2025 Was the Year AI’s Missing Layer Became Obvious.

Why decision intelligence and context graphs became the real bottleneck for enterprise AI

Happy holidays and wishing you a strong start to 2026.

After years building decision intelligence systems in production (including for Google) and founding iCustomer - decision intelligence company, I’ve watched enterprises pour energy into dashboards, insights, automation and AI copilots while quietly missing the real shift.

In 2025, that gap became impossible to ignore.

Models improved fast. Outcomes didn’t.

AI didn’t fail this year.
What failed was the missing layer between data and execution, the layer where decisions are made, governed, learned from, and owned.

And yes, we all know “good AI needs good data.” Every enterprise leader understands that by now.
But this isn’t a data problem alone. It’s a decision activation problem: turning unified context into decisions, decisions into actions, and actions into learning reliably, repeatedly, in production.

---

The Validation Block: the industry converged

In recent weeks, independent thinkers across very different domains arrived at the same conclusion: the platform shift isn’t “LLMs inside SaaS.” It’s a new execution layer for how work happens.

• Ivan Zhao (Notion) - Steam, Steel, and Infinite Minds Notion

• Aaron Levie (Box) - Jevons’ Paradox for Knowledge Work LinkedIn

  • Alternate (same idea as a thread/post): X (formerly Twitter)

• Jaya Gupta & Ashu Garg (Foundation Capital) - AI’s trillion-dollar opportunity: Context graphs Foundation Capital

The Rise of the Decision Intelligence Layer

This isn’t “AI added to SaaS.”
It’s enterprise software being rebuilt around AI-native execution.

Old world
Systems of record → reports → humans decide → tools execute

New world
Intent → context → decision → execution → learning (continuous)

CDPs, dashboards, copilots? Transitional artifacts.

The real platform shift is the Decision Intelligence Layer the middle layer between intent and infrastructure, where decisions are made, executed, and learned from.

Five signals the market is re-architecting around decisions

1. LLMs commoditize. Decision intelligence compounds.
   The leverage lives in workflows, governance, exception handling, and learning from outcomes—not in clever prompts.

2. Context isn’t metadata.
   Context is decision-time understanding: inputs, intent, constraints, history, and permissions. Most enterprises store what happened. Almost none store why decisions happened.

3. Agents only matter inside execution paths.
   Agents outside execution paths are demos. Agents inside them become systems because intent, policy, approvals, and outcomes are captured in real time.

4. The middle layer is the new abstraction.
   Every platform shift looks primitive until the right layer emerges. This is that layer bridging raw infrastructure with real execution.

5. This is a multi-year rebuild, not a one-year cycle.
   The “last mile” in the enterprise is where trust breaks and value is captured or lost. That’s where the rebuild is happening.

The One Thing Missing: Decision Lineage

At the core of decision intelligence is something most enterprises don’t have today: decision lineage.

Enterprise systems were built to store records not to capture how decisions unfold.

Decision lineage is the stitched narrative across time, built from decision traces captured at execution, showing:

• what inputs were considered

• which policies applied

• where exceptions were granted

• why an action was taken

• what outcome followed

Decision lineage is to AI what data lineage is to analytics: the ability to explain, govern, and improve outcomes over time.

Here’s what “decision traces” look like in the real world:

• Decision: shift spend from Campaign A → Campaign B

• Inputs: rising CPA, creative fatigue signal, inventory constraints

• Policy: max CPA threshold, brand safety rules, budget caps

• Exception: approved temporary CPA ceiling for 48 hours

• Action: reallocate $120K, update pacing, refresh audiences

• Outcome: CPA drops 18%, conversion rate stabilizes, learnings logged

Without decision traces, AI can act but it can’t learn, explain, or earn trust.
Without decision lineage, enterprises can’t debug, govern, or compound improvement.

Context graph + decision traces → decision lineage → decision intelligence.

That missing lineage is now the real bottleneck for enterprise AI adoption.

Credit: Nano Banana Image from iCustomer Slide

What Actually Worked in 2025

1) Agents as Roles (Not “1,000 Agents”)

What worked wasn’t shipping hundreds of one-off agents or asking operators to “build their own agents” on top of point tools.

The winners treated agents as roles: durable, accountable digital twins (Marketing Ops, Performance Marketer, RevOps, Analyst, Support Lead) built on shared context, identity, permissions, policies, history, and outcomes.

When agents share context, they become systems. When they don’t, you get demos and noise.

2) Ownership (Not “Insights”)

Teams stopped asking, “What does AI recommend?”
They started asking, “Which decisions does the system own end-to-end and where do humans stay in the loop?”

The winning pattern looked like this:

• AI monitors, prioritizes, and proposes actions

• humans approve exceptions and manage tradeoffs

• the system records what happened and why

What worked in practice: prioritization, routing, budget allocation.
Not suggestions, ownership with guardrails.

3) Loops (Not Dashboards)

Dashboards tell you what happened. The systems that mattered answered four questions continuously:

• What happened (outcomes)

• Why it happened (drivers + context)

• How to respond (next action + constraints)

• When to act (timing, triggers, thresholds)

The winners closed the loop: act → measure → learn → adjust—automatically where possible, explicitly when humans needed to weigh in.

Because decisions without memory don’t compound.

A Simple Reality Check Before You Plan 2026

As you head into 2026, ask yourself or around:

• Where does decision lineage actually live in our organization?

• Which decisions improve automatically every week?

• What context is captured during execution vs. reconstructed later?

If AI didn’t reduce decision latency or compound outcomes, AI didn’t fail.

Your context and your lineage did.
And that’s a system design problem not a model problem.

Final thought: build systems, not features or even Agents.
The rebuild has begun. The only real question is whether you’re re-architecting for it or patching legacy assumptions with new tools.