Telesight: A Telehealth Visit Copilot for the Three-Phase Encounter

Pre-visit chart preparation, intra-visit clinical decision support that respects the Five Rights, post-visit instructions and coding — augmenting the clinician across the full telehealth lifecycle.

Abstract Telemedicine reached 37.0% of U.S. adults in 2021 (43.3% among those 65+)[1] and never fully receded; Mehrotra et al. document the COVID-era visit-mix shift in JAMA Internal Medicine[2]. The telehealth visit has structural properties an in-person visit lacks — reduced ambient cues, asymmetric attention, no opportunistic chart review during the encounter — and these translate into a different copilot opportunity. Telesight is structured around the visit's three phases. Pre-visit: chart preparation using the Sinsky pre-visit-planning framework[6], AHRQ-recommended AI summarisation[7], and a written agenda. Intra-visit: real-time clinical decision support obeying Osheroff et al.'s Five Rights of CDS[8] with attention to Ancker et al.'s alert-fatigue evidence (acceptance drops 30% per additional reminder[9]). Post-visit: a teach-back-style instruction set (95% of studies show positive teach-back outcomes[11]) and AI-suggested billing codes, mindful that current LLMs achieve only 45.9% exact-match ICD-9-CM coding[15]. Telehealth visits show 29% lower no-show odds[4]; Telesight aims to widen the productivity gap further without degrading attention.

§ 1 Introduction

Telehealth visits are not in-person visits delivered through a camera. They are a structurally different encounter: clinicians cannot fluidly scan the chart while listening, patients cannot read body language, and the visit has a hard endpoint that forces denser communication. The literature confirms operational benefits — telehealth shows 7.5% no-show versus 36.1% for in-office visits[3]; adjusted odds of no-show drop 29% across 2.6 million encounters at Parkland Health[4]; telepsychiatry is equivalent to in-person across PTSD, mood, and anxiety disorders in meta-analysis[5]. The opportunity is to design the AI copilot for the visit shape that exists, not a transliteration of in-person tools.

Telesight covers the three temporally distinct phases of a telehealth visit. Each phase has its own evidence base, its own success metric, and its own failure modes. The pre-visit phase is the most under-invested in current product practice; the post-visit phase is where the literature shows the strongest patient-side returns.

1.1 Contributions

1. A three-phase visit copilot whose pre-visit agenda generation, intra-visit CDS, and post-visit instruction set are each grounded in established evidence.
2. Explicit compliance with the Five Rights of CDS[8] and the Ancker alert-fatigue threshold[9] — design constraints that production AI scribes typically ignore.
3. An evaluation harness that measures pre-visit time saved, intra-visit interruption rate, and post-visit instruction recall (teach-back proxy).

§ 2 Background and Related Work

2.1 Telehealth's Empirical Profile

The CDC's 2022 NCHS Data Brief documents that 37.0% of U.S. adults used telemedicine in the prior 12 months as of 2021[1]. The Patel/Mehrotra cohort study of 16.7M commercial and Medicare Advantage enrollees[2] documents the rapid early-pandemic shift in visit modality, with telemedicine settling at a substantially elevated plateau through mid-2020 rather than reverting to baseline. Drerup et al.[3] document the ~5× no-show advantage (7.5% telehealth vs 36.1% in-office) and Khoong et al.[4] confirm a 29% adjusted odds reduction at safety-net scale. The Shaker et al. JMIR Mental Health meta-analysis[5] establishes equivalence for telepsychiatry across PTSD, mood, and anxiety disorders.

2.2 Pre-Visit Planning

Sinsky et al.'s AAFP guidance[6] reports that pre-visit lab ordering and chart review the evening before can save approximately one hour of physician time per day, with offsetting nursing-time investment. Holdsworth et al.'s 2021 Annals of Family Medicine analysis[7] describes patchy adoption of AI/NLP pre-visit summarisation across ambulatory practice — a substantial deployment gap remains.

2.3 Intra-Visit Clinical Decision Support

Osheroff et al.[8] codified the Five Rights of CDS: right information, right person, right format, right channel, right time. Ancker et al.[9] documented the operational corollary: alert acceptance drops 30% per additional reminder, with 25% of drug alerts being same-year repeats — alert fatigue is the dominant failure mode of intrusive CDS. Tierney et al. at Permanente Medical Group[10] reported on the ambient AI scribe rollout to 10,000 physicians, finding measurable reductions in after-hours EHR time and improved patient-encounter focus.

2.4 Post-Visit Patient Understanding

Talevski et al.'s teach-back systematic review[11] reports that 95% of included studies found positive effects on patient knowledge, adherence, and disease-specific outcomes. Federman et al.[12] document the patient-side reality of the After-Visit Summary: 82.8% of patients recalled receiving it and 67.4% consulted it, but only 31.6% shared with caregivers — a clear gap voice or text follow-up can close.

2.5 Billing and AI Coding

The 2025 CPT update[14] deleted 99441–99443 and introduced 98008–98015 for audio-only E/M; Medicare did not adopt the new codes and instructs use of 99202–99215 with modifier 93 — Telesight must accommodate the policy split. The Soroush et al. NEJM AI benchmark[15] establishes a hard ceiling on autonomous AI coding: GPT-4 reaches only 45.9% exact-match on ICD-9-CM, 33.9% on ICD-10-CM, 49.8% on CPT. Telesight surfaces suggested codes; it does not autonomously code.

§ 3 Proposed Approach

3.1 Three-Phase Pipeline

3.2 Pre-Visit: Chart Preparation Agent

Triggered the evening before the scheduled visit. Generates a one-page agenda: active problems, overdue routine items, medication adherence flags, last visit's open threads. Sends a patient questionnaire (priorities for the visit, symptom updates) and aggregates the response for the clinician. The empirical anchor is Sinsky's ~30-minute-per-day figure[6]; Telesight succeeds only if total clinician time spent reviewing its output is meaningfully shorter than the savings.

3.3 Intra-Visit: Five-Rights-Gated CDS

Telesight intra-visit operates as an ambient listener with a strict prompt budget. The Tierney NEJM Catalyst report[10] on 10,000-physician Permanente rollout informs the architecture: ambient scribe transcription is the substrate, but Telesight adds a CDS overlay that interjects only when (a) a guideline gap is detected, (b) the gap is high-priority by clinical risk, (c) the interjection format is non-modal (lateral notification, not pop-up), and (d) total interjections per visit are capped at two. Each rule corresponds to one of the Five Rights.

3.4 Post-Visit: Teach-Back-Style AVS + Suggested Codes

After the visit ends, Telesight produces a plain-language AVS structured for teach-back: three to five key points with explicit comprehension checks the clinician can review and personalise. The follow-up channel (voice call, SMS, patient portal) is chosen based on the AVS readership pattern Federman et al.[12] document — only 31.6% of patients share AVS with caregivers, motivating direct-to-caregiver delivery for elderly patients. Billing-code suggestions follow the AMA E/M time-or-MDM framework[13], surfaced as recommendations subject to the Soroush ceiling[15].

§ 4 Evaluation Protocol

§ 5 Expected Contributions

1. System. An open three-phase visit copilot with phase-specific design constraints grounded in the established literature.
2. Methodology. A reproducible evaluation harness with metrics per phase rather than a single aggregate.
3. Operating envelope. Documentation of the pre-visit-savings-vs-clinician-review-overhead tradeoff and the intra-visit interjection budget.

§ 6 Limitations and Risks

Telesight's pre-visit chart preparation reduces in-visit cognitive load only if the clinician trusts and reads the output. The risk of a low-trust agent is that the clinician duplicates the work; the risk of a high-trust agent is that they miss what the agent missed. The pre-visit success metric is therefore net time saved, not gross. The intra-visit prompt budget is the most contested design choice and the most consequential — interjection patterns that work in observational pilots regularly fail when deployed at scale due to the Ancker effect[9]. The post-visit AI-coding ceiling is hard: until coding LLMs exceed ~70% exact-match — substantially above the current ~50%[15] — autonomous coding remains unsafe.

§ 7 Conclusion

Telesight treats the telehealth visit as three distinct problems rather than one. Each phase has a different evidence base, a different metric, and a different failure mode. The integration discipline — phase-specific constraints, phase-specific metrics — is the contribution, more than any individual phase implementation.