From hype to clinical translation: A tiered, readiness-based framework for artificial intelligence in gastrointestinal endoscopy

Abstract

Artificial intelligence (AI) in gastrointestinal endoscopy is maturing unevenly. We organized this review around a predefined six-domain readiness framework: Evidence maturity, regulatory or health-system pathways, real-world deployment, workflow actionability, governance and monitoring, and generalizability. By these criteria, colonoscopy computer-aided detection is the sole clear tier 1 application, supported by multiple randomized trials and Food and Drug Administration clearances, though net patient-level value remains uncertain, and three concurrent guideline panels have issued discordant recommendations on identical evidence. Computer-aided diagnosis for optical polyp characterization remains tier 2 because two rigorous meta-analyses show no net benefit for the resect-and-discard strategy in routine practice. Upper gastrointestinal second-observer systems, AI-assisted procedural quality systems, capsule endoscopy reader-assist tools, and endoscopy-based Helicobacter pylori prediction are also tier 2, each limited by pathway uncertainty or limited deployment experience. Cholangioscopy AI, therapeutic endoscopy assistance, and endoscopic ultrasound-based pancreatic lesion analysis are tier 3, where technical performance consistently outpaces translational evidence. We also propose a prospective implementation checklist for regulators and endoscopy units evaluating emerging systems, and a prioritized five-year research agenda covering pathway-defined trials, representative datasets, human-factors safeguards, and post-deployment monitoring aligned with contemporary AI reporting standards.

Key Words: Artificial intelligence; Deep learning; Gastrointestinal endoscopy; Computer-aided detection; Computer-aided diagnosis; Computer-aided quality assessment; Workflow integration; Generalizability; Governance; Resource-limited settings

Core Tip: Endoscopic artificial intelligence (AI) applications are at very different stages of translation. Colonoscopy computer-aided detection is the only system that meets readiness across all six domains we examined, yet guideline panels remain split on net patient value. Upper gastrointestinal second-observer systems, AI-assisted procedural quality systems, capsule endoscopy reader-assist tools, and endoscopy-based Helicobacter pylori prediction systems are technically strong but lack defined clinical pathways. Cholangioscopy, therapeutic, and pancreatic ultrasound AI are exploratory. Future translation will depend less on classifier accuracy and more on pathway definition, workflow integration, governance, and monitoring. We offer a tiered framework and a prospective checklist to guide implementation decisions.

INTRODUCTION

Artificial intelligence (AI) has become one of the most visible technological developments in gastrointestinal endoscopy. Most endoscopic AI applications fall into three functional categories: Computer-aided detection (CADe), which highlights suspected abnormalities during live procedures; computer-aided diagnosis (CADx), which supports real-time lesion characterization; and computer-aided quality assessment, which monitors procedural quality, inspection completeness, and workflow performance[1,2]. Implementation-focused commentaries and the 2022 European Society of Gastrointestinal Endoscopy (ESGE) expected-value statement emphasize that the clinically relevant goals of endoscopic AI now extend beyond image classification to procedural quality, standardization, workflow efficiency, governance, and patient-centered value[3-6].

The field has progressed beyond retrospective proof-of-concept image classifiers. In colonoscopy, the evidence base is now sufficiently mature that major societies and guidelines have moved to practice-facing recommendations, including the 2025 American Gastroenterological Association (AGA) Living Clinical Practice Guideline, the 2025 ESGE Position Statement, and the 2025 BMJ Living Clinical Practice Guideline[7-9]. These guidelines mark an important shift from technical validation toward implementation-facing appraisal, while also showing that even the most mature endoscopic AI application remains debated with respect to downstream patient-level value.

Outside of colonoscopy, the rationale for AI adoption remains compelling, but the evidence is less mature and more unevenly distributed. In the upper gastrointestinal tract, for example, earlier recognition of Barrett’s-associated neoplasia, superficial squamous cell carcinoma, and early gastric neoplasia may preserve eligibility for curative endoscopic resection[10,11]. Similar implementation-facing questions apply to capsule endoscopy reader-assist tools and endoscopy-based Helicobacter pylori (H. pylori) prediction, where AI may improve efficiency or visual diagnostic support. Recent syntheses describe encouraging progress across these domains while also emphasizing persistent barriers to translation, including dataset bias, false positives, workflow integration, and limited external validation[12-14]. Thus, the current landscape is not one of uniform maturity, but of markedly different levels of evidence, deployment readiness, and clinical actionability.

This uneven maturity exposes a limitation in how the field is commonly reviewed. Many prior reviews organize endoscopic AI by organ system and foreground diagnostic performance metrics such as sensitivity, specificity, or area under the curve[1,2,12]. Although useful for cataloguing technical progress, this structure is less informative for clinicians, endoscopy units, regulators, and payers deciding whether a system is ready for clinical implementation[3-6]. A model with excellent retrospective performance may still fail in practice if it depends on narrow datasets, performs inconsistently across hardware ecosystems, disrupts workflow, lacks a defined downstream action pathway, or introduces unresolved monitoring, governance, and liability concerns[3-6,15]. A second limitation is that “readiness” is often invoked without explicit criteria. Responsible implementation requires more than accuracy: It requires evidence of clinical utility, pathway integration, generalizability, lifecycle monitoring, update governance, equity, and failure management[3-6,15]. The World Endoscopy Organization consensus statement and ESGE expected-value position statement have emphasized data governance, fairness, medicolegal responsibility, workflow integration, and equity as central to responsible AI adoption[3,6]. Building on these principles, our six-domain rubric adapts implementation-relevant concepts into an application-level classification of endoscopic AI translation readiness.

These concerns are not abstract. Dataset-readiness analysis has highlighted persistent limitations in dataset size, disease diversity, annotation quality, demographic breadth, and device representation[15]. A recent bibliometric analysis showed that the published literature is heavily concentrated in China, the United States, Japan, and a small number of European centers[16]. This concentration should temper assumptions about global generalizability, particularly in South Asia, Latin America, sub-Saharan Africa, and lower-resource or rural practice environments[15,16].

To address these gaps, this narrative framework synthesis drew on peer-reviewed publications identified through targeted searches of PubMed and MEDLINE through February 2026, prioritizing randomized trials, systematic reviews and meta-analyses, multicenter validation studies, regulatory decision summaries, and major society guidance relevant to endoscopic AI translation. We critically appraise the field through a tiered, readiness-based framework centered on clinical translation. Systems are classified into tier 1, deployable or closest to routine use; tier 2, strong near-term potential but unresolved translational barriers; and tier 3, promising but exploratory. We then separate the retrospective classification from a prospective checklist intended to help investigators, regulators, and endoscopy units determine when a system has genuinely advanced to the next stage of translation.

HOW THE TIER FRAMEWORK WAS OPERATIONALIZED

We defined six domains directly relevant to clinical translation: Evidence maturity, regulatory authorization or health-system pathway, real-world deployment, workflow actionability, governance and post-deployment monitoring, and generalizability or resource fit (Table 1)[3-9,15].

Table 1 Operational criteria used for tier assignment.

Domain	Tier 1	Tier 2	Tier 3
Evidence maturity	Multicenter randomized evidence or pooled randomized data with clinically meaningful endpoints and well-described downstream consequences	Prospective multicenter, pooled randomized, or strong external-validation data, but patient-level or pathway consequences remain incomplete	Single-center, retrospective, pilot, or proof-of-concept predominance
Regulatory authorization/health-system pathway	Regulatory authorization for intended use plus either supportive or permissive practice-facing guidance (i.e., recommending in favour of or conditionally allowing the technology) or a structured evidence-generation pathway	Partial regulatory or health-system traction, pilot implementation, or pathway-facing evaluation, but convergence not yet present	No meaningful regulatory traction or practice-facing pathway
Real-world deployment	Use beyond expert development centers, including community or non-academic settings	Limited or early real-world implementation	No meaningful real-world deployment evidence
Workflow actionability	Clear clinical action path with feasible real-time integration into procedural workflow	Action path plausible but not standardized	Action path unclear, procedure-specific, or not yet clinically testable
Governance/monitoring	Explicit oversight with named metrics, update disclosure, override logging, and evidence-generation or post-market monitoring	Published protocol or institutional plan specifies some governance elements, but lifecycle monitoring remains incomplete	Governance largely undefined
Generalizability/resource fit	Evidence across heterogeneous populations, platforms, or settings, with plausible operational fit outside expert centers	Some external validation, but geographic or vendor concentration remains substantial	Narrow setting, platform, or population dependence

Tier assignment is conjunctive across domains: A system must satisfy the tier 1 threshold in all six domains to be classified as tier 1. Within any individual domain, the listed criteria represent alternative ways to satisfy that domain’s threshold, but maturity is judged across all domains. Within the combined regulatory and health-system domain, regulatory authorization, society guidance, and evidence-generation pathways were treated as complementary but non-equivalent signals; a positive signal in one category did not automatically imply maturity in the others. Borderline cases were assigned to the lower tier. The framework should therefore be read as a pragmatic, author-derived instrument for comparative translation assessment rather than as a validated taxonomy[3-6].

This framework evaluates clinical translation readiness, not technical sophistication. A technically impressive model can rank lower than a less sophisticated system that is better deployed and governed. As systems move toward implementation, expectations should converge with established AI reporting and lifecycle standards, including CONSORT-AI for randomized trials, DECIDE-AI for early live clinical evaluation, International Medical Device Regulators Forum Good Machine Learning Practice, and US Food and Drug Administration (FDA) predetermined change-control planning[17-20]. The eight questions outlined in Table 2 are intended to make that standard operational and should be applied before any system is considered for advancement to the next tier.

Table 2 Prospective checklist for advancing an endoscopic artificial intelligence system to the next tier.

Question	Why it matters	Minimum evidence before advancing tier
Does the system improve a clinically meaningful endpoint rather than only image-level accuracy	Detection gains may not translate into patient benefit if they mainly increase low-value findings	At least one prospective study with workflow-relevant outcomes; tier 1 requires multicenter randomized or pooled randomized evidence
Is there a clear action pathway once the AI output is generated	Outputs without downstream decisions create ambiguity, delay, and liability risk	Explicit linkage between AI output and biopsy, resection, documentation, referral, or review pathway
Has performance been shown outside the development environment	Single-center or single-vendor success often overestimates real-world performance	External validation across centres, operators, and ideally more than one hardware ecosystem
Will deployment preserve safe human performance	Automation bias and deskilling can offset technical gains	Human-factors plan with onboarding, override logging, periodic AI-off benchmarking, and monitoring of behaviour-level metrics
Is governance defined before launch	Undefined responsibility undermines adoption and patient safety	Named accountability, update policy, discordant-case review, and AI-specific protocol or reporting aligned with CONSORT-AI or DECIDE-AI when applicable
Is post-deployment monitoring specified	Static pre-deployment evidence cannot detect drift, latency issues, or workflow changes	Named metrics, review frequency, trigger thresholds, rollback or recalibration plan, and change-control policy consistent with lifecycle guidance
Is the dataset and validation geography sufficiently representative	Geographic and demographic concentration limits generalizability and equity	Evidence of representation across populations, settings, and device environments relevant to intended deployment
Is the system economically and operationally sustainable	Clinical value may be offset by cost, follow-up burden, or proprietary infrastructure constraints	Context-specific implementation plan addressing costs, maintenance, reimbursement, and downstream utilization

AI: Artificial intelligence.

TIER 1: COLONOSCOPY CADE IS DEPLOYABLE, BUT ITS PATIENT-LEVEL VALUE REMAINS INCOMPLETELY RESOLVED

Using the rubric above, commercially available colonoscopy CADe is the only application that clearly meets tier 1 criteria across all six domains[7-9,21-28]. This conclusion is supported by randomized trials, regulatory authorization, practice-facing guidance, and early deployment in community, ambulatory, and academic settings[7-9,28-31]. No other gastrointestinal (GI) endoscopy AI application currently shows the same convergence across all six domains.

The strongest efficacy signal comes from pooled randomized data. A 2025 meta-analysis of 28 randomized controlled trials (RCTs) involving 23861 participants found that AI-assisted colonoscopy increased the adenoma detection rate (ADR) by 20% and reduced the adenoma miss rate by 55%[28]. However, the same analysis found no significant improvement in advanced adenoma or sessile serrated lesion detection, a 39% increase in non-neoplastic resection, and a modest prolongation of withdrawal time[28]. The larger AGA pooled analysis of 41 RCTs and 32108 participants reported a 2% absolute increase in the detection of combined advanced adenomas and sessile serrated lesions, but with very low certainty[7]. This represents the central unresolved question in tier 1 colonoscopy AI: If most incremental yield comes from diminutive adenomas and low-value lesions, then the net clinical benefit remains uncertain despite higher ADR[7-9,32].

Recent guidance is cautious rather than definitive. Three contemporary major guidance documents reached materially discordant conclusions, which is itself a clinically important finding. The 2025 AGA Living Clinical Practice Guideline made no recommendation for or against routine CADe-assisted colonoscopy because the certainty was very low and the balance of desirable and undesirable effects was close[7]. The 2025 ESGE Position Statement reached the opposite directional conclusion: The ESGE panel voted 13 to 6 to issue a weak recommendation in favor of CADe for most well-informed patients already undergoing screening or surveillance colonoscopy, while acknowledging modest absolute gains, more non-neoplastic detections, and residual uncertainty around advanced neoplasia and post-colonoscopy colorectal cancer[8]. The BMJ living guideline went further in the opposite direction from ESGE, issuing a weak recommendation against routine CADe use across CADe and CADx alike; microsimulation work informing that guidance estimated only a modest incremental benefit at the cost of additional surveillance burden and procedures[9,32]. The fact that three panels reviewing the same evidence produced recommendations spanning the full spectrum, weakly against, no recommendation, weakly in favor, underscores that current evidence supports neither categorical adoption nor categorical rejection. National Institute for Health and Care Excellence draft guidance is best understood as a conditional evidence-generation pathway rather than routine adoption: Five technologies may be used in the National Health Service during a 4-year evidence-generation period, with structured monitoring and reassessment[21].

Regulatory and market signals still distinguish CADe from all other endoscopy AI domains. In the United States, GI Genius received FDA De Novo authorization as a real-time aid for lesion detection during white-light colonoscopy, not as a CADx system[22]. Additional 510(k)-cleared CADe platforms include EndoScreener, SKOUT, MAGENTIQ-COLO, CAD EYE, and CADDIE[23-27]. This broader device landscape supports classifying colonoscopy CADe as the only endoscopic AI domain with clear routine-market traction, although authorization alone does not resolve downstream value, training, or governance.

Real-world performance, however, is less uniform than randomized efficacy. A 2024 systematic review and meta-analysis of nonrandomized real-world studies found only a slight increase in ADR, no significant improvement in adenomas per colonoscopy, and no pooled ADR benefit in studies restricted to GI Genius[29]. The multicenter community-based AI-SEE randomized trial found no improvement in ADR, adenomas per colonoscopy, or serrated polyp detection, while increasing non-neoplastic detections and withdrawal time[30]. A Swedish pragmatic randomized trial similarly found no overall benefit in ADR, although detection of sessile serrated lesions improved[31]. These studies indicate that the central translational question is no longer whether CADe works in trials, but whether the benefits justify the costs, false positives, and procedural consequences in routine practice.

Human factors further complicate the tier 1 narrative. In a 2025 retrospective, multicenter observational study, continuous exposure to AI was associated with a decline in adenoma detection during subsequent non-AI-assisted colonoscopy, raising concerns about overreliance or deskilling[33]. The retrospective design limits causal inference, but the signal is mechanistically plausible and warrants prospective evaluation[4,33]. Implementation should therefore be approached as a governed service rather than simple software procurement. Safeguards proposed in the literature include onboarding on known failure modes and interface limitations, periodic AI-off benchmarking, structured review of discordant AI-human cases, override logging, update disclosure, drift monitoring, and protocol or reporting expectations aligned with CONSORT-AI, DECIDE-AI, IMDRF Good Machine Learning Practice, and FDA-predetermined change-control planning[3,4,17-20,33,34]. For these reasons, colonoscopy CADe is deployable under active governance, but its long-term patient-level value remains incompletely resolved. Tier 1 classification, therefore, reflects deployment readiness rather than resolved patient-level value, and the prospective registry work proposed in the following sections and in Table 3 should accompany, not follow, any broad rollout of current CADe systems[3-6,17-20].

Table 3 Priority translational agenda for the next 3 years to 5 years.

Priority	Why it matters	Example deliverable
Patient-important outcomes for Tier 1 colonoscopy AI	ADR alone is no longer sufficient to justify broad adoption given that most incremental yield is from diminutive lesions	Multicenter registry or pragmatic trial accompanying (not following) rollout and measuring advanced neoplasia, interval cancer surrogates, low-value resection rate, and surveillance intensity
Pathway-defining trials for tier 2 tools	Technical accuracy does not establish net clinical value	Capsule AI trial measuring reading time, false-positive burden, downstream procedure rate, and cost; H. pylori study linking AI outputs to biopsy strategy and management
Global generalizability	Current literature is concentrated in a few regions and expert centres	Prospective validation across underrepresented geographies and lower-resource settings, including South Asia, Latin America, and sub-Saharan Africa
Human-factors safeguards as a standard requirement	Automation bias can erode clinician performance, as the deskilling signal in colonoscopy demonstrates	Mandatory AI-off benchmarking, override logging, and discordant-case review as part of every rollout protocol
Governance before guideline endorsement	Deployment without accountability is operationally fragile	Minimum package of update disclosure, drift monitoring, and escalation policy before any tier 2 system receives society-level endorsement
CADx pathway clarification	Two major meta-analyses show no net benefit of current CADx in routine practice; the resect-and-discard pathway requires society-level redefinition before AI-assisted optical diagnosis can be broadly implemented	Prospective pragmatic trial comparing AI-assisted resect-and-discard vs standard practice on histopathology concordance, surveillance interval assignment, and medicolegal framework

AI: Artificial intelligence; ADR: Adenoma detection rate; H. pylori: Helicobacter pylori; CADx: Computer-aided diagnosis.

TIER 2: HIGH-POTENTIAL APPLICATIONS WITH UNRESOLVED TRANSLATIONAL BARRIERS

Upper gastrointestinal second-observer systems

Upper gastrointestinal second-observer systems are the tier 2 domain closest to tier 1[12,35-40]. A 2026 systematic review and meta-analysis of 11 randomized trials involving 57512 participants found that AI-assisted esophagogastroduodenoscopy improved neoplasm detection both per patient and per lesion, with the largest relative gains for lesions 10 mm or smaller[12]. Supporting evidence includes two important studies with different designs that should not be conflated. GRAIDS was a multicenter case-control diagnostic study in which real-time AI achieved diagnostic accuracy comparable to that of expert endoscopists and superior to non-experts[35]. ENDOANGEL-LD was a single-center, tandem, randomized trial showing a reduction in gastric neoplasm miss rates[36]. The evidence base now also includes Barrett’s esophagus: A 2023 Lancet Digital Health model-development and validation study showed strong CADe performance for early Barrett’s neoplasia across 15 centers[38], a 2024 pilot live-procedure study detected all visible Barrett’s lesions at the patient level but with modest specificity[39], and a 2024 multicenter CADx study improved characterization of Barrett’s neoplasia by general endoscopists to expert level[40].

These data strengthen the case for upper-GI AI as a near-term translation target, but they do not yet stabilize the clinical pathway. The tandem randomized trial for superficial esophageal squamous cell carcinoma showed a favorable effect on miss rates, but the investigators noted that real-world effectiveness and cost-benefit remained uncertain[12]. More importantly, a later randomized clinical trial in esophageal squamous cell carcinoma did not show improved detection with AI assistance, even among non-expert endoscopists[37]. Across upper-GI tasks, performance remains heterogeneous by organ, lesion phenotype, operator group, and study design[12,35-40]. Real-world multicenter deployment is still limited, specificity remains task-dependent, and it is not yet clear when AI outputs should alter biopsy strategy, referral patterns, or surveillance intervals. Upper-GI AI, therefore, remains tier 2: Persuasive, increasingly practical, but not yet sufficiently standardized across devices, settings, and downstream decisions.

CADx for optical polyp characterization

CADx for optical diagnosis of colorectal polyps is one of the three foundational AI categories in GI endoscopy and has a substantial clinical trial base that requires direct engagement rather than omission. The American Society for Gastrointestinal Endoscopy Preservation and Incorporation of Valuable Endoscopic Innovations thresholds established a benchmark for the clinical implementation of optical diagnosis[41]. Multiple RCTs and prospective studies have examined whether AI-assisted CADx meets these thresholds. Barua et al[42] demonstrated in a prospective multicenter study that a real-time AI system achieved sensitivity and specificity for neoplastic polyps that were non-inferior to expert visual inspection under ultra-magnifying conditions. Rex et al[43] subsequently showed in a prospective study that AI-assisted optical diagnosis by general endoscopists achieved negative predictive values consistent with Preservation and Incorporation of Valuable Endoscopic Innovations thresholds for diminutive rectosigmoid polyps. The DISCARD3 feasibility study confirmed that a resect-and-discard approach achieved at least 90% concordance between optical diagnosis and histology surveillance intervals for diminutive polyps in a screening setting, with substantial projected cost savings[44].

However, two more recent and methodologically rigorous meta-analyses from the CADx Analysis Study Group substantially complicate this picture. A 2024 meta-analysis concluded that CADx assistance modestly improved sensitivity for leaving rectosigmoid polyps in situ but also raised concerns about harms due to reduced specificity[45]. More definitively, another 2024 meta-analysis specifically examining CADx for the resect-and-discard strategy found that CADx produced no demonstrable benefit or harm compared with unassisted endoscopist optical diagnosis and explicitly questioned the clinical value of current systems[46]. CADx, therefore, remains tier 2 because downstream use depends on local surveillance rules and pathology workflows, liability around AI-assisted leave-in-situ decisions remains unsettled, and no major society currently endorses routine AI-assisted resect-and-discard as a clinical standard[3,5,6,9,45,46]. CADx is closer to implementation than many other tier 2 systems, but current evidence does not support describing it as a deployed standalone decision pathway.

AI-assisted procedural quality systems

AI-Assisted Procedural Quality Systems are classified as tier 2 because, although a prospective randomized quality-control trial[47] and a subsequent four-group parallel study[48] were both positive, the regulatory, practice-guidance, and large-scale deployment profiles remain much thinner than for colonoscopy CADe. In the prospective randomized study, a real-time automatic quality-control system improved polyp and adenoma detection by modifying withdrawal behavior and inspection completeness[47]. In the four-group parallel study, a quality-improvement system amplified the efficacy of CADe rather than serving as a standalone intervention[48].

These systems address an underappreciated translational truth: Many missed lesions reflect process failure, not only visual failure. However, the evidence base remains narrower and more methodologically heterogeneous than for CADe, and broad health-system or regulatory convergence has not yet occurred. They are best understood as near-term implementation candidates rather than mature standards of care.

Capsule endoscopy reader-assist tools

Capsule endoscopy AI is a strong tier 2 candidate because its value proposition is operationally clear: Reduce reading burden, prioritize suspicious frames, and mitigate fatigue-related misses[13,14]. A 2025 systematic review and meta-analysis found that AI-assisted capsule reading increased pooled sensitivity and negative predictive value but reduced pooled specificity relative to conventional reading[13]. That pattern supports an assistive or triage role rather than autonomous interpretation.

The specificity signal deserves more scrutiny than it has typically received. Lower specificity may shift burden from physician reading time to downstream confirmatory procedures, repeat review, or low-value follow-up, consequences that current reviews have not adequately quantified[13,14]. A 2025 review of commercially available capsule AI tools concluded that market-ready systems can markedly reduce reading time and detect many lesions effectively, but still miss a meaningful number of abnormalities and therefore cannot replace clinician review[14]. Capsule AI may improve efficiency, but its net clinical value remains contingent on whether the time saved in reading is offset by false positives, residual misses, or additional downstream investigations.

Endoscopy-based H. pylori prediction

Endoscopy-based H. pylori prediction fits tier 2[49,50]. A 2025 meta-analysis reported pooled sensitivity and specificity of approximately 0.91 in both internal and external validation settings and found higher diagnostic performance than that of junior and senior endoscopists[49]. The multicenter HOPE AI study evaluated more than 356000 images from 6207 patients across seven hospitals and reported strong internal, temporal, and geographic validation performance, with sensitivity exceeding that of senior endoscopists[50].

Even so, pathway implications remain unsettled. High diagnostic performance alone does not determine whether these systems should function as biopsy-triage tools, visual-assessment adjuncts, or structured screening workflow components[49,50]. Most validation remains geographically concentrated, and external validation outside East Asia is limited. Until broader validation exists, the framework should be interpreted as uneven.

TIER 3: PROMISING BUT STILL EXPLORATORY APPLICATIONS

Tier 3 includes applications with compelling proof-of-concept data but insufficient evidence for stable clinical pathways, broad external validation, or routine workflow integration. Cholangioscopy AI, therapeutic or third-space procedural assistance, and endoscopic ultrasound (EUS)-based pancreatic lesion analysis fall into this category[51-56]. These are clinically important targets, but the gap between technical performance and deployable benefit remains substantial.

Cholangioscopy AI is the most technically advanced within tier 3. A 2026 multicenter validation study tested a cholangioscopy AI system on new recordings from multiple academic centers and reported performance above that of brush cytology, forceps biopsy, and their combination[51], a clinically meaningful result given the long-standing sensitivity limitations of endoscopic retrograde cholangiopancreatography-based sampling. Earlier work with a real-time interpretable cholangioscopy model showed excellent discrimination and performance that exceeded both expert and novice endoscopists in the development setting[52]. Even so, cholangioscopy AI remains exploratory because these studies are diagnostic validations rather than pathway-defining trials. It remains unclear how AI outputs should alter real-time sampling strategy, whether they reduce time to diagnosis or repeat procedures, and whether performance is stable across broader device ecosystems, stricture phenotypes, and operator experience[51,52].

Therapeutic and third-space endoscopy is earlier still in translation. AI-assisted submucosal vessel detection improved endoscopist vessel recognition and shortened detection time in a prospective study, but the investigators explicitly noted that procedural benefit still requires clinical corroboration[53]. The AI-Endo workflow-recognition platform for endoscopic submucosal dissection demonstrated real-time feasibility in an animal model, but feasibility is not equivalent to readiness[54]. Until such systems show reductions in bleeding, perforation, procedure time, or adverse events in routine human cases, they should remain tier 3.

EUS-based AI for pancreatic lesions also remains exploratory. A 2024 systematic review concluded that AI in pancreatic EUS is promising but limited by retrospective designs, small datasets, and operator dependence[55]. A 2025 prospective real-time validation study of AI-enhanced EUS for solid pancreatic masses did not show superiority over experienced endosonographers[56]. The strongest argument for EUS AI at present is reduction of operator dependence and shortening of the learning curve, not established pathway change.

A PROSPECTIVE TRANSLATION CHECKLIST FOR FUTURE SYSTEMS, IMPLEMENTERS, AND REGULATORS

The tier narrative above is retrospective: It classifies where current applications sit. The checklist below is prospective: It is intended to help future authors, regulators, implementers, and endoscopy units decide whether a new system should advance to a higher tier. It should be read alongside established AI evaluation and lifecycle standards, particularly CONSORT-AI for randomized studies, DECIDE-AI for early live clinical evaluation, IMDRF Good Machine Learning Practice, and FDA-predetermined change-control planning[17-20].

FUTURE DIRECTIONS

Table 3 outlines the priority translational agenda for the next three to five years. Three themes cut across all tiers: Moving beyond surrogate endpoints, defining downstream pathway consequences of AI outputs, and embedding governance and human-factors safeguards before broad rollout[3-6,13,14,17-20,32,33,49,50].

Among these priorities, three warrant immediate attention over the next three years. First, patient-important outcomes for tier 1 colonoscopy AI, because current ADR-focused evidence cannot distinguish beneficial from low-value detection[7-9,28,32]. Second, pathway-defining trials for tier 2 systems, because technical accuracy has not yet translated into demonstrated clinical value[12-14,45,46,49,50]. Third, governance standardization before societal endorsement, because deployment without named accountability and discordant-case review undermines both adoption and patient safety[3-6,17-20,33]. Global generalizability, human-factors safeguards, and CADx pathway clarification are the subsequent priorities and are retained in Table 3.

Future studies should also report live clinical evaluations in a way that makes systems auditable and comparable across centers, including the model version used, the human-AI interface, input and output handling, override behavior, update procedures, and post-deployment monitoring triggers[17-20]. A further priority is multimodal AI. In pancreatic EUS, a randomized crossover trial of a joint model integrating EUS images with clinical information improved novice diagnostic accuracy from 0.69 to 0.90 across three external Chinese test cohorts[57]. This illustrates a plausible direction for multimodal systems supporting endoscopists, but should not yet be read as an established pathway signal. Reviews of advanced endoscopic imaging similarly suggest that integrating imaging with clinical, molecular, or contextual data may improve decision support, but will also raise the evidentiary bar for interpretability, governance, and dataset representativeness[58].

CONCLUSION

AI in gastrointestinal endoscopy is not a single translational category and should no longer be discussed as one. Commercially deployed colonoscopy CADe remains the only GI endoscopy AI application that clearly satisfies tier 1 across all six domains, but that judgment should not be misread as an unqualified endorsement. Guideline panels remain cautious, evidence generation continues, and the patient-level value proposition is still shaped by the tension between modest detection gains and the added surveillance burden[7-9,21,28-32]. This tier 1 paradox also extends to economic and operational value. Economic value and implementation feasibility remain context-dependent even for colonoscopy CADe, the most mature endoscopic AI application. Markov microsimulation of average-risk US screening colonoscopy suggests that AI-assisted colonoscopy may be cost-saving relative to standard colonoscopy[59], and Canadian modeling in fecal immunochemical test-positive patients similarly reported CADe as a dominant strategy, with higher life-year and quality-adjusted life-year gains at slightly lower cost[60]. However, these estimates remain model-dependent and sensitive to assumptions about adenoma detection gains, surveillance intensity, CADe acquisition or licensing costs, and downstream resource use[32,59,60]. Microsimulation work that explicitly incorporated surveillance burden reached a more tempered conclusion, showing a small and uncertain clinically meaningful benefit, increased use of surveillance colonoscopy, no clear increase in colonoscopy-related harms, and low-certainty evidence[32]. Thus, tier 1 classification should be interpreted as deployment readiness under governance, not as proof of resolved patient-level value, economic value, or universal operational feasibility. Prospective registry work, post-deployment monitoring, and health-system-specific implementation studies should therefore accompany, rather than follow, wider implementation of current CADe systems[3-6,17-20].

CADx for optical polyp characterization remains tier 2 because the most rigorous meta-analyses show no net clinical benefit for the resect-and-discard strategy in routine practice[45,46]. Upper-GI second-observer systems, now including a substantive Barrett’s evidence base, are also tier 2 because workflow consequences and pathway effects are not yet standardized across devices and settings[12,35-40]. Workflow-native quality systems, capsule reader-assist tools, and H. pylori prediction remain near-term candidates rather than settled standards[13,14,47-50]. Cholangioscopy AI, therapeutic or third-space assistance, and EUS-based lesion analysis remain tier 3 because validation is improving faster than evidence of pathway change[51-56].

The next phase of progress will depend less on building more accurate classifiers and more on showing, in representative live practice and under explicit governance, that specific systems improve meaningful procedural work. Investigators, regulators, and endoscopy units should demand not only performance data but also pathway definition, human-factors safeguards, update governance, and post-deployment monitoring aligned with contemporary AI reporting and lifecycle standards[3-6,17-20]. The framework and checklist offered here are intended to make that standard operational for investigators designing trials, regulators evaluating submissions, and endoscopy units making procurement decisions.