The Inventory Challenge for Marketing: Knowing Which AI Agents Work in 2026

We’ve been having the wrong AI debate for months

A few weeks ago I wrote in Tagesspiegel Background that Germany is having the wrong AI debate. We talk about models, sovereignty, and platform ecosystems. We talk far too little about what actually determines competitiveness: How do we translate existing AI into productivity, value creation, and competitive edge?

This view didn’t come from a strategy paper. It came from many conversations where CMOs and marketing leaders say almost the same thing. One line stands out. It sounds friendly, optimistic, sometimes proud. And it’s dangerous.

The line is: "We have people on the team doing great things with AI."

I hear that sentence so often that I read it as a symptom. Behind it is almost always the same picture. A few individuals tinker in private. They try ChatGPT for briefings, HubSpot AI for campaign sketches, Midjourney for visuals, custom GPTs for recurring tasks. Sometimes the results are impressive. Sometimes they even circulate within the team. What rarely exists is a central overview of what’s running, who runs it, which data it uses, and which workflows it’s embedded in.

That gap is the marketing topic in 2026. Not tool choice. Not the model. The question is whether you know your inventory.

IBM says: 1,600 agents per enterprise by end of 2026

At IBM Think 2026 in early May, a number came up that deserves a moment’s attention. IBM forecasts an average of more than 1,600 AI agents per large enterprise by the end of 2026. Not in five years. In the next seven months.

The same report offers a second number that makes the first one worrying. Today, only 18% of organizations have a complete agent inventory. 12% run a central sprawl-management platform. 70% of executives say their governance is "not fit for purpose."

Three numbers, one conclusion: Agents are arriving faster than the understanding of who operates them, what they’re responsible for, and who owns their output.

• 1,600 AI agents – Average per enterprise by end of 2026
• 18% inventory coverage – Organizations with a complete agent registry
• 12% sprawl management – Central platform for agent operations
• 70% governance inadequate – Executives consider policies insufficient

This gap isn’t only an IBM issue. In May, McKinsey noted in "Reinventing marketing workflows with agentic AI" that two-thirds of today’s marketing activities are agentically automatable. Nearly 90% of CMOs are experimenting; fewer than 10% have an end-to-end workflow in production. Salesforce renamed its Marketing Cloud to Agentforce Marketing in April. HubSpot positions itself since the Spring Spotlight as an "agentic customer platform," offering Customer Agent, Prospecting Agent, and Data Agent as platform-native building blocks, not add-ons.

Translation: You no longer need to buy agents. They arrive with the next license renewal. The open question is no longer whether you use agents. The open question is whether you still know which ones you have.

Why “great things” don’t scale

When someone on the team builds a clever custom GPT for briefings, that’s a positive moment. It’s not the start of an agentic organization. It’s the start of an inventory problem.

Three effects appear once these individual DIY solutions accumulate:

First, the stack multiplies. In April 2026 the Cloud Security Alliance showed that 82% of companies discovered at least one agent or workflow in the past year that IT or Security didn’t know about. Marketing is often the first home for these shadow agents because text, images, and email have the lowest barriers to entry.

Second, responsibility moves downward. The person who built the agent becomes the quiet owner. If that person leaves, nobody knows how the output is produced, where the data comes from, or who validates it. The result: workflows no one understands — and few want to shut down.

Third, in August 2026 the EU AI Act supervision begins. Article 4 requires companies to empower their employees with adequate AI competence. The Bundesnetzagentur will act as the central authority. Bitkom estimates compliance costs for a mid-sized company with 500 employees at €80,000–€250,000 initial plus €30,000–€70,000 ongoing per year. Compliance can only check what’s on a list. No list equals no discussion about conformity — only a discussion about penalties.

1. Stack explosion and shadow agents
   The Cloud Security Alliance showed in April 2026 that 82% of companies discovered at least one agent or workflow unknown to IT or Security. Marketing is often the first home for these shadow agents because text, image, and email have low entry barriers.
2. Responsibility shifts down and knowledge loss
   The builder becomes the silent owner. When they leave, transparency over output, data sources, and validation disappears. Workflows remain in place — but few understand them.
3. EU AI Act and compliance risk
   From August 2026 supervision begins; Article 4 requires adequate AI competence for all employees using AI on the company’s behalf. Without an inventory, compliance can’t verify. Bitkom estimates costs at €80,000–€250,000 initial and €30,000–€70,000 annually.

So an individual’s “great thing” becomes costly in three ways if it’s not integrated into organizational structure: stack explosion, knowledge loss, and compliance risk.

What we do at faive

At faive we talk a lot about agentic organizations. We also have to be one. Our own stack consists of several components we deliberately documented.

Cowork is our central working layer, where skills, MCPs, and plugins run the content pipeline, sales briefings, and internal reporting. ClickUp agents handle tracking sprint tasks, meeting outcomes, and project status. Custom MCPs connect HubSpot, Gmail, Calendar, and our cloud drive. The Content OS coordinates strategy, drafts, pipeline, and performance across multiple skills.

The list itself isn’t the remarkable part. The remarkable part is that the list exists. Each component has an owner, a documented purpose, and a validation rule. We know what the agent can and cannot do, and who steps in if it overreaches.

What we do at faive

We apply the principle to ourselves: our stack consists of deliberately documented components, not loose individual tools.

Cowork is the central working layer with skills, MCPs, and plugins for the content pipeline, sales briefings, and internal reporting. ClickUp agents track sprint tasks, meeting outcomes, and project status; custom MCPs connect HubSpot, Gmail, Calendar, and the cloud drive. The Content OS coordinates strategy, drafts, pipeline, and performance across multiple skills.

Each component has an owner, a documented purpose, and a validation rule. We know what an agent may do, what it may not do, and who intervenes if it does.

It sounds basic. It’s the discipline that separates 12 isolated agents from a marketing organization that accelerates its pipeline. Salesforce’s Connectivity Benchmark 2026 documented that organizations use on average 12 AI agents and that 50% operate in isolated silos. That’s not agentic. That’s parallel automation.

Three steps to an inventory

What you can do this week isn’t a large program. It’s three steps.

First, list which AI tools and custom applications are actually in productive use on your team. Not officially — actually. Ask in your daily or weekly. Put it in a shared spreadsheet. You’ll be surprised how many entries appear.

Second, assign an owner to each entry. One person is responsible for documenting the tool, maintaining it, and reacting if it produces problematic results.

Third, write a validation rule for each tool. What may the agent do without approval? When does a human step in? Who is the escalation point? Yale published a helpful classification in May: direct, mediated, background. An agent in a customer conversation needs real-time validation. An agent drafting a newsletter needs a two-person review. An agent reconciling data overnight needs drift monitoring. Three classes, three rules.

1. Capture the inventory
   List all AI tools and custom applications actually in use — not just officially licensed ones. Ask in daily or weekly meetings and collect entries in a shared spreadsheet.
2. Assign ownership
   Give each entry a responsible person who ensures documentation and maintenance. This person also responds if results are off.
3. Define validation rules
   Specify what an agent may do without approval, when a human intervenes, and who escalates. Use Yale’s direct, mediated, background schema with examples from your workflow.

If you take these three steps, you won’t yet have an agentic marketing organization. You will have the precondition for one.

FAQ

What distinguishes an AI inventory from a tool list?

A tool list records licenses. An AI inventory records which autonomous actions a system can take on your behalf, with which data, and under which oversight. License management is a procurement issue. Inventory is an architectural issue.

Why isn’t banning custom GPTs enough?

The Cloud Security Alliance found that companies that ban unsanctioned tools do not reduce shadow-AI use; they push it into private channels. Employees resort to personal accounts on personal devices. The solution is not prohibition but substitution: a sanctioned stack that’s good enough so no one needs a private tool.

How does this align with the EU AI Act?

Article 4 requires from August 2, 2026, adequate AI competence for employees using AI on the company’s behalf. Without an inventory you cannot meet that obligation, because you don’t know which tools need documented training. Inventory is the inexpensive compliance precondition, not the expensive compliance reaction.

Who should own the inventory — IT or Marketing?

That’s the wrong question. The inventory belongs to the organizational layer operating agentic workflows. In marketing that’s often Marketing Operations or Marketing Excellence. In smaller teams it’s the CMO. IT validates security, but it does not own the inventory.

Closing

We won’t win the model race. We can win the implementation race. That requires marketing teams that know who is authorized to decide on their behalf when they aren’t looking.

If your next quarter starts with that question on your team, it already has a clearer beginning than most others.

A marketing organization in 2026 isn’t agentic because it has many agents. It’s agentic because it knows what its agents do.