Vertical AI vs. Horizontal Assistants: Defensibility, SaaS Economics and Small-Company Opportunities

Review period: Mixed historical and undated live evidence, reviewed through October 13, 2026; a single bounded evidence period is not assigned. Scope: U.S.-focused comparison of vertical AI workflow applications and horizontal assistants, including enterprise demand, deployment evidence, economics, incumbent competition and opportunities for small software companies. Source policy: Original surveys and transaction research, government evaluations, academic studies, company financial disclosures and official product documentation; vendor claims, evidence limitations and Revoqra analysis distinguished.

Executive summary

Verdict: vertical AI workflow applications are generally a better entry strategy for a small software company than another standalone general-purpose assistant—but they are not inherently more profitable. The attractive opportunity is owning a repeatable, valuable business process, not merely adding industry terminology to a chatbot. This is a Revoqra synthesis of the purchasing, deployment and competitive evidence reviewed below.

Infor’s October 6, 2026 research identifies genuine concerns about industry fit and human review. Independent evidence supports the importance of workflow redesign and shows real deployments, but does not establish that vertical AI vendors consistently achieve superior customer retention, operating margins or customer ROI. Horizontal AI remains the larger category in Menlo Ventures’ estimated enterprise application spending. (infor.com)

The strongest competitive advantages combine workflow state, reusable integrations, domain-specific validation, permissioned data and customer distribution. Established ERP and enterprise vendors already possess several of these advantages and are embedding AI into their products. Consequently, a startup’s best opening is usually an underserved workflow around or across existing systems—not replacing those systems. (infor.com)

For small companies, the most promising default shortlist is:

1. Specialty construction contract-to-cash workflows.
2. Commercial insurance broker/MGA document operations.
3. Industrial quality and audit-evidence preparation, where founders have industry access.
4. Specialty healthcare administrative workflows, with greater compliance and integration burdens.

These are analyst-ranked opportunities, not demonstrated sector-wide profitability rankings. The decisive commercial test is whether customers pay for a repeatable workflow while implementation, support and human exception handling remain bounded.

1. What is actually being compared?

For this report:

• Horizontal assistant: a broadly applicable interface for drafting, research, information synthesis and other knowledge work across industries.
• Vertical AI application: software designed around an industry-specific process, its records, rules, permissions, exceptions and required outputs.
• Departmental AI: role-specific tools, such as coding assistants, that operate across industries. These should not automatically be classified as vertical.

The boundary is increasingly blurred. Microsoft 365 Copilot already integrates with Microsoft Graph and Office applications; it is not simply an isolated chat window. Conversely, a vertical deployment can use a general-purpose foundation model: Hiscox’s underwriting deployment uses Google’s Gemini within its own business process. A vertical application therefore does not require a proprietary foundation model. (learn.microsoft.com)

The useful comparison is between two business architectures:

• Assistant architecture: help a person produce or find an answer.
• Workflow architecture: move a defined work item through an accountable process and produce an accepted business outcome.

Revoqra analysis: Workflow ownership can support differentiation and switching costs. However, complexity becomes an advantage only when it is sufficiently reusable to be productized; otherwise, it becomes a consulting obligation.

2. Infor’s October 2026 findings: useful demand evidence, not a profitability test

What the original research says

Infor released its second Enterprise AI Adoption Impact Index wave on October 6, 2026. YouGov conducted the survey on Infor’s behalf during August 5–19, 2026, interviewing 2,111 business decision-makers across seven countries, including 550 in the United States. (infor.com)

The global findings come from the 2,111-person sample; the U.S. October findings use the 550-person subgroup. These are reported perceptions and estimates, rather than audited operating results. Infor explicitly cautions that the findings should not be interpreted as outcomes achieved through Infor products. (infor.com)

Important qualifications

1. Commercial sponsorship matters. The survey accompanied an Infor product announcement promoting industry-specific AI. That does not invalidate the responses, but it makes independent corroboration important. (infor.com)
2. The waves are not automatically a matched panel. April covered 1,024 respondents in four markets; October covered 2,111 in seven. The public material reviewed does not establish matched respondents across waves, so global changes should not be treated as clean longitudinal measurements. (infor.com)
3. The strongest economic questions remain unanswered. The public summaries do not provide comparable vertical-versus-horizontal implementation costs, renewal cohorts, realized contract prices or vendor profits; the full report is offered through a registration form. (infor.com)

Assessment: Infor supports the hypothesis that production AI needs business context, governance and review. It does not demonstrate a category-wide vertical SaaS profitability advantage.

3. Independent purchasing evidence: vertical demand is real, but horizontal demand is stronger than the thesis sometimes assumes

Measurement differences matter

• Menlo surveyed 495 U.S. enterprise AI decision-makers during November 7–25, 2025, combined with a bottom-up market model. Its spending estimates exclude AI features bundled into existing software, which can understate incumbent competition relevant to a startup. (menlovc.com)
• Ramp’s overall transaction research covers 70,000-plus American businesses. Its adoption definition is an observed payment in a month, not a representative nationwide survey of AI use. The precise eligible monthly denominator is not supplied with the cited update. (ramp.com)
• McKinsey surveyed 1,719 respondents in 97 countries during May 4–June 8, 2026, weighting responses by national GDP contribution. Its results are reported organizational experience, not audited financial attribution. (mckinsey.com)

Buying versus building: two apparently conflicting findings can coexist

Menlo reported that 76% of AI use cases were purchased rather than built internally. McKinsey subsequently found that 32% of respondents’ organizations had declined to buy at least one software product or feature because they could build it with agentic coding tools. (menlovc.com)

These percentages concern different units, periods and samples: use cases versus organizations declining at least one purchase. They cannot be subtracted or treated as a trend series.

Revoqra interpretation: Enterprises will buy finished, dependable workflows while increasingly building small features themselves. The vulnerable startup is a thin application whose principal asset is a prompt and interface. The stronger startup supplies maintained integrations, validation, exception resolution and operational accountability.

4. Documented deployments and measured results

What works—and what remains unproven

These studies are not like-for-like product comparisons. They differ in users, countries, products, observation windows, task definitions and controls. Comparing daily minutes across them would not establish which category has better ROI.

Vertical benefits can be valuable without being immediately financial

A separate before-and-after study of 263 clinicians across six health systems reported burnout falling from 51.9% to 38.8% after 30 days with an ambient scribe. Without a randomized comparison, that change should not be interpreted as a definitive causal effect or a demonstrated financial saving. (media.jamanetwork.com)

The Peterson Health Technology Institute’s March 2025 assessment likewise found promising documentation and cognitive-load benefits while describing financial impact as unclear. Retention of clinical staff is a plausible value mechanism, but improved burnout scores alone do not measure avoided employee departures. (phti.org)

Domain branding does not guarantee reliability

A 2024 preregistered study using more than 200 legal queries found hallucination rates of 17–33% among the evaluated specialist legal research products, although they improved on general-purpose GPT-4. This is a historical benchmark, not an October 2026 estimate for current product versions. Its enduring lesson is that retrieval and specialization do not eliminate the need for domain-specific evaluation and verification. (arxiv.org)

5. Comparison across the six requested business dimensions

The following table is Revoqra synthesis, not a set of measured category averages.

Published prices: useful signals, not a matched comparison

Microsoft’s 2026 commercial notice described Copilot Business at $21 per user monthly, while separately discussing changes to Microsoft 365 bundle pricing. Treat this as a published 2026 benchmark, not an all-in October enterprise quote. Nabla’s subscription documentation, updated October 4, 2026, lists $119 monthly for an individual subscription, with group rates negotiated separately. (techcommunity.microsoft.com)

CoCounsel Legal’s pricing process depends on customer characteristics and offers one-, two- and three-year plans; existing customers are directed to sales for pricing. These examples demonstrate different packaging, not a measured sector-wide vertical price premium. (sales.legalsolutions.thomsonreuters.com)

Revoqra pricing recommendation: Use a predictable base subscription plus bounded workflow volume where appropriate. Pure token resale exposes the product to model-price competition; unlimited work exposes the vendor to uncontrolled exceptions. Outcome pricing requires sufficiently reliable attribution and appropriate incentives.

6. Where sustainable competitive advantages can emerge

Four stronger assets than an industry-specific prompt

1. Workflow state and accepted outputs. Knowing which work item is pending, what evidence is missing, who must approve it and what constitutes completion is more useful than retaining a conversation alone. Infor and Oracle explicitly organize agents around governed business processes and transactional context. (infor.com)
2. Maintained domain integrations. Mapping customer-specific records, permissions and document relationships can be difficult to replicate reliably. Document Crunch’s native Procore integration illustrates embedding domain functionality inside an existing workflow rather than forcing a separate destination. (help.documentcrunch.com)
3. Permissioned exception and outcome data. The useful learning asset is not simply a collection of PDFs; it is labeled errors, approved corrections and subsequent outcomes. This is a Revoqra inference. Access rights matter: customer data cannot automatically be pooled or repurposed, particularly where HIPAA limits permitted uses. (hhs.gov)
4. Reusable validation and accountability. Tests, approval paths, evidence traceability and operational support can make deployment dependable. The specialist legal study shows why domain-specific interfaces alone are insufficient. (arxiv.org)

Sector assessment

Regulation cuts both ways. Compliance can increase switching costs after deployment, but it raises the new vendor’s entry costs too. HHS does not certify or endorse particular HIPAA technology products; a compliance label is not itself a unique moat. A no-training promise, such as OpenAI’s Business data commitment, is also not a substitute for the applicable healthcare contractual and operational requirements. (hhs.gov)

7. Competition from ERP and established enterprise software vendors

The incumbent threat is not merely that large vendors can build a chatbot. It is that they already control records, permissions, installed workflows and purchasing relationships.

These are announced or documented capabilities and commercial positions, not independent demonstrations of their ROI or ease of implementation.

Procore’s 2026 annual report also describes acquiring smaller, often marketplace-integrated companies to accelerate its roadmap. This illustrates a broader strategic possibility: incumbents can build, bundle or acquire specialized functionality. Partnership is a distribution option, not protection from competition or a guaranteed exit. (s21.q4cdn.com)

More plausible openings for small companies

Revoqra analysis: Favor workflows that involve:

• Multiple companies or software systems rather than one incumbent’s internal records.
• A narrow specialty with distinctive exceptions.
• A clearly accountable output, such as an approved submission or complete evidence package.
• Reusable integrations that do not require replacing the system of record.
• A customer channel the founder can reach economically.

Cross-system positioning alone is not sufficient: a generic connector can be copied. The durable asset must include domain validation, workflow adoption or data rights that are harder to reproduce.

8. Profitability and retention: what the disclosed numbers actually demonstrate

Established vertical platforms provide useful—but imperfect—proxies

Comparability limits: These are three purposively selected companies, not an industry sample. They include mature software and non-AI revenue; the periods and revenue-segment definitions differ. FY2027 labels here refer to quarters ending in July 2026, not future calendar-2027 results. They should not be averaged into a vertical-AI margin estimate.

The disclosures demonstrate that workflow software can support strong subscription economics while services and operating expenses materially change profitability. ServiceTitan’s professional-services-and-other segment is broader than AI implementation, so its loss should not be attributed specifically to AI onboarding. (investors.servicetitan.com)

Similarly, NRR is not customer renewal rate. Intapp explicitly attributes its figure to upselling and cross-selling. Expansion can coexist with customer churn; the headline disclosures do not supply the underlying customer-cohort N or an AI-only retention cohort. (investors.intapp.com)

Young AI companies do not automatically inherit mature SaaS margins

Bessemer’s January 2026 vertical-AI compendium reports approximately 65% gross margin for a selected LLM-native portfolio cohort. The cohort size and observation period were not verifiable from the accessible extract, limiting its use as a current market benchmark. (bvp.com)

Its separate 2025 study reported approximately 25% average gross margin among ten selected rapidly growing AI startups. These are not matched vertical-versus-horizontal cohorts, but they demonstrate why exceptional revenue growth should not be confused with healthy delivery economics. (bvp.com)

Ramp’s September 2026 update reported its blended effective token-price index falling approximately 41%, from a March peak of $1.15 to $0.68 per million tokens. This is an observed index influenced by model mix, not a fixed-quality comparison or an October API tariff. Cheaper inference can help, but does not resolve expensive human exceptions or bespoke delivery. (ramp.com)

Illustrative unit economics: the exception path can dominate

Revoqra-derived hypothetical scenario—not customer data, an observed average or a forecast. Assume one customer processes 1,000 documents monthly and pays $2,500. All inputs below are illustrative.

The review calculation is 200 documents × five minutes ÷ 60 × $60/hour = $1,000. With assumed acquisition cost of $12,000, recurring gross-profit payback from billing commencement is six versus twelve months, before onboarding and fixed overhead.

An additional 40 hours of onboarding at an assumed loaded $100/hour costs $4,000 unless recovered separately. Moving review to the customer can preserve vendor margin, but the customer must still subtract that labor when calculating ROI.

Conclusion: An expensive workflow can support pricing power and still produce poor SaaS economics. The relevant unit is gross profit per accepted work item—not model cost per token or ARR alone.

9. Ranked vertical opportunities for small software companies

Ranking assumptions and methodology

This is Revoqra analyst judgment, not a survey ranking or statistical forecast. It assumes a small team with a domain expert or partner, limited capacity for long enterprise sales and no need to train a frontier model.

• Willingness to pay: 5 = highest expected value/budget capacity for the specified buyer.
• Technical defensibility: 5 = strongest potential combination of domain validation, integrations and permitted data advantages—not simply coding difficulty.
• Acquisition difficulty: 5 = hardest. Scores assume the specified smaller or midmarket buyer, not a national insurer, hospital chain or bank.
• Margin ranges: conditional steady-state delivery targets, not observed sector medians or promised achievable results. They include recurring direct delivery costs but exclude sales, R&D and G&A; unbilled onboarding lowers first-year contribution. They are planning judgments informed by the disclosed economics above, not calculated population estimates.

The order rewards bootstrappability and repeatability, not maximum theoretical market size. Without industrial relationships, opportunity 3 should move below more accessible niches. A founder’s existing distribution can materially change the order.

How to make each opportunity more defensible

1. Construction: sell a financial workflow, not construction chat.

Start with one trade and document family. Assemble approved evidence from contracts, project records and correspondence; track outstanding requirements and deliver a reviewable payment or change-order package. Pilot metrics should include attributable collections, completeness and staff effort—not only document-search speed. Procore’s expanding AI makes generic project Q&A a weaker differentiation strategy. (support.procore.com)

2. Insurance: organize submissions before attempting autonomous underwriting.

Focus on document normalization, missing information, renewal differences and traceable review. Keep underwriting authority with the customer. The potential asset is repeated exception handling and permitted feedback, not merely summarizing policies. Hiscox shows that prepared communications and pricing can still pass through an underwriter. (hiscoxgroup.com)

3. Industrial quality: prepare evidence; do not begin by controlling production.

An initial application can organize evidence, link requirements and draft nonconformance documentation for approval. Avoid an initial promise of autonomous product release or production control. Test whether the same mappings and checks work across several customers; extensive customer-specific logic can destroy the proposed margin advantage. Infor’s relevant agents make the competitive boundary important. (infor.com)

4. Healthcare administration: choose a specialty and a payer problem.

The 2025 CAQH Index, released February 19, 2026, drew on 600-plus provider organizations and health plans representing 63% of insured lives and estimated $21 billion of remaining administrative automation opportunity. That is not an AI revenue TAM or money automatically recoverable by startups. (globenewswire.com)

A narrower product should measure net additional collections or completed authorizations after staff review and fees. Nonclinical positioning does not remove HIPAA obligations when the relevant PHI-processing relationship exists. (hhs.gov)

5. Accounting: own the evidence workflow rather than compete for final professional judgment.

Document requests, classification, missing-evidence detection and review preparation can be bounded. The product should preserve source links and approval history. A generic tax assistant faces incumbent products already incorporating domain AI; customer acquisition and renewal must be tested across the practice’s real workload cycle. (thomsonreuters.com)

6. Legal: target an accepted matter output.

Prefer a specific evidence package or chronology over broad legal research. Require source traceability and attorney approval. As a business-model inference, efficiency is easier to monetize where the firm benefits from higher throughput or fixed-fee economics; reducing billable time is not automatically a revenue gain in an hourly practice. Reliability remains a product obligation, not a marketing claim. (arxiv.org)

Attractive markets that are not default bootstrap opportunities

Revoqra assessment: Large health-system clinical deployments, institution-wide financial workflows, regulated life-sciences core systems and ERP replacement can offer substantial long-term value. They are poorer default choices for a small company lacking sector relationships and delivery capacity. The documented deployments and established-platform economics show both the value of domain integration and the advantages already held by larger organizations. (jamanetwork.com)

Likewise, a general-purpose assistant can be defensible when its owner already has a distribution platform or privileged enterprise context. The weak small-company proposition is the undifferentiated standalone wrapper—not horizontal AI as an entire category.

10. Practical investment gates for a small software company

The following are recommended operating tests, rather than source-reported findings:

1. Name the budget owner and completed outcome. A user enjoying the demo is not the same as a buyer funding an accepted work product.
2. Secure several paying design partners in one narrow segment. Confirm that the same process—not merely the same industry label—recurs.
3. Track all delivery labor. Record onboarding, connector work, support and exception review separately from model expense.
4. Benchmark against the customer’s existing assistant and suite features. Use the same inputs, permissions, tasks and acceptance criteria.
5. Evaluate ordinary cases and costly exceptions. Report sample sizes, rejection rates, review minutes and post-approval errors; avoid cherry-picking successful tasks.
6. Start with controlled actions. Use traceable outputs, explicit approvals and reversible integrations before expanding authority.
7. Test recurring demand and renewal. Distinguish a continuing workflow from a one-time cleanup project.
8. Productize before broadening. If every new customer needs bespoke engineering or vendor-operated manual work, price it as a service or redesign it rather than reporting SaaS-like economics.

Measure customer ROI and vendor economics separately

A useful customer calculation is:

First-year ROI = (realized first-year benefits − fully loaded first-year costs) / fully loaded first-year costs

Count benefits carefully:

• Labor savings only to the extent they become avoided cost, redeployed capacity or demonstrable output.
• Incremental revenue at contribution margin, not its full face value.
• Verified loss avoidance rather than hypothetical exposure.
• Faster cash collection through its financing/liquidity benefit, not by treating all accelerated cash as new profit.

Include licenses, usage, implementation, customer training, internal review, security work and ongoing maintenance. Show first-year and recurring-year economics separately; do not hide upfront delivery costs through an unexplained amortization assumption.

The evidence reviewed makes this discipline important: reported individual productivity does not reliably establish enterprise financial impact, and even documented vertical deployments can show modest benefits without a net cost-benefit conclusion. (mckinsey.com)

Final assessment

Vertical AI is a better default differentiation strategy for a new small software vendor, not a blanket superior asset class.

The strongest opportunity is a narrow, repeatable workflow where the company can obtain legitimate data access, maintain useful integrations, validate output and demonstrate customer value without assuming responsibility for unlimited manual exceptions.

Infor’s central industry-fit concern is credible. Its research does not establish that independent vertical vendors will capture the economics: customers can buy horizontal tools, configure incumbent platforms or build internally. The sustainable startup advantage must survive those alternatives.

Confidence assessment:

• High: industry context, integration and human review materially affect production usefulness.
• Moderate: selected vertical workflows offer better small-company differentiation than an undifferentiated assistant.
• Unproven: vertical AI as a category has superior retention, operating profitability or customer ROI.

Sources

Adoption and purchasing

• Infor: October 2026 original announcement, global methodology and U.S. findings. (infor.com)
• McKinsey: 2026 global AI survey and research methodology. (mckinsey.com)
• Menlo Ventures: 2025 enterprise AI spending, purchasing and market-model methodology. (menlovc.com)
• Ramp: AI Index definitions and September 2026 transaction update. (ramp.com)
• CAQH: Original 2025 Index release describing administrative automation evidence. (globenewswire.com)

Deployments and independent evaluations

• Microsoft Copilot field experiment: Original randomized workplace research. (arxiv.org)
• U.K. DWP: Government Copilot evaluation published January 2026. (gov.uk)
• JAMA: Multisite scribe adoption study and clinician-burnout evidence. (jamanetwork.com)
• PHTI: Assessment of early ambient-scribe economic evidence. (phti.org)
• Hiscox and Morgan Stanley: First-party production and adoption disclosures. (hiscoxgroup.com)
• Specialist legal AI study: Original preregistered reliability evaluation. (arxiv.org)

Economics, competition and requirements

• Financial disclosures: Intapp FY2026; ServiceTitan and Veeva quarters ending July 31, 2026. (investors.intapp.com)
• Bessemer: Selected vertical-AI portfolio and rapidly growing AI-company benchmarks. (bvp.com)
• ERP documentation: Infor October release, SAP commercial guidance and Oracle Fusion announcements/pricing distinctions. (infor.com)
• Construction platforms: Procore AI, Document Crunch integration and Procore acquisition strategy. (support.procore.com)
• Professional software and pricing: CoCounsel documentation, Microsoft commercial notice, Nabla subscriptions and OpenAI Business terms summary. (sales.legalsolutions.thomsonreuters.com)
• HHS: Official HIPAA cloud-computing guidance. (hhs.gov)

Unresolved evidence gap: The reviewed sources substantiate demand, deployments and delivery-cost risks, but do not supply matched vertical-versus-horizontal AI cohorts sufficient to settle comparative customer retention or operating profitability. The opportunity ranking is therefore a conditional investment framework, not a statistically validated forecast.