Agent Slop: Why Most AI Projects Create Chaos, Not ROI
The Context Graph thesis is inevitable, but context is just the map. To survive in production, you need the missing layer: Decision Infrastructure. That's what quiet winners are building.
TL;DR
Enterprise AI is splitting into two tracks. A small set of teams are quietly winning with measurable lift & ROI proof. A much larger set is stalling not from bad models, engineering bandwidth, but from missing decision infrastructure.
Gartner’s latest prediction: 40% of agentic AI projects will be canceled by 2027 due to unclear value and risk. And if 2025 taught us anything when 30% of GenAI projects got abandoned after POC it’s that they’re probably underestimating the complexity.
This isn’t because agents are useless.
It’s because most enterprises skipped the unsexy part:
The Decision Layer.
Not another tool. The infrastructure that makes AI-driven decisions governable, explainable, and improvable so learning compounds instead of resetting every quarter.
What Happens After the Demo: The Reality Vendors Hide
Agents are cool. The demos are intoxicating. Then reality hits:
Outputs need human verification
Approvals get messy
Exceptions explode
Teams spend more time cleaning up “first drafts” than doing the work
That’s what I call Agent Slop: high-velocity activity that creates operational debt.
I watched a Fortune 500 marketing team deploy an agentic campaign tool last quarter. Three weeks in, they had 47 active campaigns. Only 12 had clear owners. Five were duplicates. The agent was “working” but the decision system was drowning.
The VP of Marketing told me:
“We automated ourselves into chaos.”
AI didn’t create their chaos.
It revealed it and accelerated it.
The Question Winners Ask (That Losers Don’t)
Winners don’t ask: “What can the agent do?”
They ask: “What decisions are we making repeatedly that we can’t defend or improve?”
That shift from automation-first to decision-first is the difference between Agent Slop and Agent ROI.
Context Graphs Are Necessary. But Insufficient.
Jaya Gupta (with Ashu Garg) has a strong thesis: context graphs may be the next trillion-dollar opportunity. Dharmesh Shah correctly notes that most systems capture what happened, not why.
I agree with the direction. I disagree with the framing.
The issue isn’t implementation. It’s the assumption that structure equals meaning.
Context graphs come from a computer-science instinct: model the world as nodes and edges. That works when the environment is logical and stable.
Business decisions are rarely purely logical. They’re shaped by perspective, incentives, timing, and pressure. Two executives can look at the same graph and reach opposite conclusions and both can be rational.
Here’s the uncomfortable truth: “Business decisions are political theater with a spreadsheet costume.”
If you’re a CxO, you don’t need a system that reconstructs reality (more data).
You need a system that reconstructs interpretation (defensibility).
Context without perspective isn’t context. It’s just data.
We Already Solved This Problem. Then We Broke It.
In the pre-agent era, we had a decisioning layer: Decision Infrastructure 1.0.
I lived this world building decision COEs and rule-led decisioning in the Pega era for companies beyond FAANG, like Amex, Target, GE, and Palo Alto Networks. The ambition was always the same: make high-value decisions repeatable and defensible.
Then I spent the last decade productizing this for marketers, helping build the CDP era the modern system of record for marketing, as MDM wasn’t built for activation or real-time enablement. We got much better at unifying customer truth and querying that data.
But here’s what most enterprises didn’t build which is a no-brainer in the agentic AI era:
A real system of agency the layer that turns truth into governed decisions.
So what became the default decisioning layer?
Spreadsheets + tribal knowledge.
That’s why agents create slop. We’re dropping a high-speed engine (AI agents) into an organization where the transmission (decisioning) is broken.
You can’t fix throughput when the system can’t shift gears.
The 3 Systems Every Enterprise Needs (Only 1 in 10 Has #2)
To win with agentic AI, you don’t need another SaaS App pitch. You need the operating model.
It comes down to three systems:
1) System of Record (truth)
Identity, consent, events, profiles, customer/account data the foundation the CDP/data cloud era made mainstream.
2) System of Agency (decisions)
The missing layer: Bespoke Signals + decisioning + unified analytics + agentic orchestration.
This is where the enterprise decides based on goals, budget, signals and decision-first marketing instead of channel-first:
What to do next
Why it’s doing it
How it measures success
What constraints govern risk
Who approves what
What it learns for next time
This is also where Decision Traces (receipts) live so decisions are explainable, governable, and improvable.
Example: A retailer I worked with had 14 different “audience selection” processes across teams & tools. Same data warehouse. No shared decision logic. Every campaign started from scratch.
We consolidated that into one governed decision loop same intent (revenue), same evidence sources (warehouse), standardized constraints (margin floors, inventory rules), consistent measurement.
First-quarter lift: 23%. By Q3: 41%.
Not because the data got better. Because the system was learning, not just executing.
3) System of Experience (execution)
Where content is rendered and campaigns are launched often inside walled gardens like Meta, Google, LinkedIn, Reddit, and AI platforms like ChatGPT, or your owned martech stack like Adobe, Salesforce, and the expanding ecosystem of content engines, ESPs, CEPs and ABM tools.
These platforms will keep owning distribution and pixels. That’s fine.
The mistake is letting them manage your agency.
Remember: Walled gardens and marketing clouds optimize for their yield, not yours. If you leave the decision logic inside their black box, you are renting your own brain.
When Meta decides which creative to serve, Google determines your bid strategy, or Salesforce scores your leads they’re making decisions with their objectives and one-size-fits-all models, not bespoke logic for your business.
Let the walled gardens own the pixels. Don’t let them own the brain.
That’s how you compound all that spend towards tangible outcomes.
The 60-Second Test Every CEO Should Run (Yours Will Fail)
Here’s the simplest test for whether your agent strategy will scale:
If you can’t explain a high-impact AI-driven decision in 60 seconds, you can’t scale it.
The CEO/CFO cares about three things:
Accountability - who owned and approved the decision?
Control - what guardrails prevented downside risk?
Repeatability - can we do it again next week and get better, not just different?
If you can’t answer these in 60 seconds, stop buying agents. You have a governance problem.
No receipt → no scale.
Why Marketing Is Where This Shows Up First
Marketing is where companies deploy massive discretionary capital and where the decision system is often the weakest:
Fragmented signals
Volatile auctions
Attribution arguments
Channel fiefdoms
Incentives that reward activity over learning
Every CMO or Growth Head knows the uncomfortable truth:
We are spending a lot, but we can’t prove what to change without heroics.
The goal isn’t “autopilot” (spray-and-pray automation).
The goal is a compounding decision system one that gets smarter every cycle, where a decision made in Q1 improves the decision made in Q3 without starting from scratch.
Most GTM orgs have a delivery system (campaigns) & adding more Agents wont help.
Winners have a learning system (decision loops).
Stop Chasing Logs. Start Capturing Receipts.
If you want ROI, stop shopping for magic. Start capturing receipts.
A Decision Trace is not a log. It’s a board-ready receipt for an AI-influenced decision:
Intent: What were we trying to achieve?
Evidence: What data did we trust?
Constraints: What guardrails (margin/brand) were honored?
Action: What did we do?
Outcome: Did it work? What happened vs. expected?
Learning: What changes next time?
This is the difference between “an agent did something” and “a governed decision happened.”
Context graphs give you the “what.”
Decision traces give you the “why” and the “so what.”
And when you accumulate traces over time, you can run decision loops which is what makes performance compound.
Why Campaigns Die But Loops Compound
Most GTM organizations operate in sprints. Winners operate in loops.
From: Campaigns → Programs
From: One-off experiments → Learning loops
From: Tribal knowledge → Institutional memory
Here’s the test: can your organization answer this without a Slack spiral?
“We spent $500K on this audience segment last quarter. What did we learn, and what are we changing this quarter as a result?”
If the answer requires heroics stitching dashboards, chasing the analyst who left, reverse-engineering a spreadsheet you don’t have Agent Slop yet.
You have Decision Debt.
Agents will just make it faster and more expensive.
Your Moat Isn’t Data. It’s Decisions That Compound.
Jaya and Dharmesh are right to elevate “why.”
But enterprise winners won’t be the ones with the biggest graphs.
They’ll be the ones with:
Decision Velocity (speed with control)
Decision Memory (receipts that persist beyond org charts)
In the agent era, your moat isn’t data.
It’s decisions that compound.
If You Can’t Articulate the Policies & Guardrails, You Have a Governance Vacuum
If you could delegate one high-stakes marketing decision tomorrow with a guaranteed decision receipt written back into your central data warehouse what would it be?
Not “what would be cool.”
What decision are you making repeatedly that you can’t defend or improve?
Because if you can’t articulate the guardrails that would make you trust it, you don’t have a decision problem.
You have a governance vacuum.
And no agent will fix that.
Thanks for reading!
Abhi Yadav
Building decision OS for Compounding Marketing
Working on Decision Loops across Adtech + Martech? Building Audience Context Graphs? Tired of agents that create more work than value? DM me- I trade notes with everyone