How Leaders Can Use AI to Solve Real Business Problems

Harvard Business Review · article · 2026-07-07 source ↗ Reader 🖥️

Change Management & Adoption Code Was Never the Bottleneck Domain Expertise as AI Force-Multiplier Enterprise AI Adoption & Readiness Explainable AI Frontier-Lab Hype

Successful AI adoption is determined less by which model or vendor an organization chooses than by whether it first identifies a specific, well-understood business problem, hands ownership to domain experts who genuinely care about that problem, builds trust through explainability and honest communication, and resists the frontier labs' self-interested narrative that missing the AI 'wave' is existential.

practical · argumentative — An HBR IdeaCast interview (host Adi Ignatius with journalist/author Josh Tyrangiel) structured as an argument for action: leaders should stop treating AI as an indispensable strategy imposed from outside and instead run it as a disciplined, problem-first, domain-expert-owned operational program. Two extended healthcare case narratives (Cleveland Clinic) function as the argument's evidentiary spine, not as decoration.

Structure

Part	Role
Framing: Tyrangiel's reporting arc from hype/doom split coverage to a settled 'dazzling but not perfect' view	context
Critique of frontier labs as self-interested narrators (the 'boat' argument, biotech CEO/Sam Altman anecdote)	evidence / argument
Reframing of AI backlash as a rational response to a 25-year cycle of tech overpromising	argument
Cleveland Clinic case studies: bed/transfer scheduling tool and sepsis-prediction tool	evidence
Claude Code 'empire' anecdote and 'get me a human' customer-service anecdote	evidence / counter-evidence (where problem-first discipline is absent)
Talent/retraining discussion (training up vs. broad retraining)	evidence
Direct prescriptive advice for 'regular' (non-tech, non-medical) executives	conclusion

Key terms

solving for AI
Adopting or pushing AI wherever possible because it is AI, rather than starting from a specific business problem AI might solve — the inverted, backwards approach Ignatius opens the conversation by naming.
the boat ("don't miss the boat")
Tyrangiel's reframing of Sam Altman's urgency rhetoric to a biotech CEO: it is the labs' boat — built on debt they need repaid — not necessarily the listener's boat. Urgency framing from a lab is a sales signal, not a disinterested forecast.
code (broad sense)
Any domain with dense if-then structure — software, but also law, mathematics, and policy/procedure manuals — which LLMs are 'inherently good at translating between.' Tyrangiel's answer to 'what AI opportunity are traditional companies not yet considering.'
spirit of R&D vs. spirit of certainty
The posture leaders should bring to AI rollout: framed as an open experiment with defined budgets and workforce participation, not as a pre-decided rollout being sold with confidence. 'There has never been yet a single perfect AI deployment with no problems.'
training up vs. (broad) retraining
Two distinct interventions Tyrangiel refuses to conflate: 'training up' merges an already-motivated, domain-expert person's skill with a new tool (his Cleveland Clinic example); 'retraining' reassigns someone to a wholly different kind of work, which he treats as a much harder, largely abandoned policy problem in the US.

Propositions

explicitSuccessful AI implementations are determined primarily by whether an organization first identified the specific problem it was trying to solve and held realistic expectations about solving it — not by which AI model or vendor it chose.conf: high
supportsWhy OpenAI and Anthropic Are Becoming Consultants · p1 — Both claim organizational/problem-definition readiness, not model capability, is the real bottleneck to enterprise AI value.
explicitFrontier AI labs' leaders are self-interested narrators of AI's necessity — heavily leveraged, for-profit companies who need customers to believe AI is indispensable to justify investor expectations — and are not better qualified than domain experts to judge how AI should be applied inside an unrelated industry.conf: high
extendsOrdinary Engineers, Not Heroic Inventors · p12 — O'Reilly's critique of procurement/moonshot-driven AI strategy gets a concrete mechanism here: labs' debt-driven need to sell urgency is why 'choosing a vendor' framings misdirect effort.
explicitPublic backlash against AI (booing of tech leaders, hostility toward data centers) is largely a rational response to a 25-year pattern in which the same companies and leaders overpromised on the internet, smartphones, and social media, and delivered real harms (misinformation, shortened attention, addiction).conf: high
explicitAI initiatives succeed when ownership sits with a domain expert who is personally invested in the specific problem (e.g., a clinician acting as product manager), rather than when they arrive as vendor-driven partnerships or centrally imposed technology mandates.conf: high
explicitFrontline professionals will not act on an accurate but unexplained AI recommendation in a high-stakes, alert-saturated environment; adding explainability to the same underlying model changed clinician behavior and was followed by a 41% reduction in sepsis mortality over a year.conf: high
extendsTo Adopt AI at Scale, Employees Need to Trust Agents · p12 — Supplies a concrete before/after mechanism (unexplained flags ignored, explained flags acted on, mortality fell 41%) for the abstract 'calibrated trust' design goal.
explicitEven a well-tuned, explainable predictive model has an irreducible last-mile gap where embodied human judgment (smell, visual appearance) catches cases the model misses, because human illness patterns can defy the statistical patterns the model was trained on.conf: high
explicitGenerative coding tools let individual engineers ship far more, far faster, while removing the friction that used to force alignment with product/customer requirements — so speed of code production does not by itself mean the actual brief was solved.conf: high
extendsElizabeth Ayer on Mastodon — "The fact that we are *not* seeing wildly improving software all around us tells us everything we need to know" · p4 — Ayer's claim that code creation isn't the bottleneck gets a mechanism here: faster code production (Claude Code) actively removes the friction that used to force alignment with the real brief, making the non-code work more, not less, load-bearing.
explicitCustomer-facing AI/chatbot deployments frequently fail to address the customer's actual need, evidenced by 'get me a human' being reportedly the most common message typed into bot-driven customer service.conf: high
explicitEffective AI-related workforce development is a collaborative integration of an already-motivated person's existing expertise with a new tool ('training up'), which is a fundamentally different — and easier — problem than imposed, whole-job retraining; the US's prior federal retraining infrastructure (roughly 5 million people over ~40 years) ended in 2022 with no replacement effort.conf: high
explicitLeaders should stop referring to 'AI' as one monolithic thing when communicating with their workforce, because vague messaging fails to target the actual problem being solved and increases unaddressed anxiety.conf: high
explicitRadical transparency with employees about AI's uncertain effect on their jobs ('maybe, we don't know') builds more trustworthy information flow than corner-office secrecy, because it invites people into the pilot rather than making them subjects of it.conf: high
explicitThe businesses best positioned to extract near-term value from LLMs are those with substantial 'code' in the broad sense — software, but also law, mathematics, and structured policy/procedure documents — because LLMs are inherently strong at translating between rule-based systems.conf: high
implicitThe technical staff best at building and tuning AI models are often poorly equipped to understand human systems, customers, and the 'soft skills' the tech world has devalued — and these soft skills are actually decisive for business success with AI.conf: medium

Arguments

Leaders should not defer to frontier AI labs as authoritative, disinterested advisors on how their own business should use AI. strength: medium

Explicit premises:

Labs call themselves 'labs' but are for-profit companies with massive bottom-line ambitions.
Labs carry heavy investor debt and need customers to believe AI is indispensable and easy to roll out.
The technical skill of building/tuning models is different from, and does not confer, the skill of understanding a specific business's customers and systems.

Implicit premises:

A self-interested party will systematically frame its own product as more urgently necessary than an objective assessment would support.
'Soft skills' (understanding human systems, customers) are the actual bottleneck to business success with AI, not modeling skill.

Inferences:

Procurement/urgency pitches from AI vendors should be weighted with the same skepticism applied to any interested seller, and domain-expert judgment should outrank vendor claims in adoption decisions.

AI initiatives succeed when a domain expert who cares deeply about the specific problem owns the deployment, rather than when it arrives as a vendor partnership or a top-down technology mandate. strength: high

Explicit premises:

At Cleveland Clinic, every AI pilot is run by a clinician acting as product manager.
The CEO explicitly rejected becoming 'the Microsoft of healthcare' via broad tech partnerships.
Both cited tools (bed/transfer scheduling, sepsis prediction) originated from a specific clinician's or team's operational frustration with a named problem, not from a vendor's product roadmap.

Implicit premises:

Caring about a problem plus baseline technical literacy predicts successful AI deployment better than deep technical expertise alone.
Generic, large-scale technology partnerships underperform bespoke, problem-first builds even when the counterfactual partnership is not directly observed.

Inferences:

Any leader evaluating an AI initiative should first ask who owns the problem and whether they care enough to iterate through the deployment's inevitable early failures, before selecting a vendor.

Technical accuracy alone is insufficient for an AI tool to be adopted; the tool must also earn the trust of the people meant to act on its output. strength: high

Explicit premises:

The initial sepsis-risk flags had no explanation attached and were ignored by doctors and nurses despite being present.
ICUs are already saturated with alerts, so an unexplained flag competes poorly for attention.
Adding explainability to the same underlying flags preceded a 41% reduction in sepsis mortality over a year.

Implicit premises:

People will not act on a high-stakes recommendation from an opaque source, especially amid alert fatigue.
Explainability functions primarily as a trust-transfer mechanism between model and human, not merely as an interpretability nicety.

Inferences:

Model accuracy metrics (precision/recall) are an incomplete measure of an AI tool's real-world value; deployment success should be assessed only after a trust/explainability layer is in place.

Narratives

The biotech CEO who met Sam Altman, was told 'you do not want to miss the boat,' and came away asking 'okay, what boat?'
→ Supports the proposition that labs' urgency framing is self-interested rather than disinterested technical wisdom.
Cleveland Clinic's hospitalist-built bed/transfer scheduling tool (built with Palantir, incorporating doctors' audio notes) that increased international transfer revenue and cut ER wait times by 90 minutes.
→ Supports the proposition that domain-expert-owned, problem-specific AI builds outperform generic vendor partnerships, and that operational/back-office AI use can directly improve patient care via the hospital's thin margins.
The sepsis-prediction tool (Bayesian Health, integrated with Epic) that doctors initially ignored until explainability was added to its red/yellow/green flags, after which sepsis mortality fell 41%; still, about 10% of cases were caught only by a clinician's non-pattern sensory judgment.
→ Supports both the explainability/trust argument and the proposition that human and machine judgment remain complementary, not substitutable.
A friend's software shop where engineers, empowered by Claude Code, build entire products solo and confidently present work that ignores parts of the brief and looks 'the ugliest thing I've ever seen.'
→ Supports the proposition that raw AI-driven speed in code production is not equivalent to solving the actual (customer-facing) problem.
'Get me a human' as the most common message typed into AI-driven customer service bots.
→ Supports the proposition that customer-facing AI is often deployed to cut cost rather than to solve the customer's felt problem.

Questions addressed

✅ What determines whether an organization's AI adoption succeeds or fails?
Problem-first framing, domain-expert ownership, explainability/trust-building, honest workforce communication, and realistic expectations — not model choice or spend.
✅ Do the people running frontier AI labs know more than domain experts about how AI should be applied in a specific outside industry?
No — they know more about models and tuning, but that is a different and narrower kind of knowledge than domain/customer understanding.
✅ Is public resistance and hostility toward AI justified, or is it just fear/ignorance to be overcome?
Largely justified — a rational response to a 25-year pattern of tech overpromising by the same companies and leaders now promoting AI.
✅ How should leaders approach workforce training for AI?
Distinguish 'training up' (integrating tools with already-motivated expert people) from broad retraining (reassigning people to different work) — the former is tractable now; the latter is a largely abandoned US policy problem.
⭕ What replaced the US's federal job-retraining infrastructure after it ended in 2022, and what should a leader do about broad retraining today?
⭕ What should a 'regular' (non-tech, non-medical, less-resourced) company actually do differently, given that both flagship proof points come from a single elite, physician-led health system?

Gaps

incompleteBoth extended success narratives come from Cleveland Clinic — a single, atypically well-resourced, physician-led nonprofit system with an idiosyncratic margin structure (2.2% is called 'heroic') and a CEO personally opposed to vendor partnerships. Ignatius himself raises the objection ('That's a hospital. I don't know how that relates to me. I'm a regular company') but Tyrangiel's answer stays at the level of a diagnostic checklist (where's your code, where's your data) rather than a second, matched case study from a resource-constrained commercial company. The generalization from N=2 elite examples to 'virtually any' traditional business is asserted, not demonstrated.
Read next: Look for a similarly detailed AI-adoption case study from a smaller, margin-constrained non-tech company (retail, manufacturing, logistics) to test whether the problem-first / domain-expert-ownership prescription actually transfers outside a flagship, well-capitalized institution.
incompleteThe claim that the 41% sepsis-mortality reduction followed specifically from adding explainability to the red/yellow/green flags is never disentangled from other concurrent changes over the same year — deeper Epic integration, continued model tuning, and growing clinician familiarity with the tool. No phased-rollout or counterfactual data isolates explainability's specific causal share of the outcome.
Read next: Look for a study or vendor case study (e.g. from Bayesian Health) that isolates the explainability intervention from concurrent integration/tuning changes to see whether the causal attribution holds up.
illogicalTyrangiel's central argument is that labs are financially self-interested and therefore their claims about AI's necessity should be discounted — but he does not apply the same skepticism to his own two showcase narratives, both of which come from Cleveland Clinic's own leadership self-presentation during a reporting engagement arranged specifically to demonstrate its AI program. A flagship health system also has clear incentives (donor relations, reputation, competitive positioning against actual 'Microsoft of healthcare' rivals) to present a flattering AI narrative to a journalist, and that possibility is never raised.
Read next: Compare independent reporting or peer-reviewed outcomes data on the Cleveland Clinic sepsis and scheduling tools against the self-reported figures given here, to check whether the same self-interested-source skepticism Tyrangiel applies to labs survives applied to his own sources.
incompleteTyrangiel notes the US federal job-retraining program 'died in 2022' with 'no effort... to bring it back up,' but never explains why it ended, what (if anything) replaced it at the state or private-sector level, or what a leader relying on external retraining infrastructure should actually expect today. This is left as an assertion of absence rather than a mapped current landscape.
Read next: Read current US labor-policy reporting on workforce retraining funding post-2022 to establish what infrastructure (if any) leaders can actually draw on for broad (non-'training up') retraining.

How Leaders Can Use AI to Solve Real Business Problems

Structure

Key terms

Propositions

Arguments

Narratives

Questions addressed

Gaps

Related documents