The Outlook Guys Found a $100 Million Problem Hiding in Your Doctor's Phone Queue

The first thing to understand about Pretty Good AI is that its name is a feint. The software is allowed to sound conversational, but it is not allowed to improvise with a patient's chart. The company builds voice and workflow agents for medical practices running athenaOne. They pick up calls, identify what a patient wants, follow that practice's rules, complete routine work in the electronic health record and hand uncertainty to a person. The ambition is not to imitate a doctor. It is to become the competent operator every front office wishes it could clone at 9:03 on Monday morning.

That makes Pretty Good AI part phone system, part workflow engine and part deployment consultancy. Scheduling is the obvious entrance. Behind it sit prescription refill requests, referral status, insurance eligibility, prior authorization follow-up, medical-records requests, patient balances, waitlists, no-show recovery and after-hours coverage. The company packages those jobs as modular building blocks across patient access, capacity, revenue protection and operational measurement.

The customer is not a patient shopping for an app. It is the practice manager, operations chief or information-systems leader trying to keep a multi-location medical group moving. The focus is particularly sharp: organizations already running athenaOne. That constraint excludes a large portion of healthcare, but it gives Pretty Good AI permission to learn one system deeply. Company materials describe hundreds of bidirectional API connections. The point is not that the agent knows an appointment was requested. The point is that the appointment appears in the correct schedule, under the correct type, with the correct notes and a visible audit trail.

The call is only the wrapper

Voice AI demos tend to obsess over the voice. Does it pause naturally? Can it handle an interruption? Does it sound less like a motel alarm clock? Those details matter, especially when patients may be anxious, older or calling from a noisy car. Pretty Good AI's more defensible idea is that the call is simply the wrapper around an operational transaction.

Consider a patient who wants to reschedule, correct an insurance record and ask whether a referral arrived. A generic bot can collect three messages. A useful operations layer must authenticate the caller, find the right chart, understand provider and location rules, inspect open referrals, update structured fields and create a clean exception for anything it cannot finish. A pleasant transcript is not the outcome. A correct chart is.

This is where the company separates itself from EHR-agnostic tools. Broad compatibility sounds reassuring on a sales slide. In production, every EHR exposes different objects, permissions and quirks. Pretty Good AI has chosen depth over breadth, supporting a large share of athenaOne's available interfaces and building around the ugly edge cases. It is a strong wedge and a real dependency. If athenahealth changes access, improves its own automation faster or a customer switches systems, the specialization can turn from moat to wall.

What failed first? The spreadsheet

The founders' most useful admission is that the model was not usually the first thing to break. The practice was. One location calls an appointment type “new patient.” Another calls the same visit “NP,” while a third reserves the label for a different specialty. A doctor works Tuesdays except on alternating weeks. A referral is “received” to one employee and “ready” to another. Humans patch over these disagreements with memory, Slack messages and heroic front-desk intuition. Software makes them visible.

“The bottleneck is alignment, not intelligence.”Kevin Henrikson, on the lesson from AI deployments

That changed the team's mind about implementation. Kevin Henrikson has written that he once expected the breakthrough to come from better prompts or models. The company now assumes roughly 80 percent of deployment is alignment: auditing the spreadsheet, normalizing appointment definitions, mapping providers to locations and documenting escalation paths. The language model may be the glamorous component, but configuration determines whether the work gets done.

This is the most portable part of the playbook. Before buying an agent, take ten common requests and write the rules as if the new hire had no tribal knowledge. Which inputs are required? Which system is authoritative? What may be changed automatically? What always needs a human? Who owns the exception, and how quickly must it be handled? If two experienced employees disagree, the agent is not ready. You have discovered an operations problem wearing an AI costume.

The numbers, with an asterisk

Pretty Good AI's clearest production example is a single specialty practice with more than 30 locations. According to the company, the deployment processed more than 70,000 calls per month, contained about half of them within roughly 30 days and brought peak patient hold times from more than two hours to below five minutes. Routine scheduling, refill and records work could be resolved without a staff member, with exceptions escalated.

Customer-reported results from one 30-plus-location practice. Workflow, staffing, seasonality and call mix can change the outcome.

Those figures are a case study, not a randomized trial or a promise. Still, they point to the right buying metrics. Count completed workflows, not calls answered. Audit what happened to uncontained calls. Measure abandoned calls, time to human handoff, incorrect writes, after-hours bookings and staff hours returned. Then read the chart. If an insurance correction landed in a free-text note instead of the policy record, the agent did not complete the job.

The economics are less public. Pretty Good AI does not publish a rate card. Pricing is tailored to the work automated, and the company says every contract is month-to-month, with no cancellation fee. That is meaningful in enterprise healthcare, where one-to-three-year commitments are common. It also creates a blunt weekly test: did the phone queue shrink enough to justify another invoice? According to information supplied by the company, it is bootstrapped and profitable. No institutional funding round or valuation has been publicly announced.

Old teammates, new operating system

Pretty Good AI was founded in 2025 in San Francisco by veterans of unusually large systems. JJ Zhuang and Kevin Henrikson helped found Acompli, the mobile email company Microsoft acquired in 2014 and turned into the foundation of Outlook Mobile. They later held senior technical roles at Instacart, where logistics, demand and exceptions arrive continuously. Company-provided background also names Anand Bollini as a co-founder.

The personal reason is closer to home. Henrikson has described his father's diabetes and his wife's repeated battles with breast cancer. Zhuang grew up with physicians in his family. Both encountered a healthcare system where access still depends on hold music, voicemail and phone tag. Their thesis is that clinicians' time is scarce while administrative friction consumes it by the teaspoon.

The culture advertised on the careers page matches the product strategy: small team, low ego, ship in days, stay close to customers. Engineers are expected to listen to staff, map written and unwritten rules, work with APIs and remain beside a practice during launch. This is less like dropping off software and more like installing a tiny operating system inside an organization that cannot close for renovations.

Where the bet works - and where it does not

The product fits practices with enough repetitive volume to justify automation, rules that can be made explicit and leadership willing to clean up inconsistencies. Multi-location groups are especially attractive because the same agent can absorb simultaneous demand without adding another row of headsets. A clinic that loses patients after hours or buries trained medical assistants in routine calls has an obvious starting point.

It will not work everywhere. Pretty Good AI explicitly positions the service for healthcare organizations, not patient medical advice. Clinical interpretation should remain with clinicians. Practices with undocumented policies, chaotic records or no owner for escalations will automate their confusion. Patients may also resist an artificial voice, particularly in sensitive situations. The answer is not to trap them in a smarter menu. It is to disclose appropriately, make human transfer easy and reserve automation for work where consistency beats improvisation.

The competitive field is crowded. Hyro, Syllable, Assort Health, Notable, Infinitus, CallMyDoc, OhMD and others attack pieces of patient access or revenue operations. Horizontal voice platforms let technical teams build their own agents. athenahealth itself is adding more AI. Conventional call centers and another receptionist remain valid alternatives. Pretty Good AI's response is not a general claim that its model is smarter. It is the depth of its athenaOne execution, the breadth of administrative blocks and a contract short enough to punish weak results.

The boring future is the point

The company calls itself an AI operations layer, a phrase that risks sounding larger than life. In practice, its best work should feel smaller: the phone gets answered, the slot gets filled, the referral moves, the insurance record is correct and the human sees the tricky case with context attached. No grand reveal. Just fewer loops.

That is also the idea worth copying. Pick a painful channel where demand arrives faster than people can process it. Integrate deeply with the system where the final action must live. Turn unwritten habits into explicit rules. Give uncertainty a clean exit. Measure completed work. Then make the product earn renewal before the customer forgets why it bought it. Pretty good, in this market, is a demanding standard.