Five real dilemmas.
Make the call yourself.
Across the practice: data foundation, decision intelligence, AI-led engineering, trust & compliance, and adoption. Each one actually happened, sector-scrubbed. No trick options; both choices are things intelligent, well-meaning people did in the room.
A multi-million-dollar Lakehouse build has stalled for six months. Every division defines "active customer" differently. That's twelve versions of the same metric. Engineering is fully staffed and idle, waiting for a decision from above.
What's the highest-leverage next move?
Let engineering decide and move. Fast, defensible, and exactly what stalled the programme for six months in the first place: a technically correct standard nobody outside the platform team had bought into.
The foundation wasn't a technology problem. It was an unresolved political one wearing a technical costume. We built the glossary and the ownership model first. The architecture that followed took two weeks, not six months.
Single source of truth, agreed before it was built, which is the only way a data foundation for AI actually survives contact with the business.
A churn-prediction model is statistically excellent, with an AUC well above the industry benchmark. Branch managers never open the dashboard. Leadership is ready to write off the entire predictive analytics investment as "adoption failure."
What's the highest-leverage next move?
Train the managers. Reasonable enough, assuming the sophistication of the output justifies asking the audience to level up to meet it. It didn't fail because managers were undertrained; it failed because a probability score isn't a decision.
Decision intelligence isn't the model. It's the last three feet between the number and the action. We replaced "0.73 probability" with "Call these 12 customers this week; here's why each one." Same model, different question answered.
Went from unopened to cited in every regional business review, because the output finally matched the decision it was meant to inform.
11,748 tables and 1,550 ETL objects need migrating off a legacy platform. An AI code-generation tool can auto-translate roughly 80% of the transformation logic in days rather than months. The engineering lead is nervous about trusting AI-written pipelines at this scale.
What's the highest-leverage next move?
Trust the confidence score. It sounds rigorous: let the tool tell you where it's shaky. But a model's confidence and a table's business criticality are two completely unrelated variables. High-confidence, high-stakes tables are exactly where a quiet error is most expensive.
AI-led engineering still needs a human judgment layer, just relocated. We didn't gate by the AI's confidence; we gated by business consequence. Every financially or regulatorily sensitive table got a human review, full stop.
Migrated with zero reconciliation errors in the tables that mattered most, because the review effort was allocated by stakes, not by how sure the model felt about itself.
A regulator is asking pointed questions about systemic risk exposure that twelve internal teams can't answer consistently. Each team's numbers disagree. A consultancy proposes a rapid compliance-reporting layer to generate one unified answer for the regulator within 90 days.
What's the highest-leverage next move?
Ship the fast layer. Meets the deadline, looks responsive, and is the standard move under regulatory pressure. It also means the underlying disagreement resurfaces at the next audit, on a shorter fuse and with less trust in reserve.
Compliance built on top of disagreement is theatre, not trust. We were transparent with the regulator about the timeline and used it to build real, durable ownership: governance the twelve teams actually held, not a layer painted over the gap.
Full systemic-risk transparency, restored regulator confidence, and no repeat finding a year later, because the fix addressed the actual disagreement.
An AI-assisted KYC and credit-review tool is technically excellent. It cuts document review time by 70% in testing. In production, the credit analysts quietly avoid it. Adoption sits near zero after three months.
What's the highest-leverage next move?
Mandate it. Direct, fast, and technically justified given the efficiency gain. It also confirmed the analysts' actual fear, that the tool was replacing their judgment, not assisting it, and adoption became compliance theatre, not real use.
The resistance wasn't about the technology's accuracy. It was about who got to make the final call. We rebuilt it as a copilot: the AI surfaces evidence, the analyst decides. Same underlying model, completely different relationship to it.
Adoption transformed and processing time still fell sharply, because the analysts' judgment stayed the last word, not an afterthought.
Same intelligence.
Different judgment.
Across data foundations, decision intelligence, AI-led engineering, and compliance, the pattern repeats: the "obvious" answer wasn't wrong. It was intelligent, defensible, and what most capable people would choose under pressure. The outcome changed on a smaller, quieter decision each time: whose actual experience to sit with before acting. That's sensibility. Not a personality trait, a discipline, applied across the whole practice.