A thesis on the AI Accelerator, where vertical AI value actually lives, and the decade that will decide which companies compound and which disappear. Fully sourced. Intellectually honest about its own weak points.

Executive summary

The research is now unambiguous. Between 42 and 95 percent of enterprise AI initiatives fail to produce material value, depending on how the question is framed. MIT's NANDA project put the number at 95 percent. S&P Global Market Intelligence reports the abandonment rate rose from 17 percent in 2024 to 42 percent in 2025. BCG finds that only 25 percent of firms generate meaningful value from AI, while the other 75 percent spend, iterate, and stall.

These numbers describe the same phenomenon. There is a gap between AI capability and AI value capture in operating companies. We call it the Implementation Gap. This paper explains what it is, why traditional delivery models cannot close it, and what kind of firm can.

1. The Implementation Gap is real and measurable

The gap between the frontier of AI capability and the rate at which operating companies capture value from AI has widened, not narrowed, over the last 24 months. This is the opposite of what most stakeholders expected.

The evidence:

  • MIT NANDA, 2025: 95 percent of generative AI pilots at large enterprises fail to deliver measurable P&L impact.
  • S&P Global Market Intelligence, 2025: 42 percent of U.S. companies abandoned most of their AI initiatives in 2025, up from 17 percent in 2024.
  • BCG, Build for the Future 2025: 60 percent of executives report no material value from their AI programs. Only 25 percent of firms in the study generate meaningful value.
  • IDC / Lenovo, 2025: for every 33 AI proofs-of-concept launched, four reach production. An 88 percent POC failure rate.
  • McKinsey State of AI, 2024 and 2025: adoption is up, but EBIT impact remains concentrated in a small minority of companies. The median company shows less than a 1 percent EBIT lift from AI.

The signal is consistent across independent studies and methodologies. Capability is compounding. Capture is not.

2. Why traditional delivery models cannot close the gap

The Implementation Gap is not a model problem. It is a delivery problem. The dominant delivery models were built for a different kind of work.

Big consulting

Deck-heavy, strategy-forward, thin on owned engineering. Great at framing and sequencing. Structurally incentivized to sell phases of work, not outcomes. The average enterprise AI strategy engagement produces a detailed roadmap and a POC that dies inside 90 days because nobody owns running it.

Agencies and dev shops

Good at building what a client specifies. Not good at diagnosing what the client should have specified. They can hand off a working pipeline. They cannot hand off business judgment. The result is a lot of well-built systems that automate the wrong process.

SaaS vendors

Sell a product. Every product is horizontal by design. The last mile of integration, data hygiene, change management, and workflow redesign is labeled "professional services" and priced as overhead. That last mile is where the value lives. Every study of AI outcomes points to the same conclusion: the model is 10 percent of the work.

Internal build

Some firms have the talent to build in-house. Most do not. Hiring a head of AI at operating-company compensation, in a market where frontier labs pay seven figures for senior engineers, is not a realistic play. For the thousands of U.S. companies in the $10M–$500M revenue band, internal build is structurally out of reach. And a hired engineer still does not have the domain expertise that makes vertical software valuable, that lives with the operator, not the model.

Each model captures a slice. None pairs real domain expertise with senior engineering and then ships and operates the result. The gap is the whole stack (expertise, build, and operation) held together by one team.

3. Where the value lives

The value in vertical AI does not live in the model. It lives in the expertise-bound workflows that generic AI cannot touch, the parts of an industry where the right answer depends on judgment a practitioner spent a career building. Every industry has work that is software-shaped but expertise-gated, and that is exactly where state-of-the-art vertical software pays off.

Five kinds of these workflows show up in nearly every operating company we have examined. Each is a place where encoding a subject matter expert's judgment into shippable software changes the economics:

  1. Revenue workflows. Quotes that do not close on time. Proposals that sit. Intake that drops leads. Follow-up that does not happen. Dollars walk out the door because nobody encoded how the best closer actually works.
  2. Labor-heavy workflows. Senior people doing junior-people work. Manual extraction, manual reconciliation, manual routing. Expertise spent on the part of the job that does not require judgment, when it could be encoded once and run forever.
  3. Process workflows. Swivel-chair work between systems. Data re-entered three times. Status calls that exist because nothing is instrumented. The tax of running an operation without a software spine.
  4. Risk and compliance workflows. Compliance work that happens by memory. Reviews that happen at the end, not at the edge. The expert knows where the liability hides; software can enforce it at every step.
  5. Decision workflows. Decisions made late, or with the wrong data, because the system that would support them does not exist. Forecasts that are stale. Pricing that is lagging. Capacity that is misallocated. The judgment exists; it is not yet in the software.

Each kind has recognizable patterns, a recognizable build, and a recognizable range of impact. The Vertical AI Index (our quarterly benchmark, launching Q3 2026) will publish those ranges by industry.

4. The AI Accelerator

The firm that can close the Implementation Gap pairs a subject matter expert with a senior AI engineering team, encodes that expertise into a product, and ships and operates state-of-the-art software for one industry at a time. It does four things, in sequence:

Partner

Start with the expert. Find the operator who already knows where the value lives, the practitioner whose judgment is the missing input, and pair them with an engineering team that can build at the frontier. The expertise is the moat. The partnership is what gets it out of one person's head.

Design

Translate that expertise into a shippable product. Map the expertise-bound workflow, decide what to encode, and produce a written spec that ranks the build by return and by time-to-value. This is the artifact a client should expect before any code is written. It is the opposite of a deck.

Build

Own the engineering end to end. Not "oversee a vendor." Not "architect the solution." Build it. Integrations, pipelines, models, interfaces, monitoring. The firm that builds the system is the firm that understands why it works.

Scale

Run the system until it compounds, then take the pattern to the next operator in the industry. In practice, operating the software is the piece every other delivery model skips. Most AI systems fail at day 90 not because the model drifts, but because nobody owned the discipline of keeping the integrations current, the prompts updated, the feedback loop closed, and the product adjusted as the business changes.

Partner. Design. Build. Scale. That is the shape of the AI Accelerator. It is not a consultancy, not an agency, not a product, not a staffing firm. It is a fourth thing.

5. Why this shape compounds

The AI Accelerator compounds because the design, the build, and the operation share the same instrumentation. Every build produces structured data about which workflow it targeted, at what magnitude, with what software, at what cost, and with what outcome. That structured data is the flywheel.

Over time, the firm gets better at:

  • Recognizing the high-value workflows in an industry from a half-hour conversation.
  • Pricing builds against a real distribution of outcomes, not a fantasy.
  • Shipping the second version of a system faster than the first, because the patterns repeat across operators in the same vertical.
  • Publishing benchmarks the industry does not have, because nobody else is instrumented to collect them.

The writing, the index, and the work feed each other. The firm that compounds is the firm that publishes the evidence.

6. What this thesis gets wrong, honestly

We will not pretend this framing is complete. A few weak points we are willing to name:

  • Enterprise-scale work is a different animal. The AI Accelerator model described here is calibrated for operating companies in roughly the $10M–$500M revenue band. Above that, organizational politics, procurement, and regulatory complexity change the economics meaningfully. We do not claim this framing generalizes.
  • Not every workflow is AI-addressable. Strategy problems, product-market-fit problems, and leadership problems cannot be encoded away. We decline engagements where the underlying issue is not an operational one a piece of software can carry.
  • The data on AI outcomes is still young. The highest-quality studies we cite are 12–18 months old. The next two years will either confirm the pattern or complicate it. We will update this thesis as the evidence does.
  • Our own sample is biased. We see the builds that reach us. We do not see the ones that never call. Our published benchmarks will be transparent about this.

7. What happens next

Over the decade between 2026 and 2035, the gap between companies that implement AI well and companies that do not will translate directly into differences in margin, growth, and valuation. The firms that close the Implementation Gap early will compound. The firms that do not will be acquired, automated around, or disappear.

The AI Accelerator is not the only way to close the gap. It is the way we are betting on. The structure of this firm (expertise first, engineering-led, operating the software inside the client) is calibrated for the shape of the problem as we actually see it, week after week, across industries.

This paper is the first long-form argument. The Vertical AI Index, launching Q3 2026, will be the first dataset. The builds we take in 2026 and 2027 will be the first proof points. We expect to be wrong about some of this, and we will publish the corrections when we are.

If you are a subject matter expert with a workflow that should be software, and an executive on the inside who owns the change, we should probably talk.

Sources

  • MIT NANDA project, State of AI in Business 2025. (2025)
  • S&P Global Market Intelligence, AI initiative abandonment survey. (2025) Link
  • BCG, Build for the Future 2025: Are You Generating Value from AI? Link
  • IDC / Lenovo, AI POC to production study. (2025)
  • McKinsey, The State of AI. (2024 and 2025 editions)
  • Gartner, Hype Cycle for Artificial Intelligence. (2024 and 2025 editions)