Overjet + Videa AI — ROI Analysis

What we know

Videa AI is delivering measurable case-acceptance lift at Gen4 West practices. Across 42 treated practices with at least 6 months of post-implementation data, treatment acceptance percentage rose +4.0 points (count-based) and +5.1 points (dollar-based) versus comparable not-yet-treated practices — both statistically significant with clean parallel trends. Average accepted treatment dollars per practice rose +$48K/month (+47.7%). Same-day accepted treatment dollars more than doubled.

Overjet Vision evidence is inconclusive at SGA East. With only 11 treated practices having sufficient data on both PBI and Dental Intel, no case acceptance effect is detectable. Production-per-visit dropped 7% (CI excludes zero), but selection bias makes this hard to interpret — these practices were already declining faster than their peers before AI rollout.

The Brittney question, answered

In the 5/14 call, Brittney raised the question of whether AI tools could lower revenue per visit by surfacing more conservative care plans. Our data shows the opposite for Videa: treatment presented and accepted both went UP, not down. The lift came from more treatment getting accepted, not from per-visit pricing changes (PPV stayed flat). For Overjet, the small SGA East cohort can't answer the question yet — re-run after another 6 months of post-data accumulates.

Recommended actions

What we're missing

• Time clock data — only available since 10/1/25, no pre-baseline. Can't measure clinician time savings.
• Videa feature-level usage — blocked on Dentrix integration (Brittney chasing).
• Overjet utilization data — Brittney pulling.
• Insurance denial rebuttal effectiveness — no instrumentation yet; would need a manual claims-denial sample.
• 16 of 85 treated PBI practices missing from Dental Intel — different practice scope between systems. Would need name-matching expansion or separate DI grant.

Methodology in one paragraph

For each treated practice, we compare its 6-month-pre vs 6-month-post outcome (with a 1-month implementation washout) against the same calendar months for not-yet-treated practices in the same brand group. Modis was excluded as a control because it's mostly oral surgery and periodontics specialty practices, not GP dental. Per-practice effects are averaged within group with bootstrap 95% CIs (n=2,000). Parallel-trends diagnostics flag where pre-period slopes differ between treated and control. Six DI metrics + three PBI metrics analyzed per practice.

Cohort

From the AI Implementation and Pilot Tracker (255 practices), we crosswalked all rows to canonical SGA practice records via paylocity_code. Brand split:

Design: Staggered Difference-in-Differences

For each treated practice, define:

• Pre window = 6 months strictly before live date
• Post window = 6 months strictly after live date
• Washout = the live month is excluded from both

For each treated practice T, the eligible control pool is:

• Same brand group as T (SGA East or Gen4 West) — controls for org/operational structure
• Non-specialty (excludes oral surgery, periodontics, orthodontics, implant centers, etc.)
• Either never-treated or treated more than 7 months after T (not-yet-treated through T's full post window)
• Has activity in both T's pre and post windows

For each (T, C) pair we compute T_diff − C_diff (each measured over the same calendar months). The per-practice effect is T_diff minus the average control_diff. Group effects average across treated practices with bootstrap 95% CIs (n=2,000).

Why Modis is not the control group

Initial design used the 15 Modis practices as a non-AI control. Inspection revealed Modis is dominated by specialty practices: 7 oral surgery (Coastal Jaw), 4 periodontics (Pensacola, Pinellas, Jacksonville x2), and only ~4 GP dental. Specialty practices have fundamentally different revenue profiles, visit volumes, and case acceptance patterns from GP dental — not a credible control. Switched to within-network not-yet-treated practices.

Parallel-trends diagnostic

For each metric we compute the pre-period linear slope of treated vs. mean control. A "PT diff" near zero means the trends were running parallel before treatment (good for causal interpretation). A large negative PT diff for treated means treated practices were already declining faster — selection bias toward struggling practices.

Findings:

• Case acceptance metrics: PT diff is single-digit basis points (essentially zero). Trends parallel. Causal claims defensible.
• Production per visit: PT diff -$11/month for Videa cohort. 72% of treated practices were declining faster than controls. Selection bias — AI rollout was prioritized at struggling practices. Causal claims NOT defensible on PPV.
• Visits: PT diff +$15/month — treated practices were growing visits FASTER than controls pre-treatment. Direction matters: this biases visit-effect estimates downward (we're under-reporting any positive AI effect on visits).

Coverage caveats

• 16 of 85 treated practices in PBI are not in Dental Intel — different practice scope between systems
• Time clock data excluded — no baseline before 10/1/25
• Feature-usage data deferred — Brittney chasing Videa data; Dentrix integration blocked
• Insurance denial rebuttal effectiveness — no instrumentation yet
• SGA East / Overjet Vision n=13 (PBI) and n=11 (DI) — small sample limits statistical power. Wave-2 SGA East practices need ~6 more months of post-data.

Source

• Cohort: AI tracker + master crosswalk → cohorts.json
• PBI: bridge query against dataset e3fcdf32-... → panel.csv (8,320 rows)
• DI: Playwright pull from portal.dentalintel.com → monthly.csv (5,304 rows)
• Analysis: per-practice DiD with bootstrap CIs and parallel-trends diagnostic
• Meeting context: Brittney + Scott, 2026-05-14 14:21, 11 min