Dial-Up Data, Streaming AI
Why this is our first real rebuild window since the internet, and why shortcuts are the tragedy
Confession: this is not a neutral think piece. It is a field report.
I’ve spent years inside retailer ecosystems where “our data is a mess” is not a punchline, it’s the daily operating condition. Everyone admits it. Everyone also quietly believes they’re the exception. They are not.
A recent post from Clare Kitching put language to the pattern I keep watching repeat: companies rush toward AI while the data underneath is still unresolved, unloved, and often unknowable. It stuck with me, not because it was new, but because it was true enough to make me go one layer deeper.
If every AI journey begins in the data basement, what are we doing down there right now? Are we actually rebuilding, or are we just installing prettier lighting and calling it transformation?
Because AI can become the most elegant cover-up we’ve ever deployed. Or it can be the first real chance since dial-up to rebuild the foundation properly.
If you remember dial-up, you remember the feeling: the screech, the pause, the prayer that nobody picked up the phone, and the weird pride of making it work anyway. We built the early internet on friction and optimism. Enterprise data got built the same way. Connect first. Clean later. Ship it. Buffer forever.
Now we are out of patience, and AI has entered the chat.
The “mess” is not just messy. It is expensive.
Gartner estimates poor data quality costs organizations $12.9 million per year on average. Gartner That is the cost that makes it into budgets.
The cost that rarely makes it into budgets is the one that erodes trust: rework, delays, broken confidence in metrics, and the culture of “don’t ask questions, the numbers will change again.”
At the macro level, Thomas C. Redman’s Harvard Business Review piece frames the national-scale impact as $3 trillion per year. Harvard Business Review Whether you treat that number as precise or symbolic, the point holds: bad data becomes a compounding tax on every decision you try to make.
AI does not reduce that tax.
AI makes the bill due sooner.
AI is a mirror, not a miracle
AI is not a shortcut around your data reality. It is a spotlight on it.
NIST’s AI Risk Management Framework makes a core point leaders keep trying to skip: trustworthiness includes the risk that “balanced” outputs can still produce harm when context, accessibility, representativeness, and real-world conditions are ignored. NIST Publications
So when leaders say, “We’ll just use AI to solve our data problems,” I hear: we want the output without paying the cost of truth.
That is not innovation. That is denial with a budget.
And here’s the practical version:
If your definitions are inconsistent, your AI will be confidently inconsistent.
If your data represents only the easiest customers, your system will treat everyone else like an exception.
If your governance is theater, your “responsible AI” will be a slide, not a safeguard.
Data cleanup vs data capability
Most organizations approach this era like a hygiene project: cleanse, dedupe, standardize, migrate, celebrate.
That is necessary. It is not sufficient.
Data capability is different. It is the system’s ability to perceive reality accurately enough to make decisions that do not quietly harm people. Hygiene makes the data prettier. Capability makes the system wiser.
If your foundation cannot register the human signal, you will build a technically elegant machine that still misses the point of retail.
People are not tidy.
A quick retail story, anonymized, because you’ve seen it too
A retailer launches “personalized offers” and the early model looks great on paper: conversion up, basket up, leadership happy.
Then the opt-out rate spikes. Customer service sees the pattern first: “This feels creepy,” “Why do you know that,” “Stop.”
The model did not suddenly become evil. The foundation was wrong.
Household stitching merged the wrong people. Shared devices blurred identities. Legacy loyalty accounts overlapped. A few “small” definition choices turned into a system that could not tell the difference between a household and a coincidence.
The fix was not a better prompt. The fix was adult work:
redefine household and identity rules
add lineage and accountability for critical fields
introduce observability checks for drift and merge anomalies
treat opt-outs and “creepy” complaints as sentinel signals, not PR noise
re-run training with corrected representation assumptions
That is what rebuild looks like. Not glamorous. Completely transformative.
The moment we are in: rebuild window, not feature race
The OECD AI Principles, updated in 2024, are blunt about the direction of travel: trustworthy AI must respect human rights and democratic values, not just optimize performance metrics. OECD
This is why I believe this moment matters more than most companies are willing to admit:
This is the first time since the introduction of the internet that we actually have leverage to fix the foundation instead of endlessly patching it.
AI can help us do the hard, boring work faster:
untangle definitions across systems
detect contradictions earlier
make lineage visible
reduce manual reconciliation
accelerate quality checks and monitoring
But the real opportunity is bigger than clean data.
It is building structured systems that can actually see people.
Shortcuts are costing you the future
The temptation right now is to sprint: pilot fast, ship shiny, “iterate later.”
Gartner predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data. Gartner
That is not a scare tactic. That is a pattern forecast.
If your plan is speed without foundation, you are not moving fast.
You are moving toward rework, quietly, at scale.
Where the Knox AI Empathy System belongs
This is the exact layer where my work lives, and why I built the Knox AI Empathy System in the first place.
Most “AI readiness” programs treat people like downstream stakeholders. Train the model, launch the capability, then manage the fallout with comms.
My premise is simpler, and more operational:
If the foundation cannot measure human impact, then the system is not ready to be trusted, no matter how clean the tables look.
So the Knox AI Empathy System is not a separate initiative you bolt on after data work. It is a way to design the foundation so human impact becomes measurable, monitorable, and actionable, right alongside performance and cost.
It’s built to deploy as a lightweight operating layer, not a multi-year transformation program, and it can be implemented without exposing proprietary scoring logic.
Not vibes. Instrumentation.
“Good enough to trust” is a decision, not a feeling
Not “good enough to demo.” Not “good enough to ship.”
Good enough to trust.
Here is a practical checklist I use in client work, and yes, it is intentionally unglamorous:
Definition stability
One definition per critical business concept, versioned, owned, and enforced (customer, household, trip, item, margin, promo, out-of-stock).Critical-field lineage coverage
Every field that changes outcomes has provenance and an accountable owner (where it came from, when it changes, who can change it).Observability and drift triggers
You can detect schema drift, distribution shift, merge anomalies, and missingness spikes before customers feel them.Representativeness beyond demographics
Your data reflects language, region, accessibility needs, device constraints, and digital divide realities, not just neat category bins. NIST PublicationsHuman-impact sentinels
Opt-outs, complaints, “creepy” signals, friction spikes, and trust loss are treated as leading indicators, with thresholds and escalation paths.Refusal rules
The system knows when not to guess. If confidence is low, it degrades gracefully or asks for help.
That is the line between “we launched AI” and “we built something people can live with.”
The Knox Foundation Pack (what makes this easy to work in)
If you want to operationalize the above without turning it into a two-year pilgrimage, this is the minimum artifact set that keeps teams aligned and makes governance real:
Critical Field Contracts
Definitions, owners, allowed transformations, and what “correct” means for the handful of fields that drive outcomes.Human Layer Integrity KPI
A simple scorecard that sits next to NPS and model performance, tracking trust signals (opt-outs, complaint sentiment, accessibility friction, perceived creepiness, error fallout) as first-class metrics.Trust Runbook
Drift triggers, escalation paths, refusal rules, and the “what we do when things go sideways” plan, written before launch, not after.
This is how the Knox system stays practical: it produces artifacts that teams can run, not ideals they can admire.
A practical rebuild plan: 12 weeks, Foundation + Feeling
This is the part most transformation programs avoid because it forces clarity. Here is a disciplined path that does not require a two-year reinvention.
Weeks 1–2: Map the decisions
Pick 3 decisions that matter (pricing, labor, personalization). Map how each decision gets made, end to end. If you cannot trace it, you cannot fix it.
Weeks 3–4: Lock critical-field contracts
Identify the 30–50 fields that actually drive outcomes. Write the definitions, owners, and allowed transformations. Kill duplicate definitions politely and permanently.
Weeks 5–6: Profile reality, not aspirations
Use automated profiling and AI-assisted analysis to surface contradictions, missingness, drift, and merge anomalies. Do not hide the mess. Put it on the wall.
Weeks 7–8: Build governance into the workflow
Governance that lives in a PDF is a bedtime story. Put checks into pipelines. Make lineage visible. Control who can change what. Tie it to release processes.
Weeks 9–10: Implement the Human Layer
Stand up the Human Layer Integrity KPI and wire in your sentinel signals: opt-outs, complaints, service friction, accessibility issues, and trust-drop indicators. Treat them as leading indicators, not emotional noise.
Weeks 11–12: Launch with a Trust Runbook
Set thresholds for quality, drift, and harm signals. Publish escalation paths. Define refusal rules. Launch with boundaries, then iterate with accountability.
That is what “AI-ready” should mean in practice.
Not perfect data.
A system that can tell the truth about itself.
What “good” actually looks like
A rebuilt foundation is not a perfect dataset. It is a system where:
definitions are stable enough that teams stop arguing in circles
data quality is measured and monitored, not guessed
lineage and accountability are visible
AI has boundaries, not vibes
human impact is measured alongside performance
This is also how you stop treating bias as a PR risk and start treating it as a design responsibility. NIST is explicit that harms can persist even when outputs appear balanced, especially when real-world conditions are ignored. NIST Publications
The question I want leaders to sit with
Most teams are asking: “How do we get AI value faster?”
The better question is: What kind of system are we becoming while we chase speed?
We have a rare window right now. A once-in-a-generation rebuild moment.
We can either use AI to paper over the basement one more time, or we can rebuild the foundation so our systems can see the people they claim to serve.
Anyone taking shortcuts right now is not being scrappy.
They’re missing the point of the era.
Sources
Gartner, “Data Quality: Best Practices for Accurate Insights” (poor data quality cost estimate, $12.9M/year). Gartner
Gartner Newsroom (Feb 26, 2025), “Lack of AI-Ready Data Puts AI Projects at Risk” (60% of AI projects abandoned without AI-ready data through 2026). Gartner
NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0) (trustworthiness, harmful bias, context and real-world impacts). NIST Publications
OECD, “AI principles” (adopted 2019, updated 2024; trustworthy AI respecting human rights and democratic values). OECD+1
Harvard Business Review, Thomas C. Redman, “Bad Data Costs the U.S. $3 Trillion Per Year” (Sep 22, 2016). Harvard Business Review
Clare Kitching, LinkedIn posts on AI-ready data and “our data is a mess” (2025). linkedin.com
