Contents
Executive thesis
The biggest AI opportunity is not making people faster. It is removing the constraints that keep revenue, capacity, cash and margin from moving.
AI capability is getting cheaper at a pace few operating plans assume. Epoch AI estimates the cost of reaching a fixed level of performance has fallen roughly 47% per quarter since 2023, about 13-fold a year across the benchmarks it studied. Yet Census-based NBER data shows only 18% of US firms used AI in at least one business function between November 2025 and January 2026. Among adopters, 57% used it in three or fewer functions, and two-thirds used it only to assist people rather than to change how work moves.
That gap between rapidly improving capability and shallow operational redesign is the opportunity, and it explains why generic productivity rarely reaches the P&L. Time saved becomes money only through a capture mechanism: avoiding a hire, selling constrained capacity, removing a cost or leakage, or accelerating revenue and cash. The strongest cases in this report sit at constraints: a quote desk, a production line, a truck network, a transaction queue, the yield of a material.
Three historical barriers are weakening at the same time: variation in physical work, legacy software without usable interfaces, and the cost of writing custom software. Reliability and verification still bind, which is why bounded workflows with exception handling are investable and indiscriminate autonomy is not. Evidence quality varies enormously across this field, so every number here carries its outcome status and its source provenance.
Capability
Cost of a fixed capability level, indexed
Adoption
US firms, November 2025 to January 2026
From productivity to EBITDA
Where saved time goes without a capture mechanism. Illustrative, not to scale.
What a visitor should leave understanding
- AI capability is moving much faster than business transformation. The cost curve and the adoption data point in opposite directions.
- Generic productivity is economically meaningless until management creates a capture mechanism. A 7,137-worker randomized study shows why time savings and financial value are not synonyms.
- The best opportunities sit at bottlenecks, material yield, response latency, asset utilization and transaction queues, rather than in the largest pools of labor.
- Three barriers are weakening at once: variation, difficult software interfaces and the cost of creating custom software. Reliability and verification have not disappeared.
- The right question has changed. Not where can we use AI, but where does this business lose money, capacity, cash, speed or growth because a human still has to interpret, decide, reconcile or route something.
How to read the evidence
Every quantitative claim carries two independent labels: what kind of outcome it is, and where it comes from. A vendor-published result is not automatically false, and an observed result is not automatically audited.
Outcome status
- ✓Observed
- Documented as having occurred.
- ƒEstimated
- Inferred, projected or calculated from explicit assumptions.
- ◇Hypothetical
- Illustrative economics. Not a realized outcome.
Source provenance
- Independent or academic
- Research groups, peer-reviewed or working papers, randomized studies.
- SEC filing
- Financial results filed by a public company.
- Company-reported
- Stated by the company itself, including management commentary.
- !Customer-confirmed, vendor-published
- A customer's statement appearing in vendor material.
- !Vendor-only
- Reported only by the seller, often anonymous.
Fifteen insights
Each one states the thesis, what is surprising, what the evidence actually says, why it matters economically, why private equity should care, and what to do on Monday.
Five real-world stories
Five businesses, five different ways AI changed the economics, with materially different evidence quality. Each traces the path from the old constraint to a financial capture mechanism.
Beliefs challenged
Five conventional beliefs, plus a sixth for professional services, and what the evidence suggests instead.
| Conventional belief | What the evidence suggests | PE decision implication |
|---|
Five actions for the next 30 days
Each action has an acceptance criterion and a value proof, so a board can tell whether it worked.
Replace the board AI scorecard
Stop leading with licenses, prompts, training counts or self-reported time savings. Every initiative should report these fields instead.
Evidence notes and caveats
Working with Local AI
This report's rule is simple: start with a bounded workflow that has a visible financial mechanism. That is where we work, alongside the operating team and the sponsor.
Where we would start. One workflow from the matrix with direct economic capture: quote-to-order, document-to-ERP, freight-bill audit, document triage or another job that can prove value inside the systems the team already runs. We agree the baseline and the capture mechanism first, so the result can be read in EBITDA, working capital or capacity rather than hours saved.
How it runs. On the cloud the company already uses, with code, models, integrations and deployment patterns owned by the company from day one. If the workload later earns it, the same system can move onto infrastructure the company controls.
Live capabilities in 21 days
Discovery to live capabilities in 21 days, on the systems you already use. We start inside one real workflow and prove it against your own numbers.
Yours from day one
Production AI agents on the cloud you already use, with code, models, integrations and deployment patterns owned by you from day one. No vendor lock-in.
Built by operators
Our founder built machine learning at scale inside C.H. Robinson. A boutique senior team, not a generalist consultancy, not an SI, not a SaaS vendor.
Bring one real workflow. We will tell you, honestly, whether AI belongs in it yet, and if it does, how fast we can ship.
Sources
Every claim above links to one of these records. Where the original URL for a cited source was not available in the research materials, it is labeled rather than substituted with a vendor page.
| Title | Publisher | Date | Source type | Primary |
|---|