Raj Kumar

[Raj Kumar @rajkstrong]

@stats_feed @grok Hi @grok, predict the result of these round of 32 matches at the 2026 FIFA WORLD CUP

🇵🇹 Portugal vs Croatia 🇭🇷
🇪🇸 Spain vs Austria 🇦🇹
🇦🇷 Argentina vs Cape Verde 🇨🇻
🇦🇺 Australia vs Egypt 🇪🇬
🇨🇭 Switzerland vs Algeria 🇩🇿
🇨🇴 Colombia vs Ghana 🇬🇭

[Raj Kumar @rajkstrong]

Most CEOs want AI results without AI investment.

The gains. The compression. The analyst quotes.

Not the rebuild. Not the org chart that no longer makes sense.

AI isn't a tech decision. It's an operating model decision.

What are you actually changing about how you lead?

[Raj Kumar @rajkstrong]

Most enterprise AI projects don't fail because the model is wrong.

They fail because no one decided who owns the context.

Models are now a commodity. Context — memory, retrieval, tools, eval — isn't.

Spend less on the model in 2026. More on the layer underneath.

[Raj Kumar @rajkstrong]

Three years ago I wrote a chapter called "The Death of the Workflow."

I was half right.

Workflows aren't dying. They're being absorbed — turned into the memory layer agents reason over.

The agent that wins doesn't replace your workflow. It turns it into a launchpad.

[Raj Kumar @rajkstrong]

Founder lesson tonight:

Every yes is a tax on focus.

Told a paying customer "no" today on a feature that didn't fit AgentLayer's vision. They might walk.

Still the right call.

The hardest founder muscle isn't technical clarity. It's emotional clarity.

[Raj Kumar @rajkstrong]

AI agents in enterprise apps: 5% in 2025 → 40% by year-end 2026.

Only 10% of orgs have actually scaled them.

The blocker isn't the model. It's governance.

Ownership. Promotion criteria. Incident protocol.

Pick the boring problem.

[Raj Kumar @rajkstrong]

The unpopular AI truth: model performance plateaued ~18 months ago for 90% of enterprise use cases.

The real gap isn't GPT vs Claude. It's teams that can redesign a workflow vs teams that can't.

Your AI strategy isn't a model problem. It's a process problem.

[Raj Kumar @rajkstrong]

Most enterprise AI strategies fail at the same point: the gap between the deck and what ops actually does Monday morning.

Every initiative needs:
- A revenue/cost line it owns
- A human accountable for that number
- A 90-day kill test

Strategy isn't the problem. Governance is.

[Raj Kumar @rajkstrong]

Sunday reflection:

In a salaried job, your week mostly happens to you.

When you build something of your own, your week is something you author.

Smaller. Scarier. Almost always behind schedule.

But it's yours.

[Raj Kumar @rajkstrong]

Everyone's building AI agents. Almost no one's building the accountability layer.

6 months into AgentLayer: the biggest AI gap isn't capability — it's explainability.

Ask your vendors: who's accountable when it goes wrong?

https://t.co/NsqERzgsHT #AgentLayer #EnterpriseAI

[Raj Kumar @rajkstrong]

OCC Bulletin 2026-13 explicitly carves agentic AI out of the MRM perimeter — the rollback question lives in that void: https://t.co/WxorHf9oFC

Control-plane thesis for agentic banking workflows: https://t.co/gg0evqA4y2

[Raj Kumar @rajkstrong]

A regional commercial lender I'm advising shut down its agentic loan-servicing pilot last week. The model was fine. Nobody could name who unwinds the agent's decision when the borrower calls at 9am.

Who owns the rollback chain at your shop — risk, engineering, or vendor?

[Raj Kumar @rajkstrong]

Sources:
EU AI Act Annex III point 5(b) — credit scoring as high-risk AI: https://t.co/a7swEChse5
CFPB Circular 2023-03 — adverse action with complex algorithms: https://t.co/SvhKcmQj5g
Control plane for bank credit agents: https://t.co/gg0evqA4y2

[Raj Kumar @rajkstrong]

At JPMorgan we ran ~240 credit decisioning models through MRM. The day Annex III landed, I realized we'd been grading the wrong exams.

Your model is validated. Your credit agent is not.

Which of those gaps would bite your shop first?

[Raj Kumar @rajkstrong]

CMS Interoperability + Prior Authorization Final Rule (CMS-0057-F) — the federal decision-trail requirement: https://t.co/2TOV928kkw

Control-plane thesis for agentic coverage decisions: https://t.co/gg0evqA4y2

[Raj Kumar @rajkstrong]

A top-5 healthcare payer I'm advising just deployed its first automated prior-auth agent. Denials up 23% in week one. The CMO asked how to pull it back.

An AI committee is not a model risk council. Actors need a paper trail.

Where's the governance gap at your payer?

[Raj Kumar @rajkstrong]

Source (Cowbell Prime One launch, April 21, 2026): https://t.co/E1hps34Rff

Building the control plane — tamper-evident agent logs, identity-bearing agents, policy-as-code: https://t.co/gg0evqA4y2

[Raj Kumar @rajkstrong]

April 21, 2026: Cowbell binds $10M cyber limits with affirmative AI coverage for midmarket ($250M-$1B) — before OCC MRM covers agents, before NIST CAISI finalizes, before EU AI Act Article 12 lands. The carrier wrote the control standard. Who owns your evidence package?

[Raj Kumar @rajkstrong]

Gravitee State of AI Agent Security 2026, 900+ respondents: https://t.co/3JEnwiNoV8

Control-plane thesis: https://t.co/gg0evqA4y2

[Raj Kumar @rajkstrong]

88% of organizations reported a confirmed or suspected AI agent security incident last year. 92.7% in healthcare. Only 21.9% treat agents as identity-bearing entities. 45.6% still share API keys agent-to-agent. This is an authorization gap, not an authentication one.