AI in Business: Skip the Hype, Build the Machine

Short version: most corporate AI pilots flop. A tiny minority are compounding value every week. Your job is to join the tiny minority—by fixing data, picking the right workflows, and shipping faster than your org’s politics. Not sexier. Just better.

The uncomfortable truth (and why it’s good news)

MIT’s new Project NANDA report says 95% of enterprise gen-AI pilots fail to deliver measurable ROI [1][2]. Markets noticed. Execs noticed. You should too. Because the same report implies a small group is pulling far ahead. Be one of them.

This isn’t the only study pointing that direction. BCG found only 4% of companies are creating substantial value with AI; most never get past proof-of-concept [3]. Translation: value is real, but rare—and that rarity is your edge if you execute.

Meanwhile, real revenue is showing up in parts of the ecosystem: Anthropic reported its ARR jumping from roughly $1B to $5B in eight months—wild by any historical software standard [4][5]. That doesn’t make your roadmap easier, but it proves customers are paying when products work.

What the winners do differently

1) They start with data, not demos. HBR’s drumbeat is clear: AI returns correlate with data quality, integration, and governance, especially for messy unstructured data (docs, emails, notes). If your data is scattered or dirty, your AI will be too. Build the “digital core” first [6][7].

2) They pick boring problems with measurable P&L. The MIT and BCG work both suggest quick wins skew toward operations/back-office: document handling, case triage, finance, procurement, customer support. Start where minutes add up to hours and quality errors are costly [1][3].

3) They ship constantly. Shai Wininger (Lemonade co-founder) made the spicy case: if you want AI to change your company, “burn down” the old stack and ship relentlessly [8]. You don’t have to copy the pyrotechnics to copy the cadence. Elite DevOps teams deploy many times per day—that rhythm compounds learning and value [9].

4) They buy and build (on purpose). Legacy org? A credible vendor can get you out of pilot purgatory faster than an 18-month internal science project. Palantir’s AIP, for example, has public case studies showing large-scale rollouts (think thousands of locations) on tight timelines—useful pattern if your culture or stack fights change [10].

5) They empower “wild ducks.” Big companies need protected iconoclasts who cut across silos—IBM literally called them “wild ducks.” Your AI program needs a few, with air cover from the top, to break inertia and ship [11].

A practical, 90-day plan (that actually survives committee)

Weeks 1–2: Inventory pain with a stopwatch. Walk the floor. Where do humans open 50 docs to find 5 numbers? Where do queues stall? Pick 3 workflows with: (a) high volume, (b) clear acceptance criteria, (c) auditable outcomes. Write baseline KPIs (handle time, SLA, error rate, dollars recovered). Remember the 95% stat—define success up front [1].

Weeks 2–4: Stand up the “thin” data layer. You don’t need a perfect lake; you need reliable, governed access to the fields the workflow uses—plus a safe place to store AI outputs. Capture provenance for every field. (HBR: unstructured data quality is often the hidden blocker—fix it early [6].)

Weeks 3–6: Ship a “walking skeleton.” Prototype that does one thing end-to-end for real users under supervision. Log every decision. Red-team prompts. Add guardrails (approved tools, data scopes, escalation rules). If you’re a legacy shop, consider piloting on a battle-tested platform (e.g., AIP) to compress time-to-value while your data team modernizes the core [10].

Weeks 6–10: Tighten the feedback loop. Adopt DORA-style habits: deployment frequency up, lead time down, change-failure rate monitored. Small, reversible releases. The goal is daily learning cycles—even if you can’t hit “many per day” yet [9].

Weeks 8–12: Prove the business case, not the demo. Publish a one-pager with before/after on those KPIs (time saved, errors cut, dollars reclaimed). If it’s good, roll the same pattern to the next 3 workflows. If it’s not, kill it and move on. Multiple industry surveys show many gen-AI projects stall after POC—celebrate fast kills and redeploy talent [12].

Two mini case studies worth copying

Lemonade: years of AI-first habits show up in day-to-day operations. By mid-2024, ~30% of customer interactions were already handled by AI, per earnings-call coverage [13]. That’s what compounding looks like when you integrate AI into the work, not the slide deck.

Vendor-accelerated lift-off: Palantir’s AIP has public examples of large deployments on tight clocks (e.g., retail footprints in thousands of locations). Even if you don’t pick Palantir, the lesson stands: standard tooling + forward-deployed builders can beat bespoke experiments [10].

The bubble talk (and why it shouldn’t paralyze you)

Yes, there’s froth. The MIT 95% failure stat rattled investors [1][2]. At the same time, winners are scaling real revenue (see Anthropic’s ARR) [4]. Your takeaway shouldn’t be “wait.” It should be “execute where ROI is provable—faster.”

Your first three moves, starting Monday

1. Name an owner with teeth. One accountable leader who can pull people and systems across silos—your internal “wild duck” [11].

2. Pick one workflow you can measure in dollars. Back-office is fine—especially high-volume document and data tasks [3].

3. Set a shipping tempo. Weekly releases minimum; instrument with DORA metrics; collect before/after KPIs [9].

Do that, and you’re not in the 95%. You’re building the compounding machine.

Why this post, and what’s next

This piece riffs on a clip we’re embedding in the post—where we talk through Wininger’s “burn it down” argument, the MIT stat, Palantir vs. in-house, and why iteration speed is the moat [8].

---

Sources

1. Fortune — Coverage of MIT Project NANDA’s GenAI Divide report (95% of gen-AI pilots fail).

2. Yahoo Finance — Coverage of the same MIT Project NANDA findings and market reaction.

3. Boston Consulting Group — “Where’s the Value in AI?” (Oct 2024) — Only ~4% report substantial value.

4. TechCrunch — “Anthropic’s ARR surges to ~$5B,” September 2, 2025.

5. Yahoo Finance — “Anthropic says ARR climbed from ~$1B to ~$5B in eight months,” September 2, 2025.

6. Harvard Business Review — “Unstructured Data Is Your AI Bottleneck” (2024).

7. Harvard Business Review (with Accenture) — “Build a Digital Core for an AI-Ready Enterprise” (2025).

8. Shai Wininger (X) — Thread on “burning down” the old stack to ship AI faster.

9. DORA — Accelerate: State of DevOps Report 2024 (deployment frequency, lead time, change-failure rate).

10. Nasdaq/Zacks coverage — Palantir AIP enterprise deployments and large-scale rollouts.

11. IBM Archives — Thomas J. Watson Jr.’s “Wild Ducks” parable from A Business and Its Beliefs (1963).

12. McKinsey — The State of AI in 2024 (most orgs stuck in pilots; few capture material value).

13. PYMNTS — “Lemonade: ~30% of customer interactions handled by AI,” July 31, 2024.