The evidence-based case for AI in the mid-market.

Executive Summary

The headline: AI capability has advanced faster than organizational value capture. Controlled studies repeatedly find double-digit task gains, yet representative firm data still finds little realized productivity impact at most companies. The missing layer is not usually the model; it is workflow redesign, data and integration work, adoption, governance, and converting released time into cash or throughput.

Who gets the biggest ROI: not a revenue band by itself, but companies with high-volume digital work, expensive labor/SaaS/outsourcing baselines, verifiable outputs, usable data, a powerful process owner, and the ability to redesign roles. Information/software, professional services, and finance lead adoption; customer operations, IT/cyber, software engineering, and document-heavy back offices produce the most repeatable evidence.

The $20M–$200M thesis is plausible but not yet causally proven. Mid-market firms can combine enough process volume to matter with shorter decision chains and CEO-led change. But large firms still adopt and scale more often, and there is no rigorous study showing that this revenue band earns the highest ROI. Runpoint should position the segment as a mechanism-backed deployment sweet spot, not a settled empirical fact.

What changed in 2026: code generation crossed a meaningful threshold. Built-from-scratch internal tools, integrations, tests, CRUD applications, and bounded code changes are now technically feasible with agentic systems and expert oversight. But end-to-end SaaS replacement is only conditionally solved: identity, permissions, migration, auditability, uptime, maintenance, and organizational ownership remain the hard part. A precise formulation is: code production is increasingly solved; software stewardship is not.

Best strategic position: “rent the models, own the workflows” is well aligned with the evidence. Model prices, packaging, and leaders change quickly; owned data, process logic, evaluations, integrations, and source code preserve switching power and grow in value over time. Sovereignty is an economic architecture, not a requirement to train or host the frontier model.

How AI ROI Progressed

2017–2021: narrow prediction and point solutions

AI programs were centralized, data-science-heavy, and aimed at forecasting, ranking, optimization, fraud, or recommendations. ROI was normally modeled as a case-level cost reduction, risk reduction, or incremental revenue lift. McKinsey adoption moved from 20% in 2017 to 56% in 2021, but the beneficiary was usually a workflow or decision–not the whole enterprise.

2022–2024: individual augmentation

Generative AI made language and coding broadly accessible. Laboratory and field studies showed meaningful gains in writing, customer support, consulting, and bounded coding. Enterprises often measured prompts, seats, hours saved, or pilot satisfaction. Those are leading indicators, not financial returns; unreallocated time and unaccepted output have a cash value of zero.

2024–2025: production integration and the pilot trap

Adoption surged, but scaling lagged. The most consistent lessons were to redesign workflows, focus on a small number of valuable use cases, build reusable data/governance foundations, and make senior leaders accountable. The common “95% of pilots fail” headline is directionally consistent with the value-capture gap but comes from a non-peer-reviewed MIT NANDA report with an opaque denominator; it should not anchor an investment case.

2026: agents, software production, and quality-adjusted economics

Frontier systems can now execute longer coding and computer-use tasks, while open-weight models drive inference prices down. The economic unit shifts from cost per token or seat to cost per accepted business outcome. The new constraints are orchestration, evaluation, review, system access, and who owns the evolving workflow.

Why Large Task Gains Have Not Become Large Company Gains

The studies are not contradictory; they measure different layers. A worker can complete a bounded task 25% faster while the company sees no P&L change because demand is fixed, review work grows, downstream bottlenecks remain, or management never removes cost or increases throughput.

The evidence supports a four-stage funnel: capability → accepted output → changed workflow → financial realization. Most public studies prove stage one or two. McKinsey, BCG, PwC, and NBER show that stages three and four remain scarce. In McKinsey's late-2025 survey, 88% reported AI use, roughly one-third had begun scaling, 39% reported any enterprise EBIT impact, and only about 6% met its high-performer definition. NBER's representative 2026 international survey found more than 80% of firms reported no productivity or employment impact over the prior three years, even as they expected future gains.

The practical implication is severe: time saved is inventory, not ROI. It becomes economic value only when it produces more accepted output, avoids hiring, reduces overtime or outsourcing, eliminates a license, shortens revenue cycle time, or enables role redesign.

The Company Archetype Most Likely to Win

The best predictor of ROI is a system of conditions, not company size. High-return candidates have: (1) frequent, repeatable digital transactions; (2) a costly current baseline; (3) observable quality and an accepted-output definition; (4) accessible systems and reasonably clean data; (5) a CEO or functional owner who can change the process; (6) a human checkpoint where error costs require it; and (7) enough demand to monetize released capacity.

Mid-market companies can be unusually attractive because they may have enterprise-like process volume without enterprise-length decision chains. Yet they also face thinner data/IT capacity. RSM's 2025 middle-market survey found 92% experienced implementation challenges, 62% said deployment was harder than expected, and data quality and expertise were leading problems. McKinsey found only 29% of companies below $100M had reached scaling, versus nearly half of companies above $5B. Those facts challenge any simple “smaller is better” claim.

Initiative Portfolio: Where ROI Is Most Repeatable

Prioritize workflows, not generic tools. The highest-confidence opportunities combine deterministic process control with AI for the parts that benefit from language, perception, classification, generation, or exception handling. Start with a baseline and an owner, then instrument acceptance, quality, cycle time, and downstream impact.

The ranking below favors repeatability and cashability for a $20M–$200M company, not novelty. Individual copilots are easy to deploy but often weak on cash realization. Owned workflow systems and selective SaaS replacement require more implementation work but can create durable, auditable economics when they remove a meaningful license, service, or labor baseline.

Code Generation and SaaS Replacement: The 2026 Inflection

Runpoint's conviction is directionally right and should be sharpened. Frontier coding agents can now explore repositories, plan changes, write code and tests, execute tools, and iterate. METR's task-horizon work shows the duration of software tasks frontier agents can complete at a given reliability has been rising rapidly. OpenAI and Anthropic report internal agent-authored code at unprecedented scale.

But the neutral evidence is still mixed. GitHub's bounded Copilot experiment found a 55.8% speed gain; METR's early-2025 study of experienced open-source developers found AI made work 19% slower in large, familiar repositories. METR's 2026 update believes newer tools are faster but could not yet estimate the gain robustly because of selection effects. DORA's core finding is that AI amplifies the quality of the engineering system around it.

Solved enough now: built-from-scratch internal tools, CRUD and reporting apps, integrations, tests, documentation, contained refactors, and replacement of thin commodity SaaS where requirements are stable and the organization can own operations.

Not solved: deciding what to build, legacy migration, identity and permissions, security, compliance, observability, uptime, user adoption, long-term maintainability, and accountability when behavior changes. The relevant comparison is not “agent cost versus developer salary”; it is three-year quality-adjusted total cost of ownership versus the incumbent workflow or SaaS contract.

Solved, Conditional, and Unsolved

“Solved” here means a competent mid-market company can deploy the pattern repeatedly with established controls and a credible path to value. It does not mean zero implementation work or zero risk. “Unsolved” means the business and operating model is not yet standardized enough to promise repeatable results.

A Better Way to Project and Measure ROI

Use three ledgers and refuse to blend them prematurely.

1. Cashable value

Eliminated licenses, contractors, overtime, BPO spend, losses, or avoided hires. This is the strongest evidence.

2. Capacity value

Hours saved × fully loaded hourly cost × realization rate. Use a realization rate of zero until management identifies the added output, avoided work, or role redesign that captures the time.

3. Growth and strategic value

Incremental contribution margin, faster sales or product cycles, improved retention/quality, and option value. Establish a counterfactual with A/B tests, phased rollout, matched teams, or pre/post controls.

Gross annual benefit = cash savings + realized capacity + incremental contribution margin + risk-adjusted loss reduction.

Total annualized cost = build + integration/data cleanup + change/training + human review/rework + model/tool usage + security/evaluation/compliance + maintenance + migration/exit.

ROI = (benefit − cost) ÷ cost. Also report three-year NPV, payback months, and cost per accepted output.

A board-quality business case should show downside/base/upside scenarios for adoption, success rate, realization rate, model mix, review burden, and maintenance–not a single-point estimate.

Token Economics: When the Meter Matters

Token cost is usually secondary for light, high-value assistance and can dominate low-value, high-volume, long-context agents. Current GPT-5.6 API list prices illustrate the range: Sol is $5 per million input tokens and $30 per million output tokens; Luna is $1 and $6. At 100,000 runs per year, a light 5k-input/1k-output task costs about $5.5k on Sol; a long-horizon 5M-input/1M-output run profile costs about $5.5M. Model routing changes that extreme case by roughly $4.4M before caching, tool fees, or volume terms.

The “under 150 seats are subsidized, then per-token” idea should be presented as vendor packaging, not an industry rule. Claude Team has a 150-seat ceiling and current enterprise plans are usage based; OpenAI enterprise plans include baseline access and sell additional credits, while Codex moved to token-based rates in 2026. Subsidy and bundle structure can temporarily hide inference economics, but renewal and heavy usage expose them.

Open-weight models improve sovereignty and can reduce unit cost, especially for stable, high-volume, privacy-sensitive tasks. They are not free: hardware/cloud capacity, serving, monitoring, evaluation, upgrades, and quality gaps matter. The correct break-even is quality-adjusted accepted outcomes, not raw tokens.

Annual inference cost = runs × [(input tokens × input price + cached tokens × cache price + output tokens × output price) ÷ 1M + tool fees].

Local break-even volume = annual fixed local stack ÷ (API cost per accepted outcome − local variable cost per accepted outcome).

AI Sovereignty: Rent the Models, Own the Workflows

The evidence favors architectural optionality. Model quality, price, context limits, hosting, and vendor packaging move too quickly to make one model the durable asset. The assets that grow in value are the workflow specification, source code, proprietary context, integrations, permissions, evaluation set, audit history, and user adoption.

For Runpoint, sovereignty should mean:

1. A model abstraction layer with routing by task, quality, latency, privacy, and price.
2. Owned workflow logic and source code in the client's environment.
3. Portable data and context with explicit access controls.
4. An evaluation harness that makes model switching measurable.
5. A deterministic reliable business record around probabilistic steps.
6. An exit path for SaaS, model, and implementation vendors.

This position is strongest when framed as financial risk management and intellectual property that grows in value–not as an ideological preference for local hosting. A frontier model may still be the best rented component.

Runpoint Thesis Scorecard

The table separates claims that can be stated directly from claims that need qualification. This is also a useful editorial guardrail for a video: be most forceful where independent evidence is strongest.

Recommended Mid-Market Playbook

1. Choose one economic starting point. Target a workflow with at least one cashable baseline: SaaS, outsourcing, overtime, avoided hiring, loss, or constrained throughput.
2. Instrument before building. Record volume, labor minutes, acceptance/defect rate, cycle time, exception rate, cost, and downstream business outcome.
3. Redesign the workflow. Remove steps and change decision rights; do not merely insert a chat window.
4. Keep probabilistic work bounded. Use AI for generation, extraction, classification, search, or exception handling; keep permissions, state changes, approvals, and records deterministic.
5. Build the eval set early. Test the real edge cases, including refusal, escalation, security, and regression.
6. Create model optionality. Route light/high-volume work to cheaper or open-weight models and reserve frontier reasoning for cases where its acceptance lift exceeds its price.
7. Finance the rollout in gates. Prototype → shadow mode → limited production → scale. Release the next tranche only when accepted-output economics and adoption thresholds are met.
8. Measure realization quarterly. Reconcile hours saved to avoided cost, additional throughput, contribution margin, or a redesigned role.

The strongest initial candidates are likely software and information businesses, professional and business services, and document-heavy finance or insurance companies. Operational firms also fit when there is an obvious back-office or integration problem. Avoid leading with fully autonomous agents in regulated, irreversible, or weakly observable work.

Evidence Quality and Caveats

This review prioritizes representative government and academic evidence, controlled field experiments, then large cross-company surveys. Consultancy and vendor studies are useful for operating patterns but frequently survey large enterprises, rely on self-reported value, select advanced initiatives, or use changing definitions. Associations between AI adoption and revenue/productivity do not prove that AI caused the result; faster-growing, better-managed firms may adopt earlier.

The adoption series is directional because McKinsey changed questions and definitions over time. Task-study effect sizes are intentionally shown together to demonstrate the jagged frontier, but their endpoints, participants, models, and tasks are not directly comparable. Current model capabilities and prices are especially volatile; refresh the coding and token sections immediately before publication.

The middle-market evidence commonly uses $10M–$1B or employee-count bands rather than Runpoint's exact $20M–$200M segment. Conclusions about that band are therefore mechanism-based inferences, clearly labeled as such.

Selected Source Corpus

Representative and academic: U.S. Census Bureau, AI Use Among Businesses and Microstructure of AI Diffusion (2026); OECD, AI Adoption by SMEs; NBER, Firm Data on AI; Brynjolfsson, Li & Raymond, Generative AI at Work; Noy & Zhang, Experimental Evidence on the Productivity Effects of Generative AI; Dell'Acqua et al., Navigating the Jagged Technological Frontier; METR coding-productivity and task-horizon studies.

Enterprise and middle market: McKinsey State of AI (2017–2025 series); BCG AI Radar (2025, 2026); PwC AI Jobs Barometer and AI Performance Study; Deloitte State of Generative AI in the Enterprise; RSM Middle Market AI surveys (2024, 2025); National Center for the Middle Market (2026).

Engineering and economics: DORA research; OpenAI GPT-5.6 pricing and open-weight model documentation; Anthropic agent/coding research and enterprise consumption guidance; GitHub Copilot controlled experiment.

Runpoint point of view: Karp Is Right; Before the AI Project Starts; Runpoint's site and operating model. These are treated as the thesis being tested, not independent validation.