stage 02 · find the value

Opportunity Discovery

Most AI roadmaps start with capabilities looking for a use. Opportunity discovery starts the other way: with the value trapped behind a bottleneck, sized and prioritized before anything gets built.

By · V-Squared AI

Most AI roadmaps start with capabilities looking for a use. Opportunity discovery starts the other way: with the value trapped behind a bottleneck, sized and prioritized before a line of anything gets built.

Point capability at trapped value

It's the discipline of finding where customers can't get what they need, or where the business can't capture value it creates, and turning that into a prioritized set of opportunities worth pursuing. AI makes it cheap to build the wrong thing quickly; without discovery you get a portfolio of impressive pilots that don't move revenue. With it, you get an opportunity pipeline tied to how customers actually decide what they need and what they'll buy.

Where strategy earns its keep

Done well, discovery is the filter that keeps the roadmap pointed at trapped value instead of available technology. It targets the understand-the-customer and align-to-customer-needs bottlenecks. The capability is one an enterprise can build in-house, and it's defined at Endgame Engineering and taught through the Opportunity Discovery course on High ROI AI.

The filter that keeps the roadmap pointed at trapped value instead of available technology.
the engagement · vsquaredai.com

Run opportunity discovery with V-Squared

V-Squared facilitates opportunity discovery for enterprise leaders, surfacing, sizing, and sequencing the use cases that become your roadmap.

Work with V-Squared
the evidence · earnings calls, july 2026

The vendors built discovery into the engagement itself.

This stage says the roadmap has to point at trapped value rather than at available technology. In July 2026 the largest vendors described doing exactly that inside their customers, and reported what it produced.

peer-reviewedsurvey researchcompany filingsmarket dataearnings call
Microsoftearnings callFY26 Q4 · Jul 29, 2026

Two projects per customer is a discovery pattern

Microsoft announced Frontier Company, embedding 6,000 industry and engineering experts with customers to co-design and continuously improve AI systems. It ran quietly for a year first, across 330 projects with 164 customers.

Source: Microsoft FY26 Q4 earnings call, July 29, 2026.
what it confirms

That ratio is roughly two projects per customer, not one large program. Find the trapped value, size it, prove it, go again. The largest software company on earth built its biggest new organization around running discovery inside the customer's business.

Metaearnings callQ2 2026 · Jul 29, 2026

A well chosen opportunity, with the number attached

Movida, a Brazilian rental car company with roughly 400 locations, put a business agent on WhatsApp to run selection, pricing, and payment. Daily bookings rose 44% year over year in a single month, with 85% of conversations resolved without a human. More than a million businesses now use Meta Business Agents weekly.

Source: Meta Q2 2026 earnings call, July 29, 2026.
what it confirms

The booking flow was the right target because that is where customers could not get what they needed. The same technology pointed at an internal reporting workflow produces a demo. Choosing where to point it is the whole stage.

Alphabetearnings callQ2 2026 · Jul 2026

The moat is the customer's problem, not the model

Asked what the moat is when every competitor has comparable models, Pichai said the model is just an ingredient in the solution, and that customers need their own data and trajectories kept confidential with nothing flowing back.

Source: Alphabet Q2 2026 earnings call, July 2026.
what it confirms

When the model is an ingredient, the differentiator is knowing which problem to point it at and what someone will pay to have solved. That is opportunity discovery described by a vendor who no longer expects to win on capability alone.

Metaearnings callQ2 2026 · Jul 29, 2026

Underwriting the outcome requires knowing its size first

On monetizing business agents, Zuckerberg said Meta expects to evolve these products toward its advertising model, where businesses only pay when Meta delivers results for them.

Source: Meta Q2 2026 earnings call, July 29, 2026.
what it confirms

Pay-on-results only works if you identified the value in advance and can measure whether you delivered it. A vendor willing to underwrite the outcome has done discovery properly. That is the standard this stage sets before anything gets built.

Sources: Alphabet Q2 2026, Microsoft FY26 Q4, Meta Q2 2026, and Amazon Q2 2026 earnings calls and releases, July 2026. Figures are as stated by company executives. Amazon reported after market close on July 30, so Amazon figures come from the release and initial call remarks rather than a full transcript.

the wider evidence · what actually kills the budget

The failure is upstream of the build, and it has been measured repeatedly.

If AI programs failed on engineering, better engineers would fix them. Four independent datasets point somewhere else. The problem was chosen badly, defined vaguely, or never connected to value anyone would pay for.

peer-reviewedsurvey researchcompany filingsmarket dataearnings call
S&P Global Market Intelligencesurvey research2025 · 1,000+ enterprises

Abandonment more than doubled in a single year

The 2025 Voice of the Enterprise survey of more than 1,000 organizations across North America and Europe found 42% of companies abandoned most of their AI initiatives, up from 17% the previous year. The average organization scrapped 46% of its proofs of concept before they reached production.

Source: S&P Global Market Intelligence, Voice of the Enterprise: AI & Machine Learning 2025, as reported by CIO Dive and Fortune.
what it confirms

These were funded programs with staffed teams and executive sponsors. The doubling happened in the year the models got dramatically better, so the models are not what changed. Something upstream is selecting the wrong work.

RAND Corporationsurvey research2,400+ AI initiatives

Problem definition ranks above every technical cause

RAND's analysis found more than 80% of AI projects failing to deliver intended business value, roughly twice the failure rate of comparable IT projects without AI. The root causes are led by misunderstood problem definition and a technology-first mentality, ahead of data and infrastructure issues.

Source: RAND Corporation analysis of enterprise AI project outcomes.
what it confirms

Misunderstood problem definition is not a polite way of saying weak engineering. The team built what it was asked for, and what it was asked for was not the trapped value. That is the specific failure this stage removes.

MIT Project NANDAsurvey research2025 · 300+ deployments

The organizations that ran the most pilots converted the fewest

The GenAI Divide study found that enterprises above $100 million in revenue lead in pilot count and assign the most staff to AI, yet report the lowest rates of pilot-to-scale conversion. Mid-market companies moved faster, with top performers averaging 90 days from pilot to full implementation. Purchasing from specialized vendors succeeded roughly 67% of the time against about a third for internal builds.

Source: MIT Project NANDA, The GenAI Divide: State of AI in Business 2025.
what it confirms

Volume of pilots is inversely related to conversion. That is what happens when discovery gets replaced by a backlog of ideas, and it is the argument for sizing and sequencing opportunities before anything is funded.

Gartnersurvey research2024 to 2026 forecasts

The named causes are commercial, not technical

Gartner forecast that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, naming poor data quality, inadequate risk controls, escalating costs, and unclear business value. The firm separately forecast that 60% of AI projects unsupported by AI-ready data would be abandoned.

Source: Gartner research on generative AI project outcomes, 2024 to 2026.
what it confirms

Unclear business value sits in the same list as data quality and cost. It is the only one you can eliminate before a single engineer is assigned, and eliminating it is what makes the other three worth solving.

Failure rate definitions differ across these studies, so the agreement on causes matters more than the exact percentages. All four independently place the cause upstream of the build.