🟠 Some usage is monitored. Significant gaps remain.
🟢 Full visibility. Governance policies aligned to business strategy and risk standards.
Here's the uncomfortable math: most organizations are 🔴 or 🟡. And most leadership teams don't know it — because the people running AI workflows don't broadcast what they're doing, and leadership never created a mechanism to ask.
You can't govern what you can't see. And right now, most marketing leaders are making assumptions about their AI maturity based on what they've approved — not what's actually running.
This is Part 1 of 2. Next post: why ungoverned AI skills — not rogue agents — are the real enterprise risk hiding in plain sight.
You Don't Have an AI Strategy. You Have an AI Layer.
Your AI strategy looks impressive in the presentation.
But here's the one question that cuts through it: Has the manual grind actually decreased for the average analyst on your team?
Not your best prompt engineer. Not the one person who's figured out how to automate their own workflow. The average person. The one pulling data from three platforms into a spreadsheet every Monday morning before any real work can begin.
If the answer is no — you don't have an AI strategy. You have an AI layer sitting on top of an unchanged operation.
Quick self-score:
🔴 Minimal — most marketing and data work is still done manually.
🟡 Some reduction in isolated areas but the broader team is largely manual.
🟠 Noticeable reduction across several marketing and data workflows.
🟢 Significant — AI has fundamentally changed how our team processes information and produces work.
Here's what most organizations get wrong:
They measure AI adoption by tool usage. Seats purchased. Prompts run. Hours of training delivered. None of that tells you whether the manual grind has moved.
The honest test is simpler: ask your analytics team how much of their week is still spent preparing data before analysis can begin. Ask your media team how long a standard performance report takes to produce. Ask your agency how many hours go into reconciling numbers before they can tell you what happened last week.
Those answers will tell you more about your AI readiness than any dashboard showing tool adoption rates. The organizations pulling ahead aren't adding AI on top of manual processes. They're replacing the manual processes — and measuring whether it actually happened.
If your team is still cleaning data by hand, you don't have an AI strategy. You have a very expensive habit.
Confident Doesn't Mean Correct
NERD ALERT - This final post takes a peek under the hood of AI.
Grounding: AI will give you a wrong answer with the same tone as a right one. That's not a bug — it's the architecture. And it's the most important thing marketing leaders need to understand before they deploy AI at scale.
It's called the grounding problem. And it's why AI confidence and AI accuracy are not the same thing.
A language model generates responses by predicting what words should follow, given everything in its context. When it's well-grounded — meaning it's working from accurate, relevant, verified information — the output is reliable. When it isn't, the model doesn't stall or flag uncertainty. It continues. Fluently. Confidently. Incorrectly.
This is what practitioners call hallucination. But that word undersells the problem. Hallucination implies something obviously wrong. A made-up statistic. A fictional citation. Those are easy to catch.
The harder version is a response that is directionally plausible, internally consistent, and subtly wrong in ways that require domain expertise to detect. That's the one that gets published in the recap deck.
What grounding actually requires:
Grounding is the practice of anchoring AI output to verified sources before generation happens. It's not a setting you toggle. It's an architectural decision.
→ Retrieval systems that pull from your actual data, not the model's training
→ Source attribution so outputs can be traced and audited
→ Evaluation layers that check outputs against known standards before they surface
→ Human review workflows designed around the failure modes, not the happy path
This is the common thread of everything I've written in this series. Decomposition structures the question. Fan-out broadens the retrieval. Context windows supply the inputs. Retrieval vs. reasoning diagnoses the failure. Grounding is what ties it together — the discipline of ensuring AI output is anchored to something true.
The uncomfortable truth for marketing organizations:
Most AI deployments optimize for speed of output. Grounding optimizes for trustworthiness of output. Those are not the same objective, and the tension between them is where most enterprise AI implementations quietly fail.
Speed without grounding is just confident noise, faster.
The organizations that will build durable AI capability aren't the ones moving fastest. They're the ones who decided early that trustworthy output was non-negotiable — and built accordingly.
It's Not the Model. (Maybe.)
NERD ALERT - This post takes a peek under the hood of AI.
Retrieval vs. Reasoning: AI makes two fundamentally different kinds of mistakes. Most organizations can't tell them apart. That's expensive.
Retrieval: The first is a retrieval failure. The model didn't have the right information. It worked from what it had — which wasn't enough, or wasn't current, or wasn't specific to your business. The answer was wrong because the inputs were wrong.
Reasoning: The second is a reasoning failure. The model had the information. It connected it incorrectly. The logic was flawed, the inference was a stretch, or it weighted the wrong signals.
These look identical on the surface. Confident. Fluent. Wrong.
But they have completely different fixes.
Retrieval failures are solved with better data architecture — richer context, better retrieval systems, more relevant information surfaced before the model responds. This is an infrastructure problem.
Reasoning failures are solved with better prompt design, chain-of-thought instruction, and output evaluation. This is a workflow problem.
If you treat a reasoning failure like a retrieval problem, you'll rebuild your data pipeline and still get bad answers. If you treat a retrieval failure like a reasoning problem, you'll rewrite your prompts indefinitely and wonder why nothing improves.
What this means for marketing organizations:
When AI-powered campaign analysis produces a bad output, the instinct is usually to blame the model. "It just doesn't understand our business."
Sometimes that's true. But before you conclude the model can't reason about your data, ask:
→ Did it have access to the right data in the first place?
→ Was the question structured to guide its reasoning, or left open-ended?
→ Was the output evaluated against a known standard, or just eyeballed?
Diagnosing which failure mode you're in is the work. It's not glamorous. It doesn't make for a good demo. But it's the difference between AI that compounds in value over time and AI that stays permanently in pilot.
Confident and fluent is not the same as correct.
Knowing which kind of wrong you're dealing with is the first step to fixing it.
The Model Isn't the Variable. You Are.
NERD ALERT - This post takes a peek under the hood of AI.
Most people think the quality of an AI answer depends on how good the model is. When in actuality, it depends more on what you put in front of it.
This is the context window.
And it's the most underestimated variable in enterprise AI.
Here's the simple version: an AI model doesn't have memory in the way you do. Every time it responds, it works from what's currently in its context window — the information you've handed it for that interaction. Documents, data, instructions, conversation history, retrieved chunks from your database. All of it.
The model is only as good as what's in that window.
This is why two organizations can use the exact same AI model and get dramatically different results. One handed it rich, structured, relevant context. The other handed it a vague question and hoped for the best.
What this means practically:
Context window management is a design discipline. Not a prompt trick.
For marketing organizations building AI into analytics, campaign reporting, or content workflows, the question isn't just "what do we ask the AI?" It's:
→ What information does it need to answer well?
→ Where does that information live?
→ How do we get the right context into the window before the model responds?
This is exactly what retrieval systems are designed to solve — pulling the relevant information from your data and surfacing it into context before generation happens. Fan-out (which I wrote about in my last post) is one technique for doing that retrieval more thoroughly.
But retrieval is only half the problem.
The other half is what you put in by design — your instructions, your definitions, your business rules, your standards. That's not retrieved. That's architected.
Organizations that treat context as an afterthought get AI that performs like a smart intern on their first day. No institutional knowledge. No standards. Just capability without direction.
Organizations that design their context get AI that performs like a senior analyst who has read everything, remembers everything, and applies your standards consistently.
Same model. Completely different outcome.
The question isn't whether your AI is powerful enough.
It's whether you've given it enough to work with.