Thought Leadership · AI Strategy
What separates AI strategy from AI theater
A field guide to where AI is actually working in 2026 — by industry, by platform, and by direction (outward vs. inward) — and to the hardest open problem in the room: proving what it's worth.
Select a node — the matching card below expands with its goal, objective, and description.
The framework
The six core components of an AI strategy
Every durable AI strategy runs through the same six moves, in roughly this order — and cycles back to the first once something ships. Click a component (or a node in the diagram above) to see how it connects to the rest of the post.
Problem Selection
Goal: Target real pain, not hype.
Objective: Start every initiative from a measurable, expensive workflow.
Projects that begin with "we have an expensive, repetitive process" succeed far more often than those that begin with "we should use AI."
Platform Fit
Goal: Match the tool to the task.
Objective: Choose generative, agentic, or classic ML based on the problem's shape, not the trend cycle.
The best platforms aren't defined by which model they expose, but by whether they run governed, repeatable workflows safely in production.
See the platform map ↓Direction: Outward vs. Inward
Goal: Balance visible wins with durable leverage.
Objective: Run customer-facing and internal-efficiency AI as two distinct, separately measured tracks.
Internal AI is often the safer proving ground before a customer-facing bet goes live.
See the two tracks ↓Governance & Risk
Goal: Make AI trustworthy at scale.
Objective: Build logging, oversight, and audit trails before agents get write-access to real systems.
Agentic systems carry a fundamentally different risk surface than static models — governance has to catch up before scale does.
Adoption & Change
Goal: Turn pilots into habits.
Objective: Treat adoption as a people problem, not a deployment problem.
A working pilot nobody uses has an ROI of zero — roughly a third of workers admit to quietly resisting AI tools.
ROI & Measurement
Goal: Prove the value, not just the possibility.
Objective: Set baselines and KPI gates before funding, and re-measure through each stage of scale.
Only about 29% of organizations can confidently measure the ROI of the productivity gains they already report.
See why it's hard ↓01 · Adoption
Which industries are leveraging AI, and how
Adoption is no longer a technology-sector story. It's broad, but depth varies sharply by industry, and the "how" differs even more than the "how much."
Adoption is now the default, not the exception
88% of organizations use AI in at least one business function, up from 55% two years earlier — one of the fastest enterprise technology adoption curves on record, per Deloitte's State of AI in the Enterprise.
Operational AI is furthest along
Robotics, autonomous vehicles, and drones are already reshaping manufacturing, logistics, and defense operations — areas where AI augments physical, repeatable workflows.
AI across the full customer journey
Retail is among the fastest adopters, deploying AI from product discovery through post-purchase support; over 95% of customer support interactions are expected to involve AI in some capacity.
Risk-first deployment
Financial services lean on AI where regulatory and financial risk is highest — fraud detection, underwriting, and investment analysis — prioritizing precision and auditability over speed.
R&D is compounding fastest
In automotive and aerospace, AI in R&D can cut time-to-market by roughly 50% and costs by 30%. Several pharma companies have cut drug-discovery timelines by more than half.
The agentic minority is growing
23% of organizations are already scaling an agentic AI system in at least one business function, with a further 39% experimenting — per McKinsey's State of Organizations 2026.
02 · Fit-for-purpose
Different platforms for different jobs
"AI" is not one technology choice. The best enterprise platforms aren't defined by which model they expose, but by whether they let you run governed, repeatable workflows safely in production — and the right layer depends entirely on the use case.
| Use case | Platform type | Why it fits |
|---|---|---|
| Customer support, FAQs, ticket triage | RAG / Generative | Grounded chatbots over a knowledge base resolve routine volume end-to-end; deflection commonly runs 40–70%. |
| Multi-step transactions — rebooking, returns, account updates | Agentic | Requires tool-calling and system write-access, not just text generation; agentic AI is expected to have its highest impact here. |
| Forecasting, demand planning, predictive maintenance | Classic ML | Structured historical data problems — an LLM is usually the wrong tool entirely. |
| Fraud and anomaly detection | ML + feedback loop | Labeled signals with analyst corrections fed back in; precision matters more than fluency. |
| Document-heavy work — legal, claims, compliance | RAG / narrow fine-tune | Extraction and summarization with mandatory human-in-the-loop review. |
| Software development | Agentic coding assistants | AI coding assistants now write an estimated 41% of code at firms with high adoption, per McKinsey. |
A caution worth sitting with: Gartner projects that more than 40% of agentic AI projects will be cancelled by the end of 2027 over escalating costs, unclear business value, and inadequate risk controls — matching the platform to the job matters more than chasing the newest layer.
03 · Direction of value
Outward-facing AI vs. inward-facing AI
The most durable AI strategies run two tracks at once: a visible, external track for customer engagement, and a quieter, internal track for operating leverage. They are funded, measured, and governed differently.
AI for core, customer-facing business functions
External AI enhances customer experience and engagement and directly targets the top line — but it's the higher-variance bet.
- Agentic customer service that resolves, not just routes, requests end-to-end
- AI-assisted product discovery, personalization, and post-purchase support
- Bain research finds agentic AI's next wave is reinventing customer experience, not just cutting cost
- Higher visibility, higher scrutiny, harder to walk back publicly if it underperforms
AI for internal efficiency
Internal AI drives tangible cost savings by automating processes, reducing errors, and improving decisions — with a shorter, more measurable feedback loop.
- Knowledge-work copilots, coding assistants, internal search and summarization
- Fraud, risk, and anomaly detection embedded in existing workflows
- Lower-stakes environment to test, refine, and build organizational AI fluency
- Often the better starting point: prove value internally before shipping it to customers
04 · The hard part
Why it's so hard to define ROI on AI projects
This is where most AI strategy conversations stall. It isn't that AI has no value — it's that the value doesn't show up where traditional ROI models look for it.
Impact is diffuse, not a single line item
AI's effects spread across workflows, employee efficiency, collaboration, customer experience, security, and process quality simultaneously — traditional single-metric ROI models can't capture all of it at once.
Value compounds slowly, not immediately
Most organizations now expect satisfactory AI ROI within two to four years — far longer than the seven-to-twelve-month payback window typical of conventional IT investment.
2–4 years to satisfactory ROIA lot of the value is prevention, not output
Anomaly detection and risk-flagging AI create value by stopping problems before they escalate — an outcome that's structurally harder to quantify than a new dollar of revenue.
The productivity-to-financial-impact gap
79% of organizations report productivity gains from AI, but only 29% can confidently measure ROI — the translation from "faster work" to "P&L impact" is where most projects lose the thread.
79% see gains · 29% can measure ROIMessy data undermines the measurement, not just the model
AI systems depend on organizational data. When that data is incomplete, siloed, or disorganized, it's not just the model that suffers — the ROI measurement itself becomes unreliable.
05 · The fix
How to actually define ROI
The organizations that measure AI ROI credibly don't apply one generic formula. They match a payback model to each use case and treat the number as a living calculation, not a one-time approval gate.
Separate hard ROI from soft ROI — track both
Hard ROI: labor cost reduction, hours saved, incremental revenue, error reduction. Soft ROI: employee retention, skills uplift, brand strength, valuation effects. Both are real; only one shows up in a spreadsheet by default.
Set the baseline and the metric before you build
Define success with a practical metric and a measured baseline up front — not after the pilot ships. Projects that start with "we have an expensive, repetitive workflow" outperform projects that start with "we should use AI."
Budget the run cost, not just the build cost
Monitoring, retraining, governance, and human review are ongoing costs that shift the Total Cost of Ownership as the system matures — over half of finance leaders say they can't clearly demonstrate AI ROI, and an unbudgeted run cost is a common reason why.
Measure across stages, not just at launch
ROI looks different in experimentation, integration, and scaling. Early stages validate feasibility and risk; scale is where the financial outcome actually shows up. Re-measure at each stage instead of judging the whole initiative on day-one numbers.
Gate funding on a small, fixed KPI set — before approval
Financial (payback period, ROI multiple, cost per outcome), adoption (active usage, workflow integration), and quality (accuracy, escalation rate) metrics, agreed before the project starts, are what separate the roughly 12% of AI investments that show measurable returns from the rest.
~12% show measurable returns todayRelated work
Go deeper: right-sizing the investment itself
From my blog
The AI Investment Decision Guide — Right-Sizing AI for Real ROI
A companion piece that goes further into the mechanics above: the four challenges that kill AI projects, an Investment Ladder for matching spend to problem size, a use-case map of buy-vs-build-vs-fine-tune, and a full KPI scorecard to gate funding before you build.
Read the full guide →