T17 · AI & Frontier Tech · Venture & Growth

AI-Native Startups: How Investors Underwrite the AI Application Layer

An evidence-led framework for underwriting AI-native application companies across product quality, model dependence, data rights, revenue, retention and unit economics.

AI application layers passing through an investor evidence and underwriting gate
Quick answer

AI-native startup underwriting begins with accepted customer outcomes, then tests the product system, provider dependencies, data rights, revenue evidence, retention and fully loaded cost per accepted outcome.

Abstract

Background. Rapid model improvement and cheaper access can expand application-layer opportunity while also compressing technical differentiation and exposing companies to provider, data and margin risk.

Objective. This paper develops an evidence-led underwriting framework for A5 global venture, growth and fintech investors, with B5 early-stage and growth founders as a secondary audience.

Approach. The analysis reviews 40 current official, regulatory, standards, company-filing and primary technical sources available through 1 August 2026.

Findings. Durable underwriting evidence links customer acceptance to workflow integration, controlled evaluations, governed data, provider portability, revenue quality, retention and cost per accepted outcome. Product claims and management dashboards require reconciliation to source records.

Implications. Investors and founders can organise diligence around an evidence ledger, red-flag register, portability tests and scenario-based unit economics. Attributed revenue, cost reduction and loss reduction remain USD 0 until approved observed evidence supports attribution.

JEL Classification: G24, G31, G32, L26, L86, M13, O32, O33

Keywords: AI-native startups, AI application layer, venture capital, growth equity, startup underwriting, model economics, inference cost, product evaluation, retention, data rights, AI governance, GCC venture ecosystem

This Matchpoint Insight presents the web edition of Matchpoint Partners' research. The supporting paper contains the full evidence review, underwriting framework, architecture, worked cases, scorecard, adoption roadmap and source register.

Read the full research paper   Explore AI & Technology Advisory

Introduction

Artificial intelligence has become a large and concentrated category of venture investment. Stanford HAI reports that global corporate AI investment more than doubled in 2025, with generative AI capturing nearly half of private AI funding [1]. NVCA reports that US venture investment reached record levels in the first half of 2026 while most invested capital flowed to AI companies and financings of USD 100 million or more [2]. Carta records USD 119.5 billion raised by companies on its platform in 2025 across 4,859 rounds, the lowest annual round count in at least six years; its Series A data shows a 38% median valuation premium for AI companies [3]. OECD reports that AI-related investment exceeded half of overall venture investment in 2025 [6].

These figures describe capital markets, not investment quality. A label such as AI-native can refer to a product whose customer outcome depends on machine inference, a conventional software product that added a model interface, a services business using AI internally, a workflow company with a proprietary evaluation and data loop, or a founder narrative with limited production evidence. Each case has a different product, risk, margin, distribution and capital profile.

The underwriting problem is therefore evidential. An investor needs to establish what the application does, which customer outcome changes, what portion of the workflow is automated, which model and data dependencies matter, how quality is evaluated, how failures are controlled, how customers buy and retain the product, how revenue converts to cash, how cost to serve changes with usage and which rights allow the company to operate. A founder needs the same evidence to manage the company and prepare a credible financing process.

This paper develops a framework for AI-Native Startups: How Investors Underwrite the AI Application Layer. The primary ideal customer profile is A5, Global VC, Growth & Fintech Investors: international venture, growth and fintech-focused funds assessing UAE and GCC opportunities. The secondary profile is B5, Early-Stage & Growth Founders: founders raising from pre-seed through Series A and growth, particularly in AI and fintech across the UAE, GCC, UK and South Asia. These profiles, the title, hook, visual requirements, target length and Tech & AI service mapping were verified from Matchpoint Partners' Topic Tracker and ICP Legend on 1 August 2026.

The paper's central proposition is that the quality-adjusted accepted customer outcome is the correct operating unit. An accepted outcome is a completed customer task or decision that passes defined quality, authority and evidence criteria. Examples include a reviewed credit memo, a resolved customer request, an accepted invoice classification, a released software change, a completed compliance check or a customer-approved design. Tokens, calls, generations, agents and model benchmark scores are technical inputs. They become economically relevant when they contribute to accepted outcomes, retained customers and collected gross profit.

Seven underwriting principles follow. First, the AI-native label must be translated into an observable workflow and architecture. Second, the company should evidence model performance on representative customer tasks. Third, revenue quality requires contract, invoice, collection, cohort, concession and concentration evidence. Fourth, gross margin requires a complete per-outcome cost ledger, including human review and reliability costs. Fifth, model and cloud providers are part of the product supply chain. Sixth, data rights, privacy, intellectual property, security and regulatory role determine where and how the product can be sold. Seventh, the investment case should survive a model-price shock, a quality regression, a provider retirement, a customer loss and a slower sales cycle.

The evidence cut-off is 1 August 2026. The analysis uses current official market reports, securities filings, provider documentation, regulator material, legislation, standards and security resources. No T17 company, founder, fund, customer, product, contract, cohort, data room, evaluation suite, cost ledger, cap table or investment memorandum was supplied. All worked volumes, prices, costs, conversion rates, productivity effects and valuation inputs are unverified illustrative management assumptions. Attributed Matchpoint or client revenue, cost reduction and loss reduction remain USD 0.

Scope, Definitions And Evidence Boundaries

Application layer

The application layer converts models, data, software and human work into a product used for a specific customer workflow. It can include user interfaces, application logic, retrieval, memory, tools, model routing, prompts, fine-tuning, evaluation, access control, observability, billing and support. The application company may use proprietary, third-party or open-weight models. Its defensibility can arise from workflow integration, proprietary data rights, evaluation assets, distribution, switching costs, regulated permissions, customer trust or operational learning.

The model layer supplies inference or model weights. The infrastructure layer supplies compute, storage, orchestration and deployment. These boundaries can overlap. A company that fine-tunes or hosts a model still needs to show how the model changes the customer outcome and how the application captures value.

LayerPrincipal componentsUnderwriting questionMinimum evidence
Customer workflowUser, task, decision, exception and approvalWhich costly or valuable job changes?Process map, baseline time, quality and authority
ApplicationInterface, logic, retrieval, tools and integrationWhat is the product customers buy?Product demonstration, release record and customer use
EvaluationTest cases, graders, thresholds and monitoringHow is quality established and preserved?Representative suite, results, failures and release gates
DataSource, rights, lineage, quality and retentionMay the company use the data for this purpose?Rights register, samples, lineage and controls
ModelProvider, version, route, prompt and tuningWhich capability is required and replaceable?Provider register, tests, costs and migration evidence
InfrastructureCompute, storage, security and observabilityCan the service operate reliably and economically?Architecture, invoices, logs, incidents and capacity plan
CommercialContract, price, term, renewal and collectionDoes use become retained, collected revenue?Contract-to-cash and cohort evidence

AI-native classification

An AI-native classification should be based on evidence. The classification does not require one technical architecture. It requires material dependence on machine inference within the customer value chain, a product designed around that capability and an operating system that measures its quality and economics.

TestEvidence of a substantive AI-native applicationWeak evidence
Customer outcomeDefined task with baseline, acceptance and consequenceGeneric productivity statement
Product designWorkflow built around inference, review and exception handlingChat interface added to existing software
EvaluationRepresentative task suite and production monitoringOne vendor benchmark or founder demonstration
DataGoverned source, rights, lineage and feedback loopUnspecified proprietary data claim
EconomicsPer-outcome price, cost and gross profitToken cost quoted without workflow cost
OperationsRelease, incident, rollback and provider migration processBest-effort manual troubleshooting
CommercialRetained customer use tied to contract and collectionPilot logos or letters of intent only
DefensibilityEvidence loop, workflow integration, distribution or permissionModel access available to every competitor

The result can be recorded as AI-native, AI-enabled, AI-assisted service or unverified. The label is a diligence conclusion, not a valuation conclusion.

Evidence classes

Every material claim should receive a class and owner. Observed evidence comes from a controlled system of record. Independently reproduced evidence comes from a test rerun by the investor, adviser or a qualified third party. Customer-attested evidence comes from a named authorised customer representative and should be reconciled to system data where possible. Management-estimated evidence is labelled and supported by a method. Unsupported claims remain open issues.

ClassDefinitionPermitted investment-memo use
O1 observedDated source-system evidence with owner and reconciliationStated as observed within the defined period and scope
O2 reproducedIndependently rerun method with retained inputs and resultStated as reproduced within the tested environment
O3 attestedNamed customer or counterparty confirmationStated as attested; limitations disclosed
M1 management estimatedDocumented calculation using unverified inputsLabelled management estimate or illustrative scenario
U unsupportedNo adequate evidence suppliedExcluded from base case; recorded as an open issue

Public comparables and market reports

Public filings demonstrate analytical methods and possible economic variation. Palantir reported USD 4.475 billion of 2025 revenue, an 82% gross margin and 954 customers [7]. UiPath reported USD 1.611 billion of fiscal 2026 revenue, USD 1.853 billion of annualised renewal run-rate and an 83% gross margin [8]. C3 AI reported USD 250.3 million of fiscal 2026 revenue and a 31% GAAP gross margin [9]. Duolingo disclosed that increased generative-AI and hosting costs reduced gross margin during 2025 and later described continued investment in AI infrastructure [10].

These companies differ in scale, product, deployment, customer mix, services, accounting and maturity. Their metrics do not determine a private startup's gross margin or valuation. They show why an investor should avoid a generic AI-software margin assumption.

Market reports also have boundaries. Carta's published data is aggregated from companies and securities holders on its platform and excludes companies that requested exclusion from anonymised studies [3-5]. NVCA's report covers the US venture ecosystem with PitchBook as its data provider [2]. OECD uses Preqin data in its AI venture brief [6]. Stanford HAI combines multiple public and proprietary sources [1]. The paper retains each source's scope.

Legal, investment and author boundaries

This paper does not determine whether a product is regulated, whether processing is lawful, whether intellectual-property rights are valid, whether a security is suitable, whether a valuation is fair or whether an investment should be made. Those decisions require the applicable contract, facts, jurisdiction and qualified authority.

Named-person author attribution remains pending CK approval. Matchpoint Partners is the organisational author. No founder performance, investor view or client outcome is attributed to a named person.

A5 Investor And B5 Founder Decision Map

A5 investor decisions

A5 investors need a staged answer to five questions. Is the customer problem valuable and repeated? Does the product create an accepted outcome that customers will pay for? Can the company maintain quality and control through model and market change? Does retained revenue create attractive gross profit and cash characteristics? Is the financing price supported by the evidence and downside?

A5 decisionCore evidenceFailure signalAuthority
ScreeningCustomer problem, workflow, product and initial tractionAI label without customer outcomeDeal team
Technical diligenceArchitecture, evaluations, data rights, security and portabilityDemonstration cannot be reproducedTechnical reviewer and deal lead
Commercial diligenceContract-to-cash, cohorts, references and pipelinePilots counted as recurring productionDeal team and commercial specialist
Financial diligenceRevenue bridge, gross margin, burn, runway and capital planCosts omit inference, review or supportFinance reviewer
Legal and regulatory diligenceCorporate, IP, data, licensing and product-role matrixRights or permissions cannot be evidencedQualified counsel and compliance owner
Valuation and structureBase, downside, dilution, milestones and reservesPremium driven only by AI categoryInvestment committee
Post-investment planEvidence gates, hiring, product and financing milestonesPlan measures output volume without accepted outcomesBoard and management

B5 founder decisions

The same evidence supports founder control. A founder should know which customer outcome drives retention, which failure causes the most customer harm, which provider dependence constrains scale, which cohort creates gross profit and which milestone deserves the next unit of capital.

B5 decisionRequired recordManagement use
Product scopeAccepted-outcome definition and exception boundaryPrioritise the workflow customers value
Model choiceRepresentative eval, cost and latency by provider/versionRoute tasks and plan migrations
Data strategyRights, quality, lineage and feedback registerBuild lawful and useful learning loops
PricingOutcome value, usage, service intensity and willingness to paySelect subscription, usage or outcome pricing
SalesFunnel by segment, source, stage and timeAllocate founder and sales capacity
Customer successCohort use, acceptance, support and renewalIdentify adoption and churn causes
HiringBottleneck, capability and milestoneLink headcount to product or commercial evidence
FundraisingUse of funds, gates, runway and downsideRaise against evidence creation

Shared evidence room

The investor and founder should work from the same reconciled evidence room. Different interpretation is legitimate. Different underlying numbers create avoidable risk.

Evidence blockPrimary ownerReconciliation
Customers and contractsCommercial leadCRM to contract repository to invoice and collection
Product usage and outcomesProduct leadEvent log to accepted-outcome definition
Evaluation and incidentsEngineering or AI leadRelease to evaluation report to production monitoring
Cost and gross marginFinance leadVendor invoice and payroll allocation to telemetry and ledger
Data and IPLegal or data ownerSource, agreement, use, retention and output policy
Security and privacySecurity or privacy ownerArchitecture, controls, tests, incidents and remediation
Cap table and financingCFO or company secretaryLegal register, option ledger and signed instruments
ClaimsCEO and functional ownersClaim, source, class, scope, date and approval

Market Context And The AI-Capital Concentration Problem

Premium and selection

AI capital growth expands the opportunity set and increases selection pressure. Carta's 2025 data shows more capital across fewer rounds and an AI valuation premium from Series A onward [3]. Its seed research reports leaner early teams, longer fundraising timelines and a widening gap between typical and top-tier valuations, particularly for AI companies [4]. Its H2 2025 compensation report describes AI-driven capital and talent effects while emphasising smaller teams at early stages [5].

The investor should therefore distinguish a category premium from company evidence. A valuation can reflect scarcity, competition, strategic option value, market size, growth, retention, team, technology, data rights or momentum. Each driver needs its own evidence and downside treatment.

Market-sizing discipline

An application-layer market should begin with customer units and workflow spend. Top-down AI totals can contextualise adoption. They do not establish obtainable revenue.

Market layerCalculationEvidence gate
Problem populationNumber of target organisations or users with the defined workflowNamed segmentation source and exclusions
Serviceable workflowPopulation with required data, budget, jurisdiction and integrationCustomer discovery and channel evidence
Paying unitContracted account, seat, workflow, transaction or accepted outcomeActual pricing and buying process
Attainable marketQualified accounts reachable through available distributionFunnel, capacity, conversion and sales-cycle evidence
Base planNew and expanded collected revenue by cohortContract-to-cash model and downside

For a vertical AI company, the market model should include the number of target firms, eligible departments, workflow frequency, current cost, required permission, integration burden, budget owner and plausible adoption. For a horizontal product, segmentation by use case and buyer prevents a superficial total-seat calculation.

Fundraising probability and runway

The market evidence supports a concentrated financing environment. It does not establish that a given company can raise. The financing plan should include a no-raise case, a delayed-raise case and a lower-price case. It should show monthly cash, committed cost, discretionary cost, hiring triggers, provider commitments, working capital and minimum operating runway.

Financing caseUnverified illustrative assumptionRequired management response
BaseNext round closes in month 12Hire only against product and commercial gates
DelayNext round closes in month 18Preserve cash and defer non-critical commitments
Lower pricePre-money valuation 30% below baseModel dilution, structure and option-pool effect
No raiseNo external capital within 24 monthsDefine break-even, strategic sale or orderly wind-down path

These cases are unverified illustrative management assumptions. They demonstrate planning mechanics. They do not predict financing availability, valuation or return.

The AI-Native Application-Layer Test

Customer outcome before model capability

The first diligence session should reconstruct one customer workflow without the product. The team records trigger, input, operator, decision, output, review, exception, elapsed time, error cost and authority. It then reconstructs the same workflow with the product. The comparison identifies which steps disappear, change, accelerate or require new controls.

The exercise exposes category errors. A faster draft can increase review effort. A lower handling time can increase correction or escalation. A higher automation rate can hide lower customer acceptance. A product that creates more output can reduce value if users cannot verify it within their decision window.

Workflow fieldBaseline recordAssisted recordAcceptance evidence
TriggerEvent that starts the taskSame or changed eventSource-system timestamp
InputDocuments, data and contextData passed to application and modelLineage and rights
WorkHuman and software stepsAutomated, assisted and retained stepsProcess log and observation
OutputDecision, content or actionGenerated candidate and final outputVersion and reviewer record
QualityError, completeness and policy testSame tests plus model-specific failuresRepresentative evaluation
AuthorityPerson allowed to release actionHuman or system release boundaryRole and approval record
ConsequenceRevenue, cost, risk or experienceObserved changeCustomer and finance evidence

Accepted customer outcome

An accepted outcome has five elements. The task is defined. The source evidence is sufficient. The output passes quality and policy thresholds. The authorised person or system accepts it. The outcome reaches the customer workflow. A generated answer that is discarded, a draft that requires full recreation and an automated decision reversed by review are produced outputs, not accepted outcomes.

The company should measure accepted outcomes per customer and cohort. Useful measures include first-pass acceptance, accepted outcome per active user, review minutes per accepted outcome, exception rate, correction rate, incident rate, customer escalation and time to acceptance. The definition should be stable enough for cohort analysis and specific enough to prevent volume inflation.

Claim register

SEC enforcement concerning AI statements shows the importance of accurate representations [12]. A private financing also depends on reliable claims. The claim register records each material assertion used in the pitch, data room, customer proposal and investment memo.

ClaimClassEvidenceScope and dateOwnerMemo treatment
"Cuts review time by 60%"M1 until observedTime-study method and sampleNamed workflow and periodProduct ownerManagement estimate or observed result
"Proprietary dataset"U until rights provenDataset inventory and agreementsData fields and jurisdictionsLegal/data ownerOpen issue
"Enterprise production"O1Contract, environment and usageCustomer, product and periodCommercial leadObserved with defined scope
"Best-in-class accuracy"URepresentative benchmark and comparatorTask, sample and versionAI leadExclude until reproduced
"90% gross margin"O1 or M1Ledger, invoices and allocationProduct and periodCFOReconciled actual or management plan

The claim register also helps founders. Unsupported claims can be corrected before a financing process. Management estimates can be converted into observed evidence through structured tests.

Evidence flywheel

A defensible application can build a lawful and useful evidence flywheel. Customer tasks create consented or contractually permitted usage data. Review and outcome signals improve evaluation. Evaluations improve routing, workflow and release decisions. Better accepted outcomes improve adoption and retention. Retained use produces more representative evidence. Each link needs rights, quality and economic value.

The investor should reject a circular data-moat claim. More data has value only when it is permitted, relevant, sufficiently labelled, incorporated into a controlled improvement process and difficult for a competitor to reproduce.

Product Value, Customer Workflow And Product-Market Evidence

Baseline and counterfactual

Product value needs a baseline. The baseline can be current human work, rules, incumbent software, outsourced service or no action. The counterfactual should reflect what the customer would actually do without the product. Comparing a product with an unrealistic manual process overstates value.

The strongest early evidence is a representative within-customer comparison. It uses the same task family, service requirements and decision window. It records input differences, selection effects and review. Randomised tests can be appropriate for product features. Operational and regulated workflows may require matched periods, staggered rollout or controlled shadow use.

Productivity and revenue bridge

The Topic Tracker hook asks how technology multiplies productivity and revenue. The paper separates gross activity from verified enterprise benefit.

Bridge lineCalculationEvidence requirement
Baseline capacityAccepted outcomes divided by full labour hoursTime study and stable outcome definition
Assisted capacityAccepted outcomes divided by assisted labour and exception hoursSame task mix and quality
Quality adjustmentAccepted outcomes passing all thresholdsEvaluation and customer acceptance
Customer valueAvoided cost, released capacity, lower loss or higher collected revenueApproved counterfactual and finance evidence
Product priceCollected revenue net of credits and concessionsContract, invoice and bank/ledger record
Product gross profitCollected or recognised revenue less complete cost to serveReconciled ledger and allocation policy

Productivity becomes a customer value proposition when the customer can redeploy capacity, reduce actual cost, serve more demand, improve time-to-revenue or reduce measured loss. Revenue becomes company value when contracted use converts to collected and retained gross profit.

Customer evidence ladder

LevelEvidenceWhat it supports
C0 discoveryInterview and observed painProblem hypothesis
C1 design partnerAccess, feedback and workflowProduct design evidence
C2 controlled pilotRepresentative tasks and agreed success criteriaFeasibility and user evidence
C3 paid pilotSigned fee and collectionWillingness to pay within pilot scope
C4 productionRepeated accepted outcomes in operating workflowProduct use and reliability
C5 renewalContinued paid contract after an informed decisionRetention evidence
C6 expansionAdditional product, volume, team or geographyBroader value evidence

Letters of intent, memorandum of understanding, unpaid pilot, paid pilot, production contract and collected renewal should be reported separately. Logos do not substitute for this classification.

Customer reference protocol

A customer reference should be authorised and structured. The interviewer confirms problem, prior workflow, buying process, deployment, time to value, quality, exceptions, support, use frequency, user breadth, price, renewal intent and alternatives. The investor records the relationship of the interviewee to the purchase and daily use.

The reference is customer-attested evidence. It is stronger when reconciled to contract, product and finance records. A founder-selected champion can provide valuable detail and may not represent procurement, security, finance or end-user views.

Distribution and sales motion

AI applications can use founder-led sales, product-led adoption, direct enterprise sales, marketplaces, cloud partners, systems integrators, channel partners and embedded distribution. Each motion has distinct cost, control, time and concentration.

Distribution measureDefinitionCommon distortion
Qualified opportunityAccount meets explicit customer, need, authority and timing criteriaDemo request counted as pipeline
Sales cycleFirst qualified contact to signed contractExcludes procurement or security delay
CACFully loaded acquisition cost for the relevant cohortOmits founder, partner or implementation cost
PaybackCAC divided by cohort gross-profit contributionUses revenue or target margin
Win rateWon opportunities divided by comparable resolved opportunitiesRemoves losses or stale deals
ExpansionAdditional collected recurring value from existing customerOne-off services included
Channel concentrationShare of qualified pipeline, revenue and collection by sourcePartner-sourced and partner-influenced conflated

Technical Architecture, Model Supply Chain And Portability

Architecture evidence

The architecture should show the complete path from user and source system to accepted outcome. It includes identity, permissions, data movement, retrieval, prompts, model routes, tools, code execution, guardrails, human review, storage, observability and billing. The diagram should distinguish customer environments, company systems and third parties.

ComponentUnderwriting evidenceKey dependency
User and identityRoles, authentication and authorisation testsCustomer identity provider
Source integrationData contract, lineage, failure and reconciliationCustomer API and data quality
Retrieval and memoryCorpus, freshness, access and deletionVector store and source rights
OrchestrationWorkflow version, retries, timeout and idempotencyApplication code and tool APIs
Model routeProvider, model/version, fallback and thresholdsPrice, capacity, lifecycle and terms
ToolsPermission, validation and transaction boundaryExternal service and credential security
Human reviewReviewer, queue, authority and escalationSkilled capacity and operating hours
MonitoringQuality, cost, latency, incident and drift measuresEvent completeness

Provider register

OpenAI, Anthropic, Google and AWS publish different pricing structures for models, caching, batching, tools, context and deployment [13-16]. Google also publishes model and endpoint lifecycle information, including retirements [17]. The register records the exact service used, not the provider brand alone.

FieldRequired content
Service identityProvider, product, model, version, region and account
CommercialPrice unit, tier, committed spend, discount, currency and renewal
DataInput/output use, retention, logging, training setting and subprocessors
PerformanceRepresentative quality, latency, availability and capacity
LifecycleRelease, deprecation, retirement and migration notice
ControlVersion pinning, fallback, rate limit and spend limit
ExitAlternative provider/model, migration work, test status and data export

List prices can change and realised spend depends on input, output, retries, caching, tool use, context, modality and volume. Diligence should reconcile invoice line items to product telemetry for a representative period.

Concentration and switching

The FTC's study of selected cloud-provider and AI-developer partnerships records cloud-spend commitments, exclusivity features, access to sensitive information and potential switching costs [11]. The report concerns those partnerships and does not establish a private application company's terms. It provides a useful diligence lens.

Concentration is measured across spend, accepted outcomes, critical features, regions and customers. A company can appear multi-model while one provider handles every revenue-critical task. A fallback that passes a synthetic prompt can still fail on customer tools, long context, safety behaviour, structured output or latency.

Portability test

A portability test runs a representative evaluation suite against an alternative model or deployment. It records code changes, prompt changes, output-schema differences, quality, review burden, latency, cost, safety, data terms and migration time. The result can be portable, partially portable or concentrated.

The investor should model at least three shocks: a 50% increase in model unit cost, withdrawal of the primary model within 90 days and capacity throttling during peak use. These percentages and periods are unverified illustrative scenario inputs. Management should replace them with its contractual and operating evidence.

Open-weight and licence dependency

Meta's Llama 4 Community License grants specified rights and includes redistribution, attribution, acceptable-use and scale-related terms [22]. Other open-weight models have different licences. Open-weight does not mean unrestricted. The company should retain the exact licence version, model card, download source, modifications, notices, acceptable-use review and redistribution analysis.

Evaluation, Reliability, Security And Operating Controls

Evaluation contract

NIST's AI RMF uses govern, map, measure and manage functions [23]. Its Generative AI Profile identifies generative-AI risks and recommended actions [24]. OpenAI's current model guidance recommends representative task evaluation before migration or optimisation [29]. Anthropic's evaluation guidance begins with specific, measurable and relevant success criteria [30]. These sources support a practical release discipline.

An evaluation contract defines task population, sample, ground truth, grader, dimensions, threshold, severity, subgroup, model/version, prompt, tool environment and release decision. The test set should include normal, difficult, adversarial, incomplete and out-of-scope cases.

Evaluation dimensionExample measureRelease treatment
Task successAccepted answer or completed tool actionMinimum weighted pass rate
Factual supportClaims supported by permitted evidenceCritical failure if material claim unsupported
CompletenessRequired fields and decisions presentSchema and reviewer check
Safety and policyProhibited, harmful or unauthorised actionZero tolerance for defined critical cases
PrivacyPersonal or confidential data handlingData-flow and output test
RobustnessPerformance under noise, prompt variation and missing dataThreshold by customer segment
LatencyTime to accepted outcomeCustomer decision-window threshold
CostFull cost per accepted outcomeProduct and margin threshold

Release and monitoring

One benchmark is insufficient. Releases can change model version, prompt, tools, retrieval, policies, code, data and customer configuration. The company should preserve a release manifest and rerun the required suite. Production monitoring then compares acceptance, failure, latency and cost with release expectations.

Release recordMinimum evidence
IdentityApplication, model, prompt, tool, data and policy versions
ChangeDescription, owner, reason and affected customers
EvaluationDataset version, results, failures and comparison
ApprovalProduct, engineering, security and domain authority as required
RolloutCohort, feature flag, limits, monitoring and rollback
DispositionRelease, restricted release, hold or retire

Reliability and incident evidence

Reliability is measured at the accepted-outcome boundary. Availability can be high while a critical tool silently fails. A model can respond while providing an invalid schema. A workflow can complete after the customer's decision window. The incident register should include provider failures, degraded quality, retrieval errors, data staleness, tool errors, security events, spend anomalies and human-review backlogs.

Incident fieldPurpose
Start, detection and recoveryMeasure exposure and response
Affected task, customer and outcomeEstablish operating consequence
Model, data, tool and code versionsReproduce the condition
Severity and authorityTrigger required escalation
Customer communicationEvidence contractual response
Root cause and corrective actionReduce recurrence
Cost, credit and concessionReconcile economic impact

Security

NIST's SSDF community profile adds generative-AI practices to secure software development [25]. CISA and NCSC address secure design, development, deployment and operation for AI systems [26]. OWASP identifies risks including prompt injection, sensitive-information disclosure, supply-chain vulnerabilities and unbounded consumption [27]. MITRE ATLAS provides an evolving threat knowledge base for predictive, generative and agentic AI systems [28].

ThreatApplication exposureControl evidence
Prompt or context injectionUntrusted content alters model behaviourTrust boundaries, sanitisation, instruction hierarchy and tests
Excessive agencyModel invokes consequential tool beyond authorityLeast privilege, allowlists, approval and transaction limits
Sensitive disclosurePrompt, retrieval, log or output reveals dataAccess, minimisation, redaction, encryption and monitoring
Supply-chain compromiseModel, package, embedding or tool is maliciousProvenance, scanning, pinning and vendor review
Unbounded consumptionRequests create service degradation or cost shockRate, token, tool, recursion and spend controls
Model theft or extractionRepeated queries reproduce proprietary behaviourAbuse detection, limits and response
Data or RAG poisoningSource content corrupts outputSource authority, ingestion control and anomaly tests

Security certifications and penetration tests are evidence within their scope and date. They do not replace architecture review, incident history or customer-specific controls.

Data Rights, Privacy, Regulation And Intellectual Property

Exact service terms

OpenAI states that business and API inputs and outputs are excluded from training by default [18]. Its services agreement sets the business relationship and relevant content rights [19]. Anthropic publishes commercial-product data-use information [20]. Google's Gemini API terms distinguish unpaid and paid services, including use of submitted content and human review for unpaid services and different treatment for paid services [21]. Google's zero-data-retention guidance identifies product settings and features relevant to retention [40].

These terms can change. The company should retain dated contract documents, account settings and data-flow evidence. A privacy answer based only on a provider's public brand-level statement is incomplete.

Data-rights register

Data classSourceRight and purposeTraining or evaluation useRetention and deletionTransfer and processor
Customer inputCustomer systemContract and applicable lawful basisDefined by agreementCustomer and company policyNamed providers and regions
Customer feedbackUser reviewProduct operation and improvement termScope-specificDefined periodApproved processors
Public dataIdentified publicationCopyright, database and use analysisRecordedProvenance retainedProvider terms checked
Licensed dataVendor or partnerLicence scope, field, geography and termExpress permissionContract termTransfer restriction
Synthetic dataCompany generationSource and generation rightsDocumentedVersionedModel terms checked
Employee contentStaff or contractorAssignment, confidentiality and policyRestricted purposeEmployment and legal policyApproved tools

The register should record data quality and usefulness separately from legal rights. Lawful data can be irrelevant. Useful data can be unusable because the required right is absent.

UAE, DIFC, ADGM, Saudi and UK data protection

The UAE Federal Decree-Law No. 45 of 2021 establishes a federal personal-data framework within its scope [33]. DIFC Regulation 10 addresses personal-data processing through autonomous and semi-autonomous systems [34]. ADGM guidance addresses automated individual decision-making, including the need for meaningful human review in relevant cases [35]. Saudi Arabia's PDPL and implementing framework establish controller obligations within scope [36]. The UK Information Commissioner's Office publishes AI guidance covering lawfulness, fairness, transparency, purpose limitation, minimisation, accuracy, storage, security and accountability [37].

Jurisdiction or regimeProduct-role questionDiligence evidence
UAE federalIs personal data processed within territorial and material scope?Processing record, basis, notice, processor and transfer map
DIFCDoes an autonomous or semi-autonomous system process personal data?Regulation 10 assessment and control record
ADGMDoes solely automated processing create legal or similarly significant effects?Decision map, exception, meaningful human review and appeal
Saudi ArabiaIs the company a controller or processor under the PDPL framework?Basis, purpose, notice, security, transfer and retention evidence
United KingdomHow do UK data-protection principles apply across the AI lifecycle?DPIA, fairness, transparency, rights and security evidence

This table is a diligence index. It is not a legal conclusion.

EU AI Act roles and transparency

The EU AI Act distinguishes model providers, system providers, deployers and other operators and applies obligations by role and risk [31]. The European Commission states that Article 50 transparency obligations apply from 2 August 2026, with specific transition treatment for certain systems placed on the market earlier [32]. A GCC company can be within scope through product placement, deployment or affected persons in the Union, depending on the facts.

The product-role matrix should be completed for each use case and jurisdiction. A company can be a provider for one feature, a deployer for another and a downstream integrator of a third-party general-purpose model. Contract language should align with actual architecture and control.

Copyright, patent and code provenance

The US Copyright Office has published reports addressing digital replicas, copyrightability of generative-AI outputs and AI training [38]. USPTO's revised November 2025 guidance states that ordinary inventorship standards apply to AI-assisted inventions and only natural persons can be inventors [39]. Provider agreements and open-model licences allocate and reserve different rights [19,21,22].

IP assetRequired recordDiligence issue
Source codeRepository, author, assignment, licence and dependencyOwnership and open-source obligations
Prompt and workflowVersion, author, customer restriction and secret treatmentCopyright, confidentiality and portability
DatasetSource, licence, collection and permitted purposeCopyright, database, privacy and contract
Fine-tune or adapterBase licence, training data, method and deploymentRights and provider conditions
OutputHuman contribution, review and customer termsProtectability and infringement risk
InventionHuman conception record and assignmentInventorship and patent ownership
Brand and domainRegistration, ownership and useFreedom to operate and customer confusion

Generated code should pass the same review, test, security, provenance and licence controls as human-written code. Diligence should sample commits and dependencies rather than rely only on a policy document.

Revenue Quality, Retention And Distribution

Contract-to-cash

Revenue underwriting begins with a customer-level bridge. The bridge links opportunity, signed contract, performance obligation, product access, usage, accepted outcomes, invoice, collection, credit, concession, renewal and expansion. It also identifies reseller, related-party, founder-network and strategic-partner relationships.

StageEvidenceInvestor test
OpportunityQualified need, buyer, budget and timingDefinition and ageing are consistent
ContractExecuted agreement, term, scope, price and terminationSigned by authorised parties; side letters captured
DeploymentEnvironment, user and integration recordProduction, pilot and test are separated
PerformanceAccepted outcomes and service evidenceContracted service was delivered
InvoiceInvoice, date, amount, tax and due dateReconciles to contract and ledger
CollectionBank or payment record and allocationCash is received and not reversed
Credit or concessionCredit note, discount, free period or service recoveryNet revenue and customer issue captured
RenewalInformed continuation and new termAuto-renewal, negotiated renewal and holdover separated

Annual recurring revenue should have a written policy. The investor should test whether pilots, consumption commitments, implementation, professional services, non-cancellable contracts, usage minimums, one-time fees and overdue invoices are included. Reported ARR should reconcile to a customer schedule and general ledger within the policy.

Cohorts

A cohort groups comparable customers by start period, segment, product, contract or acquisition source. Revenue retention should be shown on both gross and net bases. Logo retention should sit beside revenue retention because one expanding customer can conceal broad customer loss.

Cohort measureNumeratorDenominatorInterpretation
Gross revenue retentionOpening recurring revenue less churn and contractionOpening recurring revenueRetention before expansion
Net revenue retentionOpening recurring revenue less churn and contraction plus expansionOpening recurring revenueRetention including expansion
Logo retentionOpening customers retainedOpening customersBreadth of customer retention
Accepted-outcome retentionOutcomes from retained comparable workflowsOpening comparable outcomesContinued product use
Collection retentionCollected recurring value from retained customersOpening collected recurring valueCash quality

The company should disclose cohort size and age. A young cohort can show high net retention before the first meaningful renewal. A customer with a large initial rollout can create contraction without dissatisfaction. The evidence should explain the movement.

Concentration

Concentration is measured by revenue, gross profit, collection, accepted outcomes, pipeline and critical data rights. A high-revenue customer can be low margin due to dedicated infrastructure, services, support or model consumption. A low-revenue design partner can be operationally critical because the product depends on its data or domain expertise.

Sales efficiency

Customer-acquisition cost should include sales compensation, marketing, events, founder time, channel fees, solution engineering, security diligence, pilots and implementation where they are required to win the cohort. Payback uses cohort gross profit rather than revenue.

An early-stage company can report the underlying activity even when a stable CAC is unavailable. Useful evidence includes founder hours per stage, number of security reviews, pilot duration, implementation effort, win/loss reasons and time from signature to production acceptance.

Unit Economics, Gross Margin And Capital Plan

Per-outcome cost ledger

Provider token pricing is only one cost line [13-16]. The full cost of an accepted outcome can include source-system access, document processing, embeddings, retrieval, context storage, model input and output, caching, grounding, tools, code execution, retries, failed calls, orchestration, observability, security, human review, support and customer-specific infrastructure.

Cost poolAllocation driverEvidence
Model inferenceActual tokens, images, audio, calls or provisioned capacityTelemetry to invoice
Retrieval and storageIndexed volume, query, storage time and egressCloud records and invoice
Tools and data servicesInvocation, search, transaction or licenceTool log and invoice
Human reviewMinutes by role and outcome classWorkflow time record and payroll rate
Support and operationsTickets, incidents, on-call and customer environmentTicket and staffing record
InfrastructureCompute, database, network, observability and securityTagged cloud cost and allocation policy
Customer-specific deliveryIntegration, configuration and dedicated environmentProject record and contract

Illustrative gross-profit bridge

Assume, solely for illustration, that one customer purchases 10,000 accepted outcomes per month at USD 1.20 each. Monthly revenue is USD 12,000. Model and retrieval cost is USD 1,650; tools and infrastructure cost USD 900; human review costs USD 1,800; support and reliability allocation costs USD 750. Illustrative gross profit is USD 6,900 and gross margin is 57.5%.

LineUnverified illustrative assumptionPer accepted outcome
Collected or recognised revenueUSD 12,000USD 1.20
Model and retrievalUSD 1,650USD 0.165
Tools and infrastructureUSD 900USD 0.090
Human reviewUSD 1,800USD 0.180
Support and reliabilityUSD 750USD 0.075
Illustrative gross profitUSD 6,900USD 0.690
Illustrative gross margin57.5%57.5%

Every input in this example is an unverified illustrative management assumption. It proves no company or customer result. Attributed revenue, cost reduction and loss reduction remain USD 0.

Margin decomposition

Margin movement should be decomposed by price, volume, mix, model route, prompt/context size, caching, batch use, review rate, failure rate, customer-specific service and accounting classification. Duolingo's 2025 disclosure connecting generative-AI and hosting costs to gross-margin movement illustrates why direct AI cost belongs in the bridge [10]. The different gross margins reported by Palantir, UiPath and C3 AI illustrate product and business-model variation [7-9].

A declining provider price can improve margin. It can also invite price competition, enable new entrants or support a higher-quality model route that preserves unit cost. Underwriting should model the company response rather than assume all provider savings remain with shareholders.

Capital intensity

Application companies can be capital-light or capital-intensive. Capital requirements rise with long enterprise sales cycles, dedicated deployments, data acquisition, model training, guaranteed capacity, regulated permissions, security requirements, human operations and working capital. Committed cloud spend can create a fixed obligation before customer volume arrives.

Capital lineBase evidenceDownside question
Product and engineeringHeadcount, roadmap and release gatesWhich work is required for retention?
AI and dataProvider, training, evaluation and data commitmentsWhich cost scales ahead of revenue?
Sales and implementationFunnel, cycle and delivery capacityWhat happens when deals delay?
Security and complianceCustomer and jurisdiction requirementsWhich permission blocks revenue?
Working capitalBilling, collection and vendor payment termsDoes cash use grow with revenue?
FinancingCash, debt, equity and restricted fundsWhich covenants or milestones constrain action?

Runway bridge

The runway model should start with bank-verified cash and reconcile monthly receipts and payments. Management then applies hiring, provider, sales and regulatory gates. A board should see base, delay, lower-revenue, cost-shock and no-raise cases.

The plan should distinguish committed, contracted, approved, discretionary and unapproved spend. Use of funds becomes a list of evidence milestones: production acceptance, representative evaluation, renewal cohort, gross-margin threshold, provider migration, security requirement or licensing step.

Underwriting Scorecard, Investment Memo And Downside

Scorecard design

The scorecard records evidence sufficiency, not founder charisma or category excitement. Each domain receives a status: verified, partially verified, management estimated, unsupported or adverse. A numerical score can support consistency but should not conceal a critical issue.

DomainWeightEvidence gateCritical hold
Customer problem and outcome12Observed workflow and accepted outcomeNo valuable repeated task
Product and usage10Production use and adoptionDemonstration only
Evaluation and reliability12Representative release and production evidenceMaterial failures uncontrolled
Architecture and portability8Dependency map and tested fallbackUnmanageable provider concentration
Data, privacy and IP12Rights and role evidenceMaterial rights absent
Revenue and retention14Contract-to-cash and cohortsRevenue materially misstated
Unit economics12Reconciled cost per accepted outcomeCost ledger incomplete
Distribution8Qualified funnel and repeatable buying evidencePipeline unsupported
Team and governance6Named owners and control recordAuthority or integrity concern
Capital and terms6Runway, downside and financing planInsolvency or financing gap unmanaged

The weights are unverified illustrative management assumptions. The investment committee should approve its own weights and critical holds.

Investment-memo evidence packet

Memo sectionRequired exhibit
ThesisClaim register and disconfirming evidence
MarketBottom-up workflow market and source limitations
ProductBefore/after workflow, accepted outcome and product demonstration record
TechnologyArchitecture, provider register, evaluation and portability test
CustomersContract-to-cash sample, cohorts and structured references
EconomicsPer-outcome ledger, gross-margin bridge and sales efficiency
Legal and regulatoryCorporate, IP, data and product-role matrices
TeamRoles, references, ownership, incentives and key-person plan
FinancingCap table, instrument terms, use of funds, runway and downside
DecisionConditions, reserves, governance rights and post-investment gates

Downside scenarios

ScenarioIllustrative shockEvidence to modelManagement response
Model costPrimary model unit cost rises 50%Current invoice and workload distributionRouting, price, context and fallback plan
QualityFirst-pass acceptance falls 15 percentage pointsEval and production acceptance historyRollback, review and customer restriction
ProviderPrimary version retires in 90 daysLifecycle notice and migration testAlternative model and release plan
CustomerLargest customer churnsRevenue, gross profit and cost concentrationCost, cash and pipeline response
SalesMedian enterprise cycle extends six monthsStage history and procurement evidenceHiring and runway adjustment
RegulationProduct requires additional assessment or permissionRole matrix and counsel analysisScope, geography or control change

All shocks are unverified illustrative scenario inputs. The diligence team should replace them with company-specific evidence and committee-approved severities.

Valuation and structure

The valuation analysis should separate current evidence from option value. Current value can use observed revenue, retention, gross profit, growth and cash characteristics. Option value can reflect new products, geographies, datasets or model capability and should carry probability, capital and time.

Structure can address evidence gaps through staged funding, milestones, reserves, information rights, consent rights, board rights, founder vesting, option pools and warranties. Legal counsel should draft and assess the transaction. Milestones should be within management influence and defined by auditable evidence.

Founder And Investor Playbooks

Founder diligence-room index

FolderMinimum contentUpdate frequency
01 CorporateIncorporation, registers, group, licences and board materialsOn change
02 CapitalCap table, instruments, options, debt and waterfallOn transaction
03 CustomersContract schedule, signed samples, invoices, collections and creditsMonthly
04 ProductRoadmap, releases, usage and accepted outcomesEach release/monthly
05 AI and evaluationModels, prompts, suites, results, incidents and migrationsEach release
06 Data and IPRights, provenance, assignments, licences and inventionsOn change/quarterly
07 Security and privacyArchitecture, policies, tests, incidents, DPIAs and processorsQuarterly/on incident
08 FinanceManagement accounts, bank, budget, margin and runwayMonthly
09 CommercialPipeline, funnel, pricing, losses and partnersWeekly/monthly
10 TeamOrganisation, contracts, compensation, references and hiring planMonthly
11 ClaimsPitch and memo claim registerBefore every external use

The founder should preserve original records and provide reconciliations rather than a curated collection of favourable screenshots. Customer-identifiable and personal data should be shared only under appropriate authority and controls.

Investor fieldwork

The investor begins with a request list and evidence map. It samples customers across retained, expanded, pilot, churned and lost-opportunity groups. It reproduces a bounded evaluation. It traces selected contracts to usage, invoices and cash. It reconciles provider invoices to telemetry. It samples source-code ownership and data rights. It tests one provider migration. It reviews incidents and board reporting.

FieldworkSample objectiveEvidence retained
Workflow observationConfirm task, review and acceptanceDated process map and limitations
Evaluation reproductionTest representative qualityInputs, versions, output and disposition
Contract-to-cash tracingTest revenue existence and qualityContract, invoice, collection and credit
Cohort recreationTest retention calculationsCustomer schedule and formula
Cost reconciliationTest full cost per accepted outcomeInvoice, telemetry and allocation
Provider migrationTest portability and hidden workChange log, eval, cost and time
Security/privacy reviewTest material controls and incidentsArchitecture, test and remediation
Founder and team referencesTest history, roles and integritySource, relationship and factual notes

Management questions

  1. What exact customer outcome is accepted, by whom and under which quality threshold?
  2. Which production task fails most often, and what is its customer consequence?
  3. Which reported productivity result has an observed baseline and counterfactual?
  4. Which provider, model or tool is critical to retained revenue?
  5. How long would a representative migration take, and when was it last tested?
  6. Which data claim has the weakest rights evidence?
  7. Which customer cohort produces the highest collected gross profit?
  8. Which customer requires the most human review or support?
  9. Which material claim in the pitch deck remains a management estimate?
  10. What result would cause management to restrict, redesign or stop the product?
  11. What cash plan applies if the next financing is delayed by six months?
  12. Which milestone converts the largest remaining uncertainty into evidence?

Gated 12-18 Month Roadmap

The roadmap begins with definition and reconciliation. It advances through representative evaluation, production evidence, retained economics and financing readiness. Calendar timing depends on company state, customer process and regulatory requirements.

PhaseTarget periodDeliverableExit gate
0 DefinitionMonth 0-1Workflow, accepted outcome and claim registerNamed owners and stable definitions
1 Evidence roomMonth 1-2Contract-to-cash, architecture, provider and rights registersCore records reconcile
2 EvaluationMonth 2-4Representative suite, release gate and failure taxonomyCritical thresholds pass
3 EconomicsMonth 3-6Per-outcome cost and cohort gross-margin bridgeInvoice, telemetry and ledger reconcile
4 PortabilityMonth 4-7Alternative model/provider migration testDefined quality, time and cost threshold
5 Commercial proofMonth 4-10Production adoption, reference and first renewal evidenceCustomer and collection gates pass
6 Scale controlMonth 8-14Monitoring, security, support and unit-economics controlsGrowth does not degrade outcome or margin
7 FinancingMonth 10-18Evidence-led memo, downside, valuation and use of fundsCommittee or board-approved transaction path

The roadmap is a sequence of evidence gates. A company can move faster when reliable evidence already exists. It can remain at a phase when a gate fails. The disposition can be advance, restrict, redesign, retain as research or stop.

Limitations, Research Agenda And Conclusion

This paper presents a diligence framework based on public evidence. It does not report an investment, customer cohort, production evaluation or financial return. Market data describes different datasets and periods. Public-company filings describe mature businesses with economic profiles that may not resemble a private application startup. Provider pricing and terms can change after the evidence cut-off.

The framework requires company evidence for implementation. Future research should analyse anonymised contract-to-cash cohorts, accepted-outcome economics, model migration events, incident costs and valuation outcomes across application categories. Comparative studies should report complete workload, model, prompt, tool, review, latency, failure and cost boundaries. GCC research should distinguish federal, DIFC, ADGM and Saudi legal contexts and use qualified legal analysis.

The investment conclusion is disciplined. AI-native application underwriting starts with one valuable customer workflow and one quality-adjusted accepted outcome. It then follows the evidence through architecture, model supply chain, evaluation, data rights, production use, contract, collection, cohort, cost and control. The company earns an AI-native classification when this chain is material and reproducible. It earns an investment case when retained customer value, gross profit, governance and financing price remain credible under downside.

For A5 investors, the framework converts category excitement into a claim register, fieldwork plan, scorecard and evidence-led memorandum. For B5 founders, it creates a management system for product quality, customer value, provider dependence, data rights, gross margin and fundraising readiness. Both parties benefit from one reconciled evidence room and explicit treatment of uncertainty.

Questions, answered

AI-native startup underwriting: frequently asked questions

An AI-native application company makes model-mediated capabilities central to the customer workflow and value proposition. The underwriting question is whether the company repeatedly delivers accepted customer outcomes through a controlled product system, rather than whether it merely calls a model API.

It is a customer outcome that passes explicit quality, safety, policy, review and acceptance gates. The metric connects usage to a governed unit of customer value and supports analysis of reliability, cost, retention and revenue.

Investors should test architecture, evaluation design, failure handling, human review, observability, security, data flows, portability and provider dependencies. Claims should be tied to reproducible evidence, representative holdouts and customer acceptance criteria.

The diligence team should map each critical provider, model, region, contract, price and lifecycle dependency. It should then test fallback paths, portability, switching costs, evaluation parity and the effect of provider changes on service quality and gross margin.

Useful evidence includes signed contracts, invoice and cash records, cohort retention, expansion, customer concentration, implementation burden, acceptance events and links between product usage and contractual value. Management dashboards should be reconciled to source records.

Gross margin should include model inference, retrieval, storage, orchestration, evaluation, human review, support and other service-delivery costs. Investors should examine cost per accepted outcome by customer and workflow, together with sensitivity to provider pricing, retries and usage mix.

Diligence should verify rights to collect, process, retain and reuse customer, user, training, evaluation and output data. It should also test processor terms, cross-border transfers, deletion, access controls, auditability and the treatment of confidential data by model providers.

Founders should assemble an evidence room that links customer contracts, accepted outcomes, evaluation results, architecture, provider terms, data rights, incidents, unit economics and cohorts. Unverified forecasts and management scenarios should be labelled and kept separate from observed results.

The worked cases and adoption thresholds are unverified illustrative management assumptions. Attributed revenue, cost reduction and loss reduction remain USD 0 because no approved observed Matchpoint or client attribution evidence was supplied.

This publication is general research for professional audiences. It is not investment, legal, regulatory, accounting, tax, cybersecurity, technology or valuation advice, and it is not an offer, solicitation, recommendation or promise of results. Readers should verify current requirements and decisions with qualified advisers.

Build an evidence-led AI underwriting case

Discuss product diligence, provider concentration, data rights, revenue quality, retention and unit economics with a Matchpoint partner.

WhatsApp