After 30 days with an ambient AI scribe, the share of clinicians reporting burnout fell from 51.9% to 38.8% across six US health systems (JAMA Network Open, Olson et al. 2025). The time story is more modest: a five-system study of 8,581 clinicians measured 16.0 fewer minutes of documentation per 8 scheduled patient hours (JAMA, multisite study 2026), and in a randomized trial one scribe trimmed average note time by 41 seconds while the control group dropped 18 seconds on its own (NEJM AI, Lukac et al. 2025). Note quality is the open question, with omissions in 18% of reviewed AI drafts at UC Davis (JMIR Medical Informatics 2026). For the wider documentation market around these tools, see our medical transcription statistics, and for the workload that drives the demand, our emergency physician burnout and hours statistics. We aggregated data from NEJM AI, JAMA, JAMA Network Open, JMIR Medical Informatics, JAMIA, The Permanente Medical Group, the American Medical Association, Healthwatch England, and KLAS Research.
TL;DR
- A UCLA randomized trial gave 238 physicians one of two ambient scribes or usual care; only one scribe beat control on note time (NEJM AI, Lukac et al. 2025).
- Adopters at five academic systems saved 16.0 documentation minutes per 8 scheduled patient hours, but after-hours work did not change significantly (JAMA, multisite study 2026).
- Penn clinicians cut after-hours work from 50.6 to 35.4 minutes per day (JAMA Network Open, Duggan et al. 2025).
- Burnout fell from 51.9% to 38.8% after 30 days of scribe use (JAMA Network Open, Olson et al. 2025).
- Practitioners above the high work-exhaustion threshold fell from 56.1% to 35.4% in a stepped-wedge trial (NEJM AI, Afshar et al. 2025).
- Only 21.1% of 8,581 clinicians had adopted a scribe, from 7.1% to 63.6% by site (JAMA, multisite study 2026).
- AI-drafted notes contained omissions in 18% and hallucinations in 11.5% of 356 reviewed notes (JMIR Medical Informatics, UC Davis 2026).
- Scribe adopters produced 1.81 more relative value units per week, about $3,044 per physician per year (JAMA Network Open, UCSF 2026).
- 61.8% of 12,153 Canadian adults were reluctant about future AI scribe use, and only 28.3% knew the tools were in use (JAMIA 2026).
- Patient consent fell from 81.6% after basic information to 55.3% after detailed disclosure of data storage and vendor involvement (JAMA Network Open, NYU Langone 2025).
- 81% of UK adults want to be asked for consent before an AI scribe is used (Healthwatch England 2026).
1. Randomized Trials: What Controlled Evidence Shows on Note Time
Randomized evidence arrived late in 2025, and it resets expectations. In the UCLA three-arm trial published in NEJM AI, Nabla users shaved 41 seconds off the average note (4 minutes 30 seconds to 3 minutes 49 seconds), but usual-care physicians also shaved 18 seconds without any AI at all, and the DAX Copilot arm did not separate from control. The net effect versus control was 23 seconds per note (41 minus 18), far smaller than the raw before-and-after drop alone would suggest, which is exactly why control groups matter in this category.
Usage explains part of it. Even inside a trial, physicians switched the scribe on for only about a third of their visits, and roughly 15% of physicians assigned a scribe never used it. The University of Wisconsin stepped-wedge trial (a design in which every clinic eventually gets the tool, in randomized waves) saw much heavier use, which is one reason its effects look larger.
| Metric | Value | Source |
|---|---|---|
| Physicians randomized to DAX Copilot, Nabla or usual care (UCLA) | 238 across 14 specialties | NEJM AI, Lukac et al. 2025 |
| Change in average time-in-note per note: Nabla arm vs usual-care arm | -41 seconds vs -18 seconds | NEJM AI, Lukac et al. 2025 |
| Visits in which the assigned scribe was actually used: DAX Copilot arm vs Nabla arm | 33.5% (8,271 of 24,696) vs 29.5% (6,981 of 23,653) | NEJM AI, Lukac et al. 2025 |
| Practitioners in the Wisconsin and Illinois stepped-wedge trial and their note-weighted ambient AI utilization | 66 practitioners, 16,593 encounters; 71.0% utilization (SD 31.6%) | NEJM AI, Afshar et al. 2025 |
| Share of note content typed manually, before vs during scribe use (Penn Medicine, 46 clinicians in 17 specialties) | 11.2% to 7.9% (P < .001) | JAMA Network Open, Duggan et al. 2025 |
| Same-day appointment closure, before vs during scribe use (Penn Medicine) | 66.2% to 72.4% (P < .001) | JAMA Network Open, Duggan et al. 2025 |
Context: the DAX Copilot arm at UCLA fell 23 seconds per note, which was -1.7% versus control (95% CI, -9.4% to +5.9%; P=0.66). A non-significant result in a two-month trial is not proof of no effect, but it is a reminder that scribe products are not interchangeable.
2. After-Hours EHR Time: The Pajama-Time Promise Under Measurement
“Pajama time” is the EHR work clinicians do outside scheduled clinic hours, typically finishing notes at home in the evening. It is the outcome health systems most want to move, and it is where the evidence disagrees most. In the five-system JAMA study, after-hours work changed by only -3.1 minutes per 8 scheduled patient hours, a result the confidence interval (-6.8 to 0.5) cannot distinguish from zero, and the UCLA trial likewise found no significant after-hours change versus control.
The positive results come from the heavily used Wisconsin trial, single-system studies, self-reported surveys, or longer follow-up. Penn’s pre-post study logged a drop of about 15 minutes per day, and a Washington University preprint found that pajama time did not move at all on day zero but was 19% lower by day 150. The practical reading: after-hours savings appear to depend on habit change that takes months, not on switching the tool on.
| Metric | Value | Source |
|---|---|---|
| Work outside work hours per 8 scheduled patient hours, adopters vs nonadopters (5 academic systems) | -3.1 minutes (95% CI -6.8 to 0.5), not significant | JAMA, multisite study 2026 |
| Work outside work, stepped-wedge trial | -0.50 hours/day (95% CI -0.90 to -0.09; P=0.03) | NEJM AI, Afshar et al. 2025 |
| After-hours work per day before vs during scribe use (Penn Medicine, 46 clinicians) | 50.6 to 35.4 minutes (P=.02) | JAMA Network Open, Duggan et al. 2025 |
| Median after-hours documentation time, before vs after adoption (Providence, 1,547 clinicians) | 22.0 to 20.6 minutes; sustained decline of 0.38 minutes per month (P=.02), no immediate change | JAMA Network Open, Husa et al. 2026 |
| Pajama time at day 150 vs before adoption (220 primary care clinicians) | -19% (IRR 0.81, interval 0.73 to 0.90); no change at day 0 | medRxiv preprint, Washington University 2026 |
| Self-reported weekly hours writing notes outside clinic hours, before vs after 30 days | 4.95 to 4.05 hours (-0.90 hours; P < .001) | JAMA Network Open, Olson et al. 2025 |
Outlier note: the Wisconsin trial’s after-hours result became non-significant (-0.19 hours, 95% CI -0.40 to 0.03) once the top 3% of observations were excluded, so a few heavy evening workers drove most of the effect.
3. Burnout, Cognitive Load, and Work Exhaustion
Burnout is the outcome where ambient scribes look strongest, and it moves faster than any time metric. In the six-system Olson study, burnout prevalence dropped 13.1 percentage points in 30 days (51.9% to 38.8%; aOR 0.26, 95% CI 0.13 to 0.54; P < .001), while note-related cognitive load fell from 7.10 to 4.46 on a 10-point scale. One plausible reading is that the relief comes less from minutes saved than from not having to type while listening, which would explain why burnout improves even where log-based time barely changes.
The caveat is design. The Olson study had no control group, and the Mass General Brigham and Emory surveys drew response rates between 11% and 30%, so the most satisfied users may be overrepresented. The randomized trials also show improvement, but smaller: at UCLA, the Mini-Z burnout composite improved by under 3 points on a 10-to-50 scale for both scribes. Physicians already expect this: 70% see AI as a way to automate tasks that contribute to burnout (AMA, Physician Survey on Augmented Intelligence 2026).
| Metric | Value | Source |
|---|---|---|
| Clinicians reporting burnout vs severe burnout, before vs after 30 days with the scribe (6 systems, 186 clinicians in the burnout models) | Burnout 51.9% to 38.8%; severe burnout 18.4% to 12.2% (post hoc, P = .01) | JAMA Network Open, Olson et al. 2025 |
| Note-related cognitive task load, 10-point scale, before vs after | 7.10 to 4.46 (P < .001) | JAMA Network Open, Olson et al. 2025 |
| Practitioners above the high work-exhaustion threshold, trial start vs end | 56.1% to 35.4% | NEJM AI, Afshar et al. 2025 |
| Mini-Z burnout composite (10 to 50), change vs control | DAX +2.83 (95% CI +1.28 to +4.37); Nabla +2.69 (95% CI +1.14 to +4.23) | NEJM AI, Lukac et al. 2025 |
| Emory clinicians reporting a positive effect of documentation on well-being, absolute change at 60 days | +30.7 percentage points (11% survey response rate) | Mass General Brigham, You et al. JAMA Network Open 2025 |
| Mass General Brigham survey response rate at 42 days vs 84 days | About 30% vs 22% | Mass General Brigham, You et al. 2025 |
| Job satisfaction gain at Hattiesburg Clinic: first scribe vendor vs second vendor | +17% vs +13% | AMA, physician AI uses report 2025 |
Context: professional fulfillment in the Wisconsin trial rose 0.14 points (P=0.04), which the authors counted as not significant because each co-primary outcome had to clear P < 0.025.
4. Adoption: How Far Scribes Have Spread Inside Health Systems
Access is not usage. At five large academic systems, only 21.1% of 8,581 ambulatory clinicians had adopted an AI scribe, ranging from 7.1% at Mass General Brigham to 63.6% at Yale New Haven Health, and only about 32% of adopters used it for half or more of their notes. The national survey picture is consistent: documentation is a top AI use case, but fewer than a third of physicians use AI for visit notes.
Where uptake is high, it is habitual. Among 102 surveyed Permanente adult and family medicine physicians, two-thirds used the scribe 5 or more days per week and 63% used it in every in-person visit (The Permanente Medical Group, NEJM Catalyst 2025). The adoption gap between systems is therefore less about whether physicians like the tool and more about how each organization rolls it out: at 4 of the 5 academic systems, adopters opted in themselves, and rollout timing ran from June 2023 to August 2025.
| Metric | Value | Source |
|---|---|---|
| US physicians using AI to document billing codes, medical charts or visit notes | 28% (survey of nearly 1,700 physicians) | AMA, Physician Survey on Augmented Intelligence 2026 |
| US physicians using AI to create discharge instructions, care plans or progress notes | 30% | AMA, Physician Survey on Augmented Intelligence 2026 |
| Clinicians who adopted an AI scribe at 5 academic systems, overall and lowest vs highest site | 21.1% (1,809 of 8,581); 7.1% at Mass General Brigham vs 63.6% at Yale New Haven Health | JAMA, multisite study 2026 |
| Adopters using the scribe for 50% or more of their notes | About 32% (577 adopters) | JAMA, multisite study 2026 |
| Providence clinicians who were active ambient AI users, January 2024 vs March 2025 | 0.1% to 28.8% | JAMA Network Open, Husa et al. 2026 |
| UCSF physicians who adopted a scribe vs share of encounters they handled with it | 44.6% of 1,565 physicians vs 15.2% of encounters | JAMA Network Open, UCSF 2026 |
| Permanente physicians who adopted the scribe in the first 10 weeks of a 2023 rollout | 3,442 of 10,000 offered, across 303,266 encounters | AMA, Permanente pilot report 2024 |
Context: at Providence, active users were only about 8% of all clinicians across the whole study window, a reminder that a system-wide “rollout” can coexist with low average penetration.
5. Productivity, Billing, and the Return-on-Investment Question
The financial case is real but small. UCSF adopters generated 1.81 more relative value units (RVUs, the Medicare unit that measures billable physician work) per week, which the authors translated into about $3,044 per physician per year at 2025 Medicare rates, with no difference in claim denials (JAMA Network Open, UCSF 2026). These are gross revenue estimates, not a measured return on investment.
The billing signal deserves scrutiny. Some of the RVU gain may be better capture of work that was always done, and some may be coding intensity. A KLAS validation of one vendor showed Level 3 established-visit codes falling 18.2% while Level 4 codes rose 7.3% and Level 5 codes rose 5.0% at McLeod Health, and the UCSF authors explicitly flag that future work must separate accurate coding from upcoding.
| Metric | Value | Source |
|---|---|---|
| RVUs per week, scribe adopters vs nonadopters (UCSF, 1.2 million encounters) | +1.81 (95% CI 0.86 to 2.75; P < .001) | JAMA Network Open, UCSF 2026 |
| Patient encounters per week, adopters vs nonadopters (UCSF) | +0.80 (95% CI 0.05 to 1.56; P = .04) | JAMA Network Open, UCSF 2026 |
| Additional weekly visits and exploratory monthly evaluation and management revenue per adopting clinician (5 academic systems) | +0.49 visits (95% CI 0.17 to 0.81), a 1.7% increase; $167.37 per month (95% CI $86.52 to $248.21) | JAMA, multisite study 2026 |
| Median RVUs, before vs after adoption (Providence) | 382.3 to 399.1 (P < .001); no significant change in appointments per day | JAMA Network Open, Husa et al. 2026 |
| Work RVUs at day 150 vs before adoption (primary care) | +2% (IRR 1.02) | medRxiv preprint, Washington University 2026 |
| Diagnosis-coding (ICD-10) compliance score, 0 to 10: ambient AI notes vs human-written notes | 6.87 vs 5.94 (P<0.001) | NEJM AI, Afshar et al. 2025 |
| Time in notes and net monthly gain per provider at McLeod Health (vendor-specific validation of Suki) | -26.8% time in notes; $1,004 net gain per provider per month; Level 3 established-visit codes -18.2% | KLAS Research, Suki ROI Validations 2026 (vendor-specific report, via HIT Consultant) |
Context: the KLAS figures describe customers of a single vendor and are not a category benchmark; FMOL Health, in the same report, recorded a 21% reduction in time in notes.
6. Note Accuracy, Errors, and the Editing Burden
The most detailed real-world audit so far comes from UC Davis, where reviewers read 356 of 7,545 AI-drafted notes. Omissions appeared in 18% of reviewed notes and hallucinations (content the AI wrote that was not said in the visit) in 11.5%, while 5.3% carried errors rated as posing serious or imminent risk (JMIR Medical Informatics, UC Davis 2026). Put differently, 19 of 356 sampled drafts needed a correction that mattered for safety, which is why the authors conclude that clinician review remains necessary.
Error rates depend heavily on who counts and how. An August 2026 preprint audit of three commercial scribes found a verified failure in 31.3% of notes, yet also showed that between 28% and 97% of sampled notes carried a failure depending on the review standard used (arXiv preprint 2026). Earlier, more favorable results came from Permanente’s 2024 pilot, which scored 35 transcripts an average of 48 out of 50 on accuracy and related measures, and the Wisconsin trial, which found stigmatizing language in under 1% of notes (12 notes). The transcription layer underneath these notes has its own accuracy gaps by speaker, covered in our speech recognition accent bias statistics.
| Metric | Value | Source |
|---|---|---|
| Reviewed AI-drafted notes with accidental omissions (UC Davis, 356 notes) | 18% (64 notes) | JMIR Medical Informatics, UC Davis 2026 |
| Reviewed notes with hallucinated content | 11.5% (41 notes) | JMIR Medical Informatics, UC Davis 2026 |
| Reviewed notes with accidental inclusions vs with bias | 9.3% (33 notes) vs 1.1% (4 notes) | JMIR Medical Informatics, UC Davis 2026 |
| Reviewed notes with errors rated serious or imminent risk (severity 4 to 5) | 5.3% (19 notes) | JMIR Medical Informatics, UC Davis 2026 |
| Median share of AI-generated words changed by physicians vs notes left unedited | 9.0% vs 14.9% (143 of 960) | JMIR Medical Informatics, UC Davis 2026 |
| Overall note quality on the PDQI-9 scale, simulated specialist encounters (VA, 2 products) | 36.2 of 50; thoroughness lowest item at 3.4 of 5 | JAMA Network Open, VA simulation study 2026 |
| Notes with at least one verified failure, 3 commercial scribes, 142 consultations (preprint) | 31.3% [27.0, 35.6] | arXiv preprint 2026 |
| Physician-rated frequency of clinically significant inaccuracies, 5-point scale (2 = rarely, 3 = occasionally) | DAX 2.7; Nabla 2.8; one mild adverse event in the trial | NEJM AI, Lukac et al. 2025 |
Context: the VA study named the error types clinicians feared most, including fabricated physical exam findings and omitted pertinent negatives, and one hematology-oncology participant called them “big safety issues”.
7. Patient Acceptance and Consent
Inside US clinics, patients rarely say no. In the Wisconsin trial, 99.92% of patients consented to ambient recording, with only 22 refusals (NEJM AI, Afshar et al. 2025). Public surveys tell a more guarded story, and the gap is mostly about information: comfort drops when patients learn where audio goes and which companies are involved.
The NYU consent study shows the effect directly. Consent was 81.6% after basic information about the technology, but 55.3% after a detailed explanation of AI features, data storage and corporate involvement (JAMA Network Open, NYU Langone 2025). In the UK, comfort falls from 48% for a routine check to 23% for a consultation about domestic abuse (Healthwatch England 2026). Consent conversations in multilingual settings add another layer, covered in our hospital medical interpreter statistics.
| Metric | Value | Source |
|---|---|---|
| NYU pilot patients comfortable with ambient documentation at baseline (103 patients) | 74.8% (77 patients) | JAMA Network Open, NYU Langone 2025 |
| NYU pilot patients who would consent after basic information vs after detailed disclosure of AI features, data storage and corporate involvement | 81.6% (84) vs 55.3% (57) | JAMA Network Open, NYU Langone 2025 |
| NYU pilot patients who did not want their data shared with the vendor | 59.2% (61 patients) | JAMA Network Open, NYU Langone 2025 |
| Canadian adults reluctant about future AI scribe use vs aware that AI scribes are used (12,153 adults) | 61.8% vs 28.3% | JAMIA, Chandrasekaran and Moustakas 2026 |
| UK adults who want clinicians to ask for consent before using an AI scribe (4,039 adults) | 81% | Healthwatch England, YouGov poll 2026 |
| UK adults with a recent appointment who were unaware AI scribing was being used | Nearly 90% of 2,795 | Healthwatch England, YouGov poll 2026 |
| UK adults somewhat or very comfortable with an AI scribe: routine health check vs domestic abuse consultation | 48% vs 23% | Healthwatch England, YouGov poll 2026 |
| Permanente patients comfortable vs uncomfortable with the scribe (118 patients) | Two-thirds vs 8% (26% neutral) | The Permanente Medical Group, NEJM Catalyst 2025 |
Context: UK support for an NHS-wide rollout is split almost evenly, 38% in support and 37% opposed, and strong opposition (21%) is almost twice strong support (11%).
Summary: Ambient AI Medical Scribe by the Numbers
| Metric | Value | Source |
|---|---|---|
| Burnout before vs after 30 days of scribe use (6 systems) | 51.9% to 38.8% | JAMA Network Open, Olson et al. 2025 |
| Practitioners above high work-exhaustion threshold, trial start vs end | 56.1% to 35.4% | NEJM AI, Afshar et al. 2025 |
| Documentation time per 8 scheduled patient hours, adopters vs nonadopters | -16.0 minutes (about -10%) | JAMA, multisite study 2026 |
| Average note time change: Nabla arm vs control (randomized) | -41 seconds vs -18 seconds | NEJM AI, Lukac et al. 2025 |
| After-hours work per day, Penn Medicine | 50.6 to 35.4 minutes | JAMA Network Open, Duggan et al. 2025 |
| Pajama time at day 150, primary care (preprint) | -19% | medRxiv preprint, Washington University 2026 |
| Clinicians who adopted a scribe at 5 academic systems | 21.1% | JAMA, multisite study 2026 |
| Providence active ambient AI users, Jan 2024 vs Mar 2025 | 0.1% to 28.8% | JAMA Network Open, Husa et al. 2026 |
| Physicians using AI for billing codes, charts or visit notes | 28% | AMA, Physician Survey on Augmented Intelligence 2026 |
| RVUs per week, adopters vs nonadopters (UCSF) | +1.81 | JAMA Network Open, UCSF 2026 |
| Estimated added revenue per physician per year (UCSF) | About $3,044 | JAMA Network Open, UCSF 2026 |
| Reviewed AI notes with omissions / hallucinations / serious or imminent risk errors | 18% / 11.5% / 5.3% | JMIR Medical Informatics, UC Davis 2026 |
| Notes with a verified failure, 3 commercial scribes (preprint) | 31.3% | arXiv preprint 2026 |
| Patient consent to ambient recording (US trial) | 99.92% | NEJM AI, Afshar et al. 2025 |
| Patient consent after detailed vs basic disclosure (NYU) | 55.3% vs 81.6% | JAMA Network Open, NYU Langone 2025 |
| Canadian adults reluctant about future AI scribe use | 61.8% | JAMIA 2026 |
| UK adults wanting to be asked for consent | 81% | Healthwatch England 2026 |
Methodology and Sources
We prioritized randomized trials and peer-reviewed health-system studies, used press releases only when they summarized a named paper, and labeled preprints and vendor-specific reports as such. No figure in this roundup comes from VoxBooster data.
- NEJM AI, Lukac et al. 2025, Ambient AI Scribes in Clinical Practice: A Randomized Trial (UCLA), with the UCLA Health release
- NEJM AI, Afshar et al. 2025, A Pragmatic Randomized Controlled Trial of Ambient AI to Improve Health Practitioner Well-Being (University of Wisconsin)
- JAMA 2026, Changes in Clinician Time Expenditure and Visit Quantity With Adoption of AI-Powered Scribes: A Multisite Study
- JAMA Network Open, Olson et al. 2025, Use of Ambient AI Scribes to Reduce Administrative Burden and Professional Burnout
- JAMA Network Open, Duggan et al. 2025, Clinician Experiences With Ambient Scribe Technology (Penn Medicine)
- JAMA Network Open, Husa et al. 2026, Ambient AI Use and Clinician Documentation Burden, Productivity, and Efficiency (Providence)
- JAMA Network Open 2026, Ambient AI Scribes and Physician Financial Productivity (UCSF)
- medRxiv preprint 2026, Longitudinal Effects of Ambient AI Scribe Use in Primary Care (Washington University), not peer reviewed
- The Permanente Medical Group, NEJM Catalyst one-year review 2025, summarized by the Kaiser Permanente Division of Research
- AMA, Physician Survey on Augmented Intelligence 2026
- AMA, report on physicians’ AI uses and the Hattiesburg Clinic pilot 2025
- AMA, report on the Permanente 10-week pilot 2024
- Mass General Brigham, release on You et al. JAMA Network Open 2025 (Mass General Brigham and Emory)
- JMIR Medical Informatics 2026, Quality of Clinical Notes Created by Ambient Listening Generative AI (UC Davis)
- JAMA Network Open 2026, Ambient Scribe Technology in Simulated Patient Encounters Across Specialties (Veterans Health Administration)
- arXiv preprint 2026, One note in three: a verified census of three deployed AI scribes, not peer reviewed
- JAMA Network Open 2025, Informed Consent for Ambient Documentation Using Generative AI in Ambulatory Care (NYU Langone)
- JAMIA 2026, Patient attitudes toward ambient AI scribes in clinical care (Canada survey)
- Healthwatch England 2026, What does the public think about AI scribe use in healthcare (YouGov poll)
- KLAS Research, Suki ROI Validations 2026, vendor-specific report as reported by HIT Consultant
- Data watch: most effect sizes here come from nonrandomized pre-post or observational designs with self-selected adopters, so they are associations, not proof of cause; only the UCLA and Wisconsin trials are randomized. The Penn paper reports a 30.0% after-hours reduction in its abstract but -16.9% in its Table 2, and its abstract says “per workday” where the table says “per scheduled day”. The Olson study converts its weekly after-hours change into 10.8 minutes per workday, a derived estimate, and several of its authors have ties to the vendor whose scribe was studied. The VA simulation paper gives the neurology score as 42.2 in the text and 41.2 in its table. Burnout surveys at Mass General Brigham and Emory had response rates of 11% to 30%. The Washington University and arXiv studies are preprints and may change before peer review, and the KLAS figures describe a single vendor’s customers. The Permanente physician survey figures (102 respondents) and patient survey figures (118 respondents) are small samples from one medical group. The next AMA augmented intelligence survey and longer follow-up from the five-system JAMA cohort, whose most robust estimates reach only about 5 months after adoption, are the most important new data expected.
Last updated: October 24, 2026. We update this roundup quarterly, and the next refresh is expected when the AMA publishes its 2027 Physician Survey on Augmented Intelligence.