Jack Dorsey Just Fired Half His Company. D2C, Retail and Fintech Leaders, You’re Still Asking the Wrong Question.
Why data infrastructure alone was never the unlock and why decision architecture is the real competitive moat and key to showing ROI on AI transformation.
A year ago, I wrote here that retail’s AI problem wasn’t AI it was data: 35+ disconnected systems, most 1P data unactivated, teams stuck doing pipeline work instead of building intelligence. I called the shift from Data Foundation 2.0 (CDP mania) to 3.0: knowledge-first, AI-native architecture. The piece went viral data leaders nodded, it triggered great CDO conversations and then most companies did what they always do: went right back to arguing about the stack.
Meanwhile, Jack Dorsey cut 4,000+ roles from Block’s 10,000+ workforce. Shares jumped 20%+ after-hours. He didn’t call it a restructuring. He said “intelligence tools have changed what it means to build and run a company,” and that most companies will reach the same conclusion within a year. The business wasn’t in trouble. Gross profit was up 24%.
Predictably, the internet debate split: AI is killing jobs vs. this is just COVID overhiring. Oxford Economics says the “AI layoff wave” evidence is still patchy, and critics point out Block’s headcount more than tripled from 2019 to 2022. Even Dorsey partially conceded it.
Both camps miss the mechanism. When tools don’t embed decision logic, orgs compensate with people - coordinators, analysts, reviewers until headcount becomes the decision layer. That tax compounds. The correction isn’t AI replacing humans; it’s decision architecture finally doing the work headcount was hired to cover.
The question isn’t whether you’ll face a Block-style cut. It’s whether you design the architecture that makes one unnecessary. Wherever data/ customer decisions run at volume - retail, commerce, fintech, B2B2C when intelligence collapses execution cost, headcount stops tracking growth. What matters is whether you redesigned decisions or just modernized data.
Are you building from the data up? Or from the decision down?
Those are two completely different companies.
What Changed in 12 Months
When I wrote about Data Foundation 3.0, the conversation was about knowledge graphs, semantic models, AI-ready architecture. Right call.
But in 2025, “context” became the new umbrella word vendors use to sell the same thing. Every pitch deck wrapped itself in it and “context” became the new CDP. Everyone claimed to own the context layer. Nobody agreed on what it meant.
I don’t want 2026 to become another year of that debate while businesses keep buying context layers and nothing moves.
Context without a decision architecture is just a very expensive rearview mirror. You can have the most sophisticated context layer in the industry and still have a VP reading an insight deck on a Tuesday, deciding what to do about it next quarter.
The question was never: how do we serve the right context?
The question is: what does the system decide to do with it and who designed those boundaries?
That’s where most organizations are stuck. MIT Project NANDA’s 2025 “GenAI Divide” report estimates $30–40B in enterprise GenAI spend, yet 95% of organizations report zero measurable P&L impact not because models don’t work, but because organizations didn’t redesign the decision layer.
Not at the data layer. At the decision layer.
The Three Levels - And Why Most Teams Are Misreading Them
Level 1 — Systems of Record + Reporting. Tools store events, teams interpret, teams act. Reports explain last month. Humans stitch together insights across commerce, marketing, CX, and operations. This is where most teams actually operate not because they’re behind, but because this is what the market sold them for a decade.
Level 2 — Intelligence That Watches. Tools detect, summarize, predict but humans still drive every execution decision in silo. Churn scores. Cart signals. Journey simulation. More sophisticated but still fundamentally reactive. The human is still the bottleneck.
This is also where well-resourced teams get stuck. CDAOs, data science teams, warehouse-native stacks caught between what the business wants and what the data team wants to rebuild. More tools. More Data. More governance. More tickets. Quarterly roadmaps while customers churn.
Level 3 — Governed Autonomy. Agents execute within explicit decision rights, budgets, and guardrails and outcomes feed learning. Flags the at-risk customer before the signal is obvious. Routes the right intervention not the notification, the action within decision boundaries your business owns.
The gap between Level 2 and Level 3 is not a data problem. Not a context problem. Not an LLM or model performance problem. It’s a decision architecture problem for a human + agentic system and it has to be solved top-down by the business.
A quick diagnostic if you can’t answer these, you’re still at Level 2:
Do we have the top 20 decisions that drive our business written down?
Do we have decision owners and SLAs for each?
Do we have guardrails and policies versioned and documented?
Do we capture a decision log why a decision was made, not just what happened?
Do outcomes feed back and update policy?
The Real Barrier: Nobody Has Mapped the Decision Waterfall
Here’s how AI transformation actually fails:
A team gets the data foundation right. Connects product graph, customer graph, inventory, campaigns. Brings in an LLM layer. Launches an AI Assistant for marketing. Beautiful insights surface. The team reads them, goes to a meeting, escalates to a VP who puts it on next quarter’s roadmap.
The intelligence existed. The decision architecture didn’t.
This is what Block is building toward. This is what most organizations haven’t started.
The waterfall runs top-down. Here’s the operational template save this:
We call this the Decision Waterfall.
Most organizations have parts of this. Very few have all ten pieces in sequence, owned, with accountability at each level.
The teams building top-down will look radically different in 18 months. The teams still building bottom-up more data, better context, another AI tool that surfaces insights will look like what Dorsey just cut.
Why Decision Architecture Has to Be Business-Led
AI transformation is stalling because it’s being driven without decision ownership.
IT optimizes for infrastructure stability. Data teams optimize for data integrity. Both are doing their jobs well but neither role, by definition, wakes up worried about your customer’s next purchase decision. In 2024, “AI strategy” became their conversation and business leaders stepped back because it sounded technical: LLMs, inference, models, RAG, vector databases. That gap is the problem.
This doesn’t have to mean a CMO or a COO leads the charge. Some of the best decision architects are CDOs or technical leaders who think in business outcomes first. What matters isn’t the title it’s whether the person leading this can answer: what decision are we trying to replace, augment, or automate and what does the business lose if it gets it wrong?
Only someone with that business-decision fluency can map which decisions stay human and where agents act autonomously. If nobody owns that mapping, the data team defaults to optimizing pipelines, and the IT team defaults to optimizing uptime. Both rational. Neither sufficient.
The leaders who survive this chapter will be the ones who asked: What are the 20 decisions that drive our business? Which require human judgment? Which should never require it again?
That’s decision architecture. The function doesn’t matter. The fluency does.
What Leaders Should Do Right Now
Audit your decision landscape, not your data landscape. List every significant GTM decision made in a week - offer timing, suppression rules, audience prioritization, budget reallocation. Map who owns each and how long it takes.
Separate decisions into three buckets - fully human (strategic, high-stakes), augmented (human final call, agents do the prep), autonomous (repeatable, rule-bound, high-volume). In most orgs, a majority of execution decisions are repeatable enough to automate. Almost none currently are.
Hold vendors to one question: Show me an agent acting within a defined decision boundary without a human triggering it. If they can’t, they’re selling Level 2 dressed in Level 3 language.
Redesign the org around the decision map, not the tool map. Roles shift from campaign execution toward decision governance and exception handling. The efficiency Dorsey signaled doesn’t come from cutting people - it comes from eliminating work that only existed because software was dumb.
The Uncomfortable Synthesis
A year ago, I told you the data stack was the barrier. That was true. It remains foundationally true.
But the industry ran in the wrong direction. More context. More enrichment. More signals & Automations. All without the one thing that moves the needle: a decision architecture that starts with the customer, maps buyer and seller intent, and cascades deliberately through what humans own and what agents execute.
Context is not the unlock. Decision architecture is.
The companies that figure this out in the next 12 months will be smaller, faster, structurally different. The ones that don’t will have the most sophisticated data stacks in the world and a board meeting asking why revenue is flat.
Dorsey showed everyone the endgame. Gross profit up 24%. Headcount down 40%. Stock up over 20% in a day.
It’s whether you design the path or someone designs it for you.
Start building. Top down.
Where are you stuck - Level 1, Level 2, or somewhere in between? Reply or reach out. I’d like to hear what the blocker actually is.
If this reframed how you think about your stack, share it with someone who’s still equating AI strategy with model selection.
Author’s note: An LLM helped with research, citations, and light proofreading without changing content. All ideas, arguments, voice, and em dashes are mine. :)
At iCustomer, we build the decision layer most organizations are missing - intelligence that maps your customer signals to automated execution without replacing the judgment your team is actually good at. If your MAD (marketing, advertising & data tech) stack is overwhelming you but your ROI isn’t moving, that’s the gap we close with a composable decision middle layer.