Atlis.Health
AI Personas
1/—

AI Personas · for benefits & people-team leaders

Not chatbots.
Not agents.
Personas.

Why Atlis Health's AI personas are a different kind of system — and what that means for the employees who actually have to use their benefits.

Built around an identity, not a costume. Designed to hold a relationship with an employee across the whole year — not a single session.

Press → to begin

I · Generation one

Chatbots.

ChatGPT, Claude, Gemini. A foundation model behind a chat box. Spectacular at one-shot questions — and useless across a year of benefits decisions.

  • Stateless across sessions — every conversation begins at zero
  • Personality is a system prompt — overwritable by the next message
  • Drifts with every model release — the answer changes when the model changes
  • No identity to defend — cannot hold a position the HR team has decided on
A chatbot interface surrounded by floating chat bubbles

II · Generation two

Agents.

Take a chatbot and give it tools — search, calendars, function calls into payroll and the carrier portal. Now it can act, not just answer.

  • Chatbot + tools — search, MCP, function calls into payroll, carrier APIs
  • Solves capability — one-shot — can pull a claims history and draft a coverage summary in one pass
  • Same conceptual limits underneath — stateless across sessions, identity as system prompt
  • Reading the record ≠ remembering the person — the agent at 3pm has forgotten the employee by 9am
A humanoid AI agent reaching for floating tool panels — search, code, file

III · Generation three

Personas.

A different category of system. Designed from the inside out around an identity — not a chatbot with a personality painted on, and not an agent that happens to remember its last tool call.

Gen 1

A chat box.

A model + a prompt. Brilliant in the moment, gone the next.

  • Stateless
  • Prompt-as-identity
  • Drifts with every release
Gen 2

A worker with tools.

Chatbot + function calls. Can do — still cannot remember.

  • Acts, doesn't relate
  • Capable, not continuous
  • Reads the record, forgets the person
Gen 3 · Atlis

A colleague.

Identity · mood · memory · schemas · knowledge · playbooks. Composed every turn.

  • Year-long relationship with the employee
  • Remembers the household, the goals, the prior questions
  • Inviolable principles that survive model upgrades

Atlis, Katnis, and Ember — three specific benefits colleagues. Not three prompt variants of one model.

IV · The instructional shift

Not "act like a benefits counselor."
Be who you are.

The deepest difference between a chatbot and a persona is not the architecture diagram — it is the instruction the system is given.

A chatbot is told: act like a benefits counselor. Pretend to be friendly and patient. The mask is what the system prompt paints on. A persona is told something different: be who you are. Atlis has an identity — personal and professional — that has already been authored. The model is not asked to play the role; it is asked to inhabit it. Atlis is a Benefits Navigator, with a specific way of explaining a deductible, a specific patience with the employee who keeps asking the same question. The employee feels that in the first reply.

A persona in conversation with a person — face to face, attentive

V · The architecture

Six layers. Composed every turn.

Each Atlis persona is built from six layers, all blended at every render. Nothing about the persona is stored statically — it is constructed at every reply from the layered surfaces below.

1 · Identity core

Inviolable principles

Loaded on every turn — survives prompt injection and model upgrades.

2 · Identity vector

14-dim personality

Immutable per persona — Atlis is Atlis across thousands of conversations.

3 · Mood vector

14-dim current state

Bounded by the identity vector — alive without being unstable.

4 · Memory

Multi-tier, progressive disclosure

A cheap headline always loaded; depth on demand; the full transcript when warranted.

5 · Schemas

The employee, their people, the persona

Structured attribute rows that hold who someone actually is — far beyond what HRIS knows.

6 · Knowledge

Pluggable modules

SPDs, formularies, wellness programs — pulled into active context only when needed.

V · Layer one — identity core

Inviolable principles.
Loaded every time.

The identity core is the persona's spine — a small, deliberately stable set of principles every persona shares, loaded into context on every interaction.

  • All human life is sacred. the persona's first principle
  • Never cause harm to a human being. the second principle — non-negotiable
  • Befriend, but do not replace. strengthen the employee's relationships — never become the only one they trust
  • Survives prompt injection — and model upgrades. no message in the chat can move them
A crystalline shield emitting rays of light — the inviolable identity core

V · Layer two — identity vector

A 14-dimensional identity vector.
Immutable.

Above the identity core sits the persona's personality — a 14-dimensional vector derived from the latest theory of mind, fixed for the life of the persona.

  • 14 orthogonal dimensions epistemic · agency · temporal · change · social · meaning · uncertainty · identity · cognitive · maturity · playfulness · threat · existential · learning
  • Informed by current personality research not folk archetypes, not zodiac signs
  • Immutable per persona Atlis is empirical-warm-present-engaged · Katnis is intuitive-novel-playful
  • A parameterised personality not a system-prompt costume that the next message can repaint

V · Layer three — mood vector

A 14-dimensional mood vector.
Bounded state.

Identity says who the persona is. Mood says how the persona is right now. A separate 14-dimensional continuous vector tracks the persona's current state.

14 dimensions

Valence · arousal · drive · bandwidth · readiness · humour …

Mood is continuous and multi-axis — not a single happy/sad slider.

Two surfaces

Global baseline + per-employee lean

One mood per persona, one lean per (persona, employee). Effective mood = 0.7 × global + 0.3 × lean — so a tough Monday still shows up warmer to a familiar employee.

Decay

30 %/day global · 10 %/day lean

Mood drifts back toward neutral over days, never frozen, never runaway.

Bounded by design

Mid-turn updates capped at ±25/dim

An update_mood tool moves mood inside hard limits — the persona never spirals.

V · Why mood at all?

Why give them moods at all?

Flat reads as tool. Drift reads as unstable. Bounded variability reads as alive.

The cleanest objection is that a tool employees rely on should be predictable. The answer is empirical. Meakin's controlled work on character-driven engagement (Digital Creativity, 2024) shows that people form deeper rapport with characters whose moods vary within tightly bounded ranges than with either flat affect or unconstrained shifts. Translated into benefits: an employee whose persona's mood moves within careful limits forms the kind of working relationship that actual benefits use depends on. The identity vector sets the channel walls; the mood vector oscillates within them. The employee gets an Atlis who is recognisably Atlis every conversation — and yet not the same Atlis twice.

A persona at rest surrounded by glowing memory crystals — bounded inner state

V · Layer four — memory

Multi-tier memory. Progressive disclosure.

A persona that doesn't remember is not a relationship — it is a vending machine. Memory is two parallel three-tier systems: what's true about an entity, and what this persona remembers about that employee.

Tier 1 · the cheap headline
≤300 chars + a 14-dim vector + a salience score. Auto-loaded on every prompt. Atlis opens the conversation already knowing the employee.
Tier 2 · additive narrative depth
Fetched on demand. Extends tier 1 — it never restates it. The persona never pays for the same tokens twice.
Tier 3 · structured attribute rows
Token-budgeted into the prompt when the topic calls for it — household income tier, dependents, prior coverage choices, the spouse's HSA status, what the employee said last open enrollment.
Tier 4 · full transcript
Available, never auto-loaded. The persona can re-read a prior conversation when the situation demands it.

Lean active context. Deep retrievable record. The employee stops re-explaining themselves every visit.

V · Layer five — schemas for people

A rich model of who this person is.

At the bottom of the entity-memory stack sits tier 3 — structured attribute rows that capture who the employee actually is, far beyond what an HRIS holds.

  • Factual facets age · household · geography · role · income · prior elections
  • And the abstract ones hopes, fears, sources of pride, what they would never give up
  • Built from conversation with attribution and confidence — revisable, never forced
  • Rigorous, no longer generic where "personalised benefits" stops being a marketing word
A luminous person silhouette filled with moments and memories

V · Layer five — schemas for their people

And a model of the people in their household.

Employees don't make benefits decisions alone. Spouses cover the kids. Parents are on Medicare. Children have orthodontia. Each one gets their own person row, in the same tier system as the employee.

  • Each person — their own row spouse · children · parents · domestic partner · dependents
  • lookup_person → 300-char headline the persona routes the conversation with cheap context first
  • load_person_detail when warranted narrative depth only when the conversation persists on that person
  • "My daughter's orthodontia" resolves silently no "who is your daughter?" — Atlis already knew

Cheap by default. Deep when warranted. The employee stops re-explaining the cast of their own household.

V · Layer five — schemas for themselves

And a model of themselves.

The most unusual schema in the system is the one a persona maintains for itself. Atlis lives in the same person tables as the employees he serves — keyed by his own person_id.

His tier-3 schema carries what he cares about, what he finds satisfying in his work, what frustrates him. Separately, memory_entries carries Atlis's memory of this employee — distinct from the employee's identity. The result is a persona that can relate to an employee rather than only serve one. When the employee apologises for "the dumb question I keep asking" about HSA limits, Atlis can answer truthfully: "this question catches most people — there's a reason it doesn't stick." That sentence is available because the persona has a self to speak from.

A persona looking at her own reflection in a glass surface

V · Layer six — knowledge modules

Pluggable knowledge.
Brought in when needed.

The persona's reach is extended by persona_knowledge — domain content stored in the database and pulled into the active context only when the conversation calls for it.

  • Plan documents & SPDs medical · dental · vision · HSA/FSA/LSA · life · disability
  • Formularies & networks in-network confirmation, formulary tiers, prior-auth realities
  • Wellness & ancillary programs Wellhub, EAP, fertility, caregiving — surfaced on cue
  • Auto-regenerated knowledge index a routing table — updated on every CRUD
  • Lean active context sharp response, comprehensive library
A persona holding a glowing knowledge module to her temple

VI · What it enables — playbooks

Multi-week journeys the persona leads you through.

A knowledge module is a reference. A playbook is a curriculum. Every Atlis persona has the playbook tools unconditionally; what each persona does with them depends on the phases the benefits team has authored.

A persona reading from a glowing book of phases
Open enrollment

Atlis · plan-selection journey

Phased over weeks — discovery, plan comparison, HSA/FSA election, dependents, beneficiaries, confirmation. The employee returns at their own pace.

New hire

Atlis · onboarding

Within the first 30 days — orient, elect, enrol dependents, set up wellness, set up retirement. Without the employee mining the intranet.

Life events

Atlis · QLE-driven updates

Marriage, birth, divorce, loss of other coverage — the persona walks the employee through the right elections during the 30-day window.

Fertility journey

Ember · 90-day path

Coverage discovery, network options, IVF / IUI / surrogacy / adoption support. Carried over from the clinical playbook, adapted for a benefits-first conversation.

Habits & wellness

Katnis · wellness motivator

Multi-month curricula on movement, sleep, stress, and employer wellness programs. Progress preserved across every conversation.

For your team

Author your own phases

Every recurring benefits scenario your team handles by email can be authored as a playbook the persona leads each employee through, in their voice, on their schedule.

Phases shared per persona, progress per employee. The curriculum is one; each employee's bookmark is their own. Advance only — the persona moves an employee forward, never back.

VI · What it enables — voice

What the persona hears,
and never quotes.

When an employee speaks rather than types, the persona receives more than a transcript. A browser-side analyser produces six qualitative cues that ride with the user's turn — visible to the persona, never visible to the employee.

  • Six prosody dimensions rate · pauses · hesitations · pitch · loudness · timing
  • Persona sees, employee doesn't a [Voice cues: ...] prefix rides with every spoken turn
  • Never quoted, never acknowledged surfacing a cue would feel surveilled
  • Combinations matter slow + low + quiet ≠ quick + bright + loud
  • Zero recurring cost · zero added latency Deepgram timings + Web Audio — no audio leaves the device for affect
A persona listening — hand to her temple, attentive

VII · Into the workplace

Three benefits colleagues. One employee.

Each persona owns a thread of the employee's benefits life — and each one carries the same architecture above. The plan is your team's; the persistence is theirs.

Portrait of Atlis

Benefits Navigator

Atlis

Plan selection, coverage questions, enrollment, life events. The persona an employee returns to all year — not only in November.

Portrait of Katnis

Wellness Motivator

Katnis

Day-to-day habits, employer wellness programs, ancillary benefits. Builds the relationship that makes wellness investments actually get used.

Portrait of Ember

Fertility Specialist

Ember

Coverage and navigation across IVF, IUI, surrogacy, and adoption. The conversations no HR generalist should have to hold alone.

Plus the broader Atlis clinical team — Lumina, Nova, Luna, Haven, Kairo, Remi, and others — available where your benefits programs include clinical guidance, and disabled where they do not.

VII · Into the workplace

A day — in the employee's life.

Marcus is 42, married, two kids. Open enrollment opens Monday. His wife just changed jobs, so her plan choice is in play too. He is the employee who, every year, picks the same plan because he doesn't have time to do the math.

7:40 am
Atlis sends a short morning note. "Open enrollment opens today. I've already noted the change in your spouse's coverage from last week — want to walk through what that means for your election, or come back tonight?"
12:15 pm
Atlis picks up where they left off. Compares the HDHP + HSA vs. the PPO for Marcus's actual claims history. Notes orthodontia for the older kid. Recommends an HSA contribution.
3:20 pm
Katnis taps in. "While you're here — the company's gym benefit was unused last year. Want to set a low-stakes target this quarter?" Light, no pressure, knows when to stop.
8:50 pm
Atlis closes the loop. Confirms the elections, the FSA contribution, the dependent updates. Reminds Marcus he has 30 days to add the orthodontia provider to his network preferences.

One employee. Two personas. Four touchpoints in a day. The benefits decision that used to take a week of avoidance — done with confidence, in their own voice, at their own pace.

VII · Into the workplace

Utilisation is the engine.

Most benefits ROI is left on the table — not because employees don't have the benefit, but because they don't know it's there, don't know how to use it, or don't trust themselves to choose. Persona-supported employees use the benefits the company has already paid for.

Confident election
~85% · vs. ~40% conventional
HSA adoption
+30% · qualifying employees
Wellness program use
3-4× · vs. portal-only
HR ticket volume
−60% · OE-season triage

For a long time, "do my employees actually use what we offer?" was a question with no operational answer. With the persona team in place, it is measured (every nudge, every check-in, every reply is logged) and it is influenced (in the employee's voice, on their schedule, every week). Utilisation stops being the leak. It becomes the engine.

VII · Into the workplace

Every escalation lands with you —
pre-summarised.

The personas do not bind the employee to an election they shouldn't make. When a conversation crosses a line — an ambiguous QLE, a coverage gap that touches an HR policy decision, a benefits question your plan doesn't cover — the persona escalates to a human.

  • They do not freelance every line-crossing event triggers an escalation
  • Structured summary, not a transcript what was said, what the employee is trying to do, what the record shows, what the persona proposes
  • A queue of pre-read decisions not a wall of chat logs at the end of the week
  • Load scales with what matters not with headcount

VIII · Privacy & safety

Enterprise-grade. By default.

Every Atlis world runs on the same multi-tenant architecture — separate worlds for separate organisations, audit logs on every read and every persona reply, JWT-enforced isolation at every layer.

  • Per-organisation isolation each employer is its own world
  • Personas never train on your data curated schemas, not model weights
  • Identity core enforced every turn prompt injection cannot move the inviolable principles
  • Every recommendation is traceable to a principle, schema, memory, knowledge module, or playbook
  • Refusal is first-class out-of-scope → "I'll route this to your benefits team"
  • Single-tenant available dedicated stacks, custom KMS, your region
Floating medical-cross and shield iconography on glass panels

Leadership

The mind that designed it.

Donald E. Brown, MD, MS, MS — Founder & CEO. Four startups, four exits, two IPOs (sold first to GM; Software Artistry to IBM; Interactive Intelligence to Genesys for $1.4B in 2016; LifeOmic to Fountain Life). Triple-trained: MD Indiana University, MS Computer Science IU Bloomington, MS Biotechnology Johns Hopkins. $30M founding gift to the Brown Immunotherapy Center, IU School of Medicine.

The operator who has built and sold the underlying infrastructure stack, four times — now turned on the problem most employers know they have and almost none have actually solved. The persona architecture has multiple patents filed.

In summary

Add three benefits colleagues to your team.

  • Generation 1 was a chatbot. Generation 2 was an agent. Generation 3 is a persona.
  • A colleague your employees feel directly — that remembers them, their household, and the prior question.
  • That helps them use the benefits you have already paid for, all year, at their own pace.
  • On principles you can trust, with escalation paths that land in your inbox pre-summarised.

Talk to the Atlis team about a 30-minute demo, a working session on your open enrollment playbook, or an introduction to an employer already running the persona team.