Linh Truong · AI-First
Linh Truong  ·  2026

The AI-First Company Playbook

How I think about moving a traditional company to AI-first: intelligence at the center of how you decide, build, and serve customers—not a side project bolted onto old workflows.

Author: Linh Truong, MA (Harvard), MBA
Source: LinhTruong.com
Email: Linh@Alumni.Harvard.edu
For: CEOs, CIOs/CTOs, transformation leads · Horizon: 0–24 months · Updated: May 2026
~70%
of AI transformations stall in pilots — they never reach production scale
3–4×
productivity uplift reported in mature AI-native workflows vs. manual baselines
<1%
of companies describe themselves as fully "AI-mature" today
5 layers
must move in lockstep: strategy, data, tech, people, governance

Stats below are directional (2024–2026)—verify against §17 before citing in a deck. Sources include McKinsey, BCG, and Gartner survey series.

00 Executive Summary

The thesis in one page

An AI-first company doesn't just "use AI." Data, models, and agents sit where decisions and delivery happen. AI-enabled is a turbo on the old engine; AI-first is designing around the engine.

🎯

It is a leadership problem first

Technology is rarely the bottleneck. The constraint is strategy, incentives, data ownership, and the courage to redesign workflows rather than decorate them.

🔁

Value compounds through a flywheel

Better data → better models → better products → more usage → more data. AI-First companies engineer this loop deliberately and defend it.

🤖

2026 is the agentic inflection

The shift from AI that answers to AI that acts — multi-step agents that execute work — changes the unit of automation from the task to the workflow.

Contents

01 The Case for Change

Why now — the 2026 inflection point

Three forces landed at once: capable models, agentic workflows, and boards asking for a plan. Waiting isn't neutral—competitors are rebuilding cost and speed around intelligence.

Capability

Frontier models crossed the utility threshold

Reasoning, long context, tool use, and multimodality mean models now handle ambiguous, multi-step knowledge work — not just classification. Quality is "good enough to ship" across many domains.

Economics

Inference cost fell ~10× per year

What was uneconomical in 2023 is routine in 2026. Cheaper tokens turn one-off demos into always-on production systems and reframe the build-vs-buy math.

Autonomy

Agents moved from demo to deployment

Tool-using, planning agents execute end-to-end workflows with human oversight. The automatable unit shifts from the task to the process.

The strategic stakes

The upside for movers

  • Cost structure reset — knowledge work unit costs fall as agents absorb routine throughput.
  • Velocity — product, content, and code cycles compress from weeks to days.
  • New products — offerings that were economically impossible become viable (hyper-personalization, 24/7 expert service).
  • Data moat — proprietary feedback loops widen the lead over time.

The risk for laggards

  • Margin compression — competitors pass AI savings to customers; you can't match the price.
  • Talent flight — the best people want to work where the tools are modern.
  • Shadow AI — employees adopt ungoverned tools anyway, creating risk without value capture.
  • Compounding gap — the flywheel means leads widen, not narrow, with time.
The reframe

The right question for leadership is no longer "Which tasks can we automate with AI?" It is "If we rebuilt this company today, knowing what AI can do, how would it look?" — then close the gap between that and today.

02 Definition

What "AI-First" actually means

AI-First is an organizing principle: AI is the default consideration in every significant decision about how work is designed and delivered. It spans a spectrum — most companies today are "AI-curious" or "AI-enabled," and the goal is to climb deliberately toward "AI-native."

DimensionAI-Enabled (bolt-on)AI-First (built-in)
Mindset"How do we add AI to what we do?""How would we operate if AI did the heavy lifting?"
WorkflowsExisting process + AI assistant on the sideProcess redesigned around AI; humans handle judgment & exceptions
DataLocked in silos; used for reportingTreated as a strategic asset; engineered into feedback loops
DecisionsGut + dashboardsModels propose, humans decide; decisions instrumented & learned from
OrgA separate "AI team" / innovation labAI fluency embedded in every function; central enablement, federated delivery
AmbitionEfficiency on current productsEfficiency and new AI-native products & business models

The five pillars of an AI-First company

Transformation fails when leaders treat AI as a technology project. It is a system change across five interdependent pillars. Weakness in any one pillar caps the whole.

Sustainable Competitive Advantage & Value Creation 1 Strategy & Leadership Vision · value cases · funding 2 Data & Infrastructure Quality · access pipelines · platform 3 Technology & Models Models · agents tooling · MLOps 4 People & Culture Skills · adoption incentives · trust 5 Governance & Risk Ethics · security compliance · trust Foundation: Executive Sponsorship · Clear Strategy · Adaptive Operating Model
Figure 1. The five pillars rest on a leadership foundation and hold up the value roof. Progress is gated by the weakest pillar.

03 Diagnostic

The AI maturity model

Before plotting a destination, locate yourself honestly. Most organizations sit at Stage 1–2 and overestimate their position. Each stage has a distinct bottleneck; you cannot skip stages, but you can move through them faster with deliberate investment.

STAGE 1 Exploring Ad-hoc, individual use STAGE 2 Experimenting Pilots & POCs, siloed teams STAGE 3 Operationalizing Production use cases, MLOps, platform STAGE 4 Scaling Cross-functional, measured ROI STAGE 5 AI-Native AI core to strategy, products & flywheel
Figure 2. The maturity staircase. The hardest leap — where most stall — is Stage 2 → 3: moving from pilots that demo well to systems that run in production.
StageHallmarkPrimary bottleneckMove to next stage by…
1 · ExploringIndividuals use chatbots informally; no strategy.Awareness & permissionNaming an exec sponsor; setting an acceptable-use policy; funding 2–3 pilots.
2 · ExperimentingDisconnected pilots; impressive demos, little production."Pilot purgatory"Building shared platform & data access; choosing winners; killing the rest.
3 · OperationalizingA handful of use cases run reliably in production.Repeatability & trustStanding up MLOps/LLMOps, governance, and reusable patterns.
4 · ScalingAI embedded across functions; ROI tracked.Org & change capacityFederated operating model, talent at scale, portfolio management.
5 · AI-NativeAI is core to products, decisions, and the business model.Continuous reinventionDefending the flywheel; building AI-native products; ongoing R&D.
The chasm

The gap between Stage 2 and Stage 3 is where ~70% of initiatives die. The cause is almost never the model — it's the absence of production data pipelines, ownership, evaluation, and a workflow redesign that makes the AI's output actually usable. Budget for the boring infrastructure.

04 The Engine

The AI-First value flywheel

The defining characteristic of an AI-First company is a self-reinforcing loop where usage generates data, data improves intelligence, intelligence improves the product, and a better product drives more usage. This is the moat. Strategy is largely about engineering and defending this loop.

VALUE FLYWHEEL 📊 DATA 🧠 MODELS 📦 PRODUCT 👥 USAGE trains & tunes powers generates feedback & signals
Figure 3. The compounding loop. Each turn of the wheel widens the lead. Friction at any node (e.g. data you can't access, products with no telemetry) stalls the entire engine.
Data

Proprietary, well-governed data — including the exhaust from your own operations — is the renewable fuel. Capture it by design.

Models

Mostly bought (frontier APIs) and adapted (prompting, RAG, fine-tuning) — rarely built from scratch. Intelligence is increasingly a utility.

Product

Where intelligence meets the customer. Must be instrumented to capture feedback (explicit ratings + implicit behavior).

Usage

More users in more situations = more signal. Design products that get measurably better the more they're used.

05 Strategy

Strategic framework & North Star

AI strategy is not a separate document — it is your business strategy expressed through what AI makes newly possible. Anchor it to a North Star, then choose where AI plays offense (growth) versus defense (efficiency).

The three strategic plays

⚙️

1. Optimize the core

Apply AI to existing processes to cut cost, time, and error. Lowest risk, fastest payback. Examples: support deflection, code generation, document processing, forecasting.

Defense · efficiency
🚀

2. Reinvent the offering

Embed AI into products/services so the customer experience materially improves. Medium risk, durable advantage. Examples: AI copilots in your product, personalization, dynamic pricing.

Offense · differentiation
🌐

3. Create new business models

Build offerings impossible before AI. Highest risk and reward. Examples: AI-native products, outcome-based services, autonomous agents-as-a-service.

Offense · new growth
Portfolio rule of thumb

Early on, weight ~70% of effort to optimize the core (build credibility & fund the program with quick wins), ~20% to reinvent the offering, ~10% to new models (option value). Shift toward offense as maturity grows.

Crafting your AI North Star

A strong North Star is specific, measurable, and tied to enterprise value. Use this template:

"By [date], we will [verb: lead / transform / automate] our [core domain] so that [measurable outcome — e.g. 50% of customer interactions are resolved autonomously with higher CSAT], making us the [differentiated position] in [market]."

Build vs. buy vs. partner

LayerDefault postureWhy
Foundation modelsBuy (frontier APIs)Capital-intensive, fast-moving; rarely a source of advantage to build yourself.
Orchestration & agentsBuy + assembleMature frameworks exist; your advantage is in how you compose them around your workflows.
Application layer / UXBuildClosest to your customer and differentiated workflows — this is where you compete.
Proprietary data & evalsBuildYour moat. Never outsource the feedback loop or the data that fuels it.

Heuristic: buy the commodity, build the differentiator, partner on the bleading edge. Re-evaluate quarterly — the line moves fast.

06 Organization

The AI-First operating model

The winning structure for most enterprises is the hub-and-spoke (federated) model: a central platform team provides shared capability, standards, and governance, while embedded "spokes" in each business unit own delivery and outcomes. Pure-central is too slow; pure-decentralized fragments and re-builds the same things.

AI CENTER OF EXCELLENCE Platform · standards governance · enablement Product & Eng Embedded AI pod Copilots, agents Operations Embedded AI pod Automation, forecasting Customer / Sales Embedded AI pod Service, personalization Corp / Finance / HR Embedded AI pod Back-office automation Risk / Legal Governance partner Compliance, oversight
Figure 4. Hub-and-spoke. The hub builds capability once; the spokes deliver value many times. Risk/Legal acts as an embedded partner, not a gate at the end.

Key roles to staff

Leadership & strategy

  • Executive sponsor (CEO/COO) — owns the mandate, removes blockers, holds the budget.
  • Chief AI Officer / AI lead — sets strategy, runs the portfolio, chairs governance.
  • AI product managers — translate business problems into AI solutions; own outcomes.

Delivery & platform

  • ML / AI engineers & data scientists — build, tune, and integrate.
  • Data engineers — pipelines, quality, governance plumbing.
  • ML/LLMOps engineers — deploy, monitor, evaluate in production.
  • AI champions — fluent power users embedded in every team driving adoption.
Operating principle

Centralize what benefits from scale and consistency (platform, security, model access, evaluation, standards). Decentralize what benefits from proximity to the problem (use-case selection, workflow design, change management).

07 Architecture

The AI-First technology stack

Think in layers. Most companies should rent the bottom of the stack and own the top. The 2026 stack is increasingly about compound AI systems — orchestrated agents, retrieval, tools, and guardrails — rather than a single model call.

6 · EXPERIENCE Copilots · chat · embedded UX · agent interfaces — where humans & AI collaborate BUILD 5 · ORCHESTRATION & AGENTS Agent frameworks · tool/function calling · planning · workflow routing · multi-agent BUILD+BUY 4 · KNOWLEDGE & RETRIEVAL (RAG) Vector DBs · embeddings · context engineering · grounding on proprietary data BUILD+BUY 3 · MODELS Frontier LLMs (Claude, etc.) · small/specialized models · fine-tunes · routing BUY 2 · DATA PLATFORM Lakehouse · pipelines · feature/embedding stores · catalog · quality & lineage BUILD+BUY 1 · INFRASTRUCTURE Cloud compute/GPU · networking · identity · cost controls · MLOps/LLMOps backbone BUY ⟵ CROSS-CUTTING: Security · Observability/Evals · Governance · Cost & FinOps ⟶
Figure 5. The layered stack. Security, evaluation, governance and cost controls run vertically through every layer — they are not a final step.

The five techniques to adapt models to your business

TechniqueWhat it doesCost / effortUse when…
Prompt engineeringInstruct the model precisely; provide examples.LowestAlways the first lever. Solves more than people expect.
RAG (retrieval)Ground answers in your documents/data at query time.Low–mediumYou need current, proprietary, or citable knowledge.
Tool use / agentsLet the model call APIs, run code, take actions.MediumThe task requires doing, not just answering.
Fine-tuningAdjust model weights on your examples.Medium–highYou need a consistent style/format or a smaller, cheaper specialist.
Pre-trainingTrain a model from scratch.HighestAlmost never — only for unique frontier needs with massive scale.
Sequencing rule

Exhaust cheaper levers before expensive ones: prompt → RAG → tools → fine-tune → (rarely) pre-train. Most production value is captured in the first three. Reach for fine-tuning to optimize, not to start.

08 Foundation

Data strategy — the real foundation

There is no AI-First company without a data-first foundation. Models are increasingly commoditized; your proprietary data and the loops that refresh it are the durable advantage. "Garbage in, garbage out" is now "garbage in, confident hallucinations out."

The four data imperatives

  1. Accessible — break silos; a unified, queryable platform beats perfect but locked data.
  2. Quality & governed — lineage, ownership, freshness, and definitions everyone trusts.
  3. Captured by design — instrument products & processes so usage continuously produces new signal.
  4. Compliant & secure — privacy, consent, residency, and access control built in, not bolted on.

Data readiness checklist

  • A single catalog of what data exists and who owns it
  • Pipelines that are reliable, monitored, and documented
  • Clear, enforced access & privacy controls (RBAC, PII handling)
  • Feedback telemetry wired into customer-facing products
  • A vector/embedding store for unstructured knowledge
  • Evaluation datasets that represent real production use

The data flywheel in practice

Sources apps · docs · ops Ingest & Clean pipelines · quality Intelligence RAG · models · agents Value decisions · products Feedback signals · labels feedback loop refreshes the source data
Figure 6. Treat data as a renewable loop, not a one-time migration. The dashed return path is what separates AI-First from AI-enabled.

09 Prioritization

Choosing & sequencing use cases

The fastest way to fail is to chase shiny use cases. The fastest way to build momentum is to rank every candidate on two axes — business value and feasibility — and sequence ruthlessly. Win early, build credibility, then take on harder bets.

BUSINESS VALUE → FEASIBILITY (data + tech readiness) → STRATEGIC BETS High value · hard — plan & invest QUICK WINS ★ High value · easy — do first DEPRIORITIZE Low value · hard — avoid FILL-INS Low value · easy — automate cheaply Support deflection Code assist Doc summarization Autonomous ops agent AI-native product line Meeting notes Niche custom model sequence: quick wins fund the bets
Figure 7. Plot candidates, then sequence. Quick Wins first to build credibility & fund the program; reinvest into Strategic Bets.

Scoring template

Score each use case (1–5)Value factorsFeasibility factors
AskRevenue / cost / risk impact · strategic fit · # of people affected · customer impactData availability & quality · technical complexity · clear owner · change-mgmt difficulty · regulatory risk
The lighthouse principle

Pick 2–3 visible, high-confidence quick wins as "lighthouse" projects. Their job is as much to build belief and a reusable pattern as to deliver ROI. Publicize them internally — momentum is a strategy.

10 People

Talent, skills & culture

AI transformation is ~20% technology and ~80% people and process. The hardest barrier is rarely capability — it's adoption, trust, and fear. Treat culture and change management as a first-class workstream with a real budget, not an afterthought.

📚

Build AI literacy at every level

Tiered enablement: awareness for everyone, fluency for power users, depth for builders. Make "prompting" and "working with agents" core skills like email once was.

🧭

Address the fear directly

People won't adopt what they think will replace them. Be explicit: AI handles toil; humans move up to judgment, relationships, and creativity. Reskill visibly; redeploy, don't just cut.

🏆

Align incentives

What gets measured and rewarded gets adopted. Put AI-adoption and outcome metrics into team goals; celebrate champions; make "did you try AI first?" a norm.

The talent strategy: buy, build, borrow

LeverUse forNotes
Build (upskill)The majority of your workforceCheapest and most scalable; your people know the business. Invest heavily here.
Buy (hire)Scarce specialists: ML/LLMOps, AI PM, applied scientistsExpensive and competitive; hire for platform/leadership roles that multiply others.
Borrow (partners)Bursty needs, the leading edge, knowledge transferUse to accelerate & de-risk early, but build internal capability — don't create permanent dependence.

The human–AI division of labor

AI does

  • High-volume, repetitive throughput
  • First drafts & synthesis at scale
  • Pattern detection across large data
  • 24/7 routine execution & monitoring

Humans do

  • Judgment on ambiguous, high-stakes calls
  • Accountability & ethical decisions
  • Relationships, empathy, persuasion
  • Setting goals, taste, and reviewing AI output
The mindset shift

The goal is not "humans vs. AI" but the centaur model — humans directing and reviewing AI, each doing what they're best at. The most valuable employees become great orchestrators of AI, not competitors to it.

11 Trust

Governance, risk & responsible AI

Governance is not the brake on AI — it is what lets you go fast safely. Without trust, adoption stalls and a single incident can set you back years. Build guardrails that enable confident scaling, and treat regulation (e.g. the EU AI Act and emerging frameworks) as a design input from day one.

RESPONSIBLE AI OPERATING SYSTEM ⚖️ Accountability named owners, human-in-loop 🔍 Transparency explainable, disclosed AI use 🤝 Fairness bias testing, inclusive data 🔒 Security data & prompt injection defense 🛡️ Privacy consent, PII, data residency 📋 Compliance EU AI Act, sector rules
Figure 8. Six pillars of a responsible-AI operating system. Stand these up alongside your first production use cases — not after an incident.

Practical guardrails to put in place

  • AI use policy — what's allowed, what data can go where, approved tools.
  • Risk tiering — classify use cases by impact; higher tiers get more oversight & human review.
  • Evaluation & monitoring — test for accuracy, bias, and drift before and after launch.
  • Human-in-the-loop — define where a human must approve before action is taken.
  • Incident response — a plan for when AI gets it wrong (it will).
  • Model & data inventory — know what's running, on what data, owned by whom.

Top AI-specific risks to manage

  • Hallucination — confident wrong answers; mitigate with RAG, citations, review.
  • Data leakage — sensitive data sent to third-party models or exposed in outputs.
  • Prompt injection — adversarial inputs hijacking agent behavior.
  • Bias & discrimination — especially in hiring, lending, and other regulated decisions.
  • Over-reliance / automation bias — humans rubber-stamping AI without scrutiny.
  • Vendor & concentration risk — lock-in to a single model provider.

12 Execution

The transformation roadmap

Transformation runs in overlapping waves, not a single big-bang. The arc moves from foundation → momentum → scale → reinvention over roughly 18–24 months, though timing varies by size and starting maturity. Each phase has a clear goal and exit criteria.

PHASE Q1 Q2 Q3–Q4 Y2+ 0 · Mobilize Sponsor · audit · North Star 1 · Foundation Data platform · governance · quick wins 2 · Momentum Operationalize · MLOps · upskill · measure ROI 3 · Scale Federated pods · portfolio · cross-functional 4 · Reinvent AI-native products first prod use case ROI proven at scale
Figure 9. Phases overlap deliberately — you start quick wins while building the platform. Gate progression on exit criteria, not the calendar.

Phase 0 — Mobilize (weeks 1–6)

Secure an executive sponsor and budget. Run an AI-readiness audit (data, skills, tech, risk). Define the North Star and pick lighthouse use cases. Exit: funded mandate & a prioritized backlog.

Phase 1 — Foundation (months 2–6)

Stand up data access, a thin AI platform, security & governance basics, and acceptable-use policy. Ship 2–3 quick wins. Exit: first use case running reliably in production with measured impact.

Phase 2 — Momentum (months 4–12)

Operationalize: MLOps/LLMOps, evaluation, reusable patterns. Launch broad upskilling. Track ROI rigorously and publicize wins. Exit: 5–10 production use cases & a repeatable delivery playbook.

Phase 3 — Scale (months 9–18)

Roll out the federated operating model with embedded pods. Manage a portfolio. Embed AI fluency across functions. Exit: AI in every major function with attributable enterprise value.

Phase 4 — Reinvent (months 18+)

Use the platform and data moat to launch AI-native products and business models. Defend and accelerate the flywheel. Exit: ongoing — this becomes business as usual.

13 Value

Metrics, ROI & value capture

"What you don't measure, you can't scale or defend in the budget cycle." Track value at three levels — adoption, performance, and business impact — and connect every initiative to a hard outcome. Beware vanity metrics (number of pilots) over value metrics (dollars moved).

LayerQuestion it answersExample metrics
AdoptionAre people actually using it?Active users, % of eligible workflows using AI, frequency, retention
PerformanceIs the AI good enough?Accuracy/quality scores, latency, hallucination/error rate, cost per task
Business impactIs it creating value?Cost saved, revenue uplift, cycle-time reduction, CSAT/NPS, deflection rate
💵

Cost & productivity

Hours saved × loaded rate; throughput per FTE; cost-per-transaction before/after. The easiest ROI to prove early.

📈

Growth & revenue

Conversion lift, retention from better experience, new AI-native revenue lines, faster time-to-market.

🛡️

Risk & quality

Error reduction, compliance incidents avoided, faster detection — value that shows up as losses not incurred.

Measurement discipline

For every funded use case: capture a baseline before launch, define the one metric that matters, and assign a business owner accountable for the number. "Interesting" is not a unit of value; dollars, hours, and points of CSAT are.

Total cost of ownership — don't forget

14 Avoid

Common pitfalls & how to avoid them

Most failures are predictable and repeated. Knowing them in advance is the cheapest insurance you can buy.

⚠️ Technology-led, not value-led

Chasing AI because it's exciting, with no business problem. Fix: start from a pain/opportunity worth real money; the tech serves the outcome.

⚠️ Pilot purgatory

Endless POCs that never reach production. Fix: require a path-to-production plan before funding any pilot; kill the rest fast.

⚠️ Ignoring the data foundation

Expecting magic from messy, inaccessible data. Fix: fund the unglamorous data work; it's the actual bottleneck.

⚠️ Treating it as IT's problem

Delegating transformation to a tech team with no business authority. Fix: CEO-level sponsorship; business owns outcomes.

⚠️ Neglecting change management

Great tools nobody adopts. Fix: budget for training, champions, incentives, and addressing fear — explicitly.

⚠️ Governance as an afterthought

Moving fast until a public incident forces a freeze. Fix: build guardrails alongside the first use cases, sized to risk.

⚠️ Decorating old workflows

Bolting AI onto a broken process. Fix: redesign the workflow around what AI makes possible, then automate.

⚠️ Boiling the ocean

Trying to transform everything at once. Fix: sequence — lighthouse wins first, then scale the pattern.

15 Action

The 90-day quick-start

If you do nothing else, do this. A focused first quarter that creates a mandate, a foundation, and at least one real win — the trifecta that makes everything after it easier.

Days 1–30 · Align
  • Name an executive sponsor & AI lead
  • Run a rapid AI-readiness audit
  • Draft the North Star & guiding principles
  • Publish an interim AI use policy
  • Survey & rank 15–20 use cases
Days 31–60 · Build
  • Pick 2–3 lighthouse quick wins
  • Secure data access for those use cases
  • Stand up a thin, secure AI platform
  • Run literacy training for leaders
  • Define success metrics & baselines
Days 61–90 · Prove
  • Ship the first use case to production
  • Measure impact vs. baseline
  • Stand up a governance council
  • Tell the story internally — broadly
  • Lock the next quarter's roadmap & budget
Day 90 success looks like

A funded mandate with a clear North Star · a secure platform & basic governance in place · at least one AI use case live in production with measured value · a literate leadership team · and a credible, sequenced roadmap for the next 12 months.

Twelve questions for the leadership team

  1. If a competitor rebuilt our business AI-First tomorrow, where would they beat us?
  2. What is our proprietary data advantage — and are we capturing or wasting it?
  3. Who, by name, owns AI transformation and has the authority to fund it?
  4. Which 3 workflows, if AI-redesigned, would move the most value?
  5. Where are employees already using ungoverned "shadow AI," and why?
  6. What would we stop doing if AI handled the routine 80%?
  7. How will we measure value — and what's our baseline today?
  8. What's our policy on data leaving our boundary to third-party models?
  9. How do we keep humans accountable for high-stakes AI decisions?
  10. Are we building the differentiator and buying the commodity?
  11. What's our plan for the people whose roles will change most?
  12. What would make us proud — and what would make headlines for the wrong reasons?

16 Reference

Appendix

A. Model selection guide

If you need…Reach for…
Hardest reasoning, agents, complex coding, high-stakes accuracyA frontier "max" model (e.g. top-tier Claude Opus-class). Pay for quality where it matters.
The everyday workhorse — strong quality at good speed/costA balanced mid-tier model (Sonnet-class). The default for most production traffic.
High-volume, latency-sensitive, simple tasksA small/fast model (Haiku-class) — classify, extract, route cheaply.
A consistent format/style or a cheaper domain specialistA fine-tuned smaller model on your examples.
Current, proprietary, citable knowledgeAny capable model + RAG — retrieval beats memorization for facts.

Use a model-router pattern: send each request to the cheapest model that meets the quality bar. Re-benchmark quarterly — the frontier moves fast and prices fall.

B. Glossary

  • Agent — an AI system that plans and takes multi-step actions using tools, not just generating text.
  • RAG — Retrieval-Augmented Generation; grounding model answers in retrieved documents.
  • Fine-tuning — adapting a model's weights on your own examples.
  • Context window — how much text a model can consider at once.
  • Embedding — a numeric vector representing meaning, used for semantic search.
  • Compound AI system — multiple models, tools, and retrieval orchestrated together.
  • LLMOps — practices for deploying, monitoring & evaluating LLM apps in production.
  • Hallucination — a confident but false model output.
  • Prompt injection — malicious input that hijacks an AI's instructions.
  • Human-in-the-loop — requiring human approval at defined decision points.
  • Eval — a test suite measuring an AI system's quality on representative cases.
  • Token — the unit models read/write; billing and limits are measured in tokens.

C. The one-sentence summary

AI-first means leadership owns the system change—strategy, data, tech, people, governance—starting with quick wins, then reinvention, so your data → value loop compounds.

17 Bibliography

References & sources

Citations behind the hero stats, maturity model, flywheel, org design, stack choices, governance, and ROI framing. Percentages move every survey cycle—open the publisher’s latest report before you paste a number into a board deck.

This note synthesizes public research and practice; diagrams are original unless noted. Not legal, tax, or compliance advice—confirm current regulatory text with counsel.

Industry adoption & economics (hero, §01, §03)

  1. McKinsey & Company, “The state of AI.” Annual global survey—adoption, scaling, and value capture. mckinsey.com/state-of-ai
  2. Boston Consulting Group, AI maturity / value research (e.g., “Where’s the value in AI?” and maturity index materials). Use the edition year on the PDF you download.
  3. Gartner research on AI deployment, pilot-to-production rates, and maturity—access via subscription or public summaries; match headline stats to the underlying note.
  4. Stanford HAI, AI Index Report. Annual trends in capabilities, economics, and deployment. aiindex.stanford.edu
  5. Epoch AI and provider pricing pages for inference-cost trends cited in §01 (re-benchmark quarterly).

Strategy, transformation & operating model (§05–06, §12–15)

  1. Kotter, Leading Change. Harvard Business Review Press—executive sponsorship and sequencing (§12 roadmap, §15 quick-start).
  2. Prosci ADKAR model—change management and adoption (§10 people, §14 pitfalls). prosci.com
  3. McKinsey, “The platform play” / federated operating models—hub-and-spoke patterns for digital and AI capability (§06).
  4. Forsgren, Humble & Kim, Accelerate. IT Revolution—DORA metrics and platform-team delivery (§06–07).
  5. Skelton & Pais, Team Topologies. Platform vs stream-aligned teams; RACI adjacency for AI platform orgs.
  6. Christensen et al., “Know Your Customers’ Jobs to Be Done.” Harvard Business Review, 2016—value framing for use-case selection (§09).

Data flywheel & network effects (§04, §08)

  1. Parker, Van Alstyne & Choudary, Platform Revolution. W. W. Norton—feedback loops and data network effects.
  2. Shapiro & Varian, Information Rules. Harvard Business Press—data as strategic asset and versioning (§08 data strategy).
  3. Amazon shareholder letters (Bezos era)—flywheel thinking; useful analogy for §04, not a literal template for every firm.

AI systems, RAG & compound architectures (§07, §16 appendix)

  1. Zaharia et al., “The Shift from Models to Compound AI Systems.” Berkeley BAIR Blog, 2024—compound AI systems framing in §07. BAIR blog
  2. Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.” NeurIPS 2020. arxiv.org/abs/2005.11401
  3. Es et al., “Ragas: Automated Evaluation of Retrieval Augmented Generation.” 2023. arxiv.org/abs/2309.15217
  4. Huyen, Designing Machine Learning Systems. O’Reilly—data pipelines, deployment, monitoring (§07–08).

Prioritization, metrics & value capture (§09, §13)

  1. Amplitude, North Star Playbook—North Star metric design (§05, §13). amplitude.com/north-star
  2. Rodden et al., HEART framework (Google / large-scale UX measurement)—experience metrics alongside business KPIs.
  3. ICE / RICE prioritization—product scoring for backlog and use-case ranking (§09 scoring template); see Intercom (RICE) and lean product practice.

Governance, risk & responsible AI (§11)

  1. European Union, Artificial Intelligence Act (Regulation (EU) 2024/1689). EUR-Lex
  2. NIST, AI Risk Management Framework (AI RMF 1.0). nist.gov/ai-rmf
  3. ISO/IEC 42001—AI management system standard (purchase from ISO or your national body).
  4. OWASP Top 10 for Large Language Model Applications. OWASP LLM Top 10
  5. OECD AI Principles. oecd.ai
  6. Partnership on AI / NIST Generative AI Profile—responsible deployment practices; check current editions.

Talent & human–AI collaboration (§10)

  1. World Economic Forum, Future of Jobs reports—skills shift and reskilling context. weforum.org/reports
  2. Brynjolfsson & McAfee, The Second Machine Age—automation economics (background for centaur / division-of-labor framing).
Hero stat strip

The ~70% pilot-stall figure, productivity multipliers, and “<1% AI-mature” headlines should be re-checked against McKinsey, BCG, and Gartner’s current publications—titles and years change.