Personas, not chatbots
Atlis’s benefits colleagues are built around a fixed identity and each employee’s own records, and composed fresh on every turn.
Use ← → to turn the sheets.Swipe, or use the arrows below, to turn the sheets.
AtlisBenefits Navigator
KatnisWellness Motivator
ClaraMedical Claims & Benefits Resolution Advocate
AmyliaCommunity Support NavigatorThree generations
A chatbot is a prompt and a model. It forgets you between visits, and its personality is whatever the latest model release makes of the prompt.
An agent adds tools, so it can look things up and act. It still has no settled identity and no record of the person it is helping.
A persona adds a fixed identity, a memory of the person, and that person’s own records, behind tools that can only reach them.
What a persona is made of
A persona is a bundle of files, versioned with the product. Atlis’s own is below; the clinical colleagues arrive in the same shape from the engine.
An Atlis administrator can adjust a persona’s portrait, their identity and their records in the console. The change rides on top of the bundle, so everything nobody touched keeps following new releases.
persona.jsonname, title, card, voice, and which records and modules loadidentity/Core Personality · Conversational Restraint · Role · How I Answer · Boundariesdimensions.jsonsixteen identity dimensions, each with a confidenceknowledge/what they know: one module always loaded, three read when neededassets/their portraittheir programan eight-phase playbook, published from the console’s catalog
Composed every turn
Nothing is left over from one turn to the next except what is written down. Each turn, the engine assembles what the persona reads, in this order.
- Dimension 1: 30
- Dimension 2: 25
- Dimension 3: 60
- Dimension 4: 35
- Dimension 5: 72
- Dimension 6: 75
- Dimension 7: 35
- Dimension 8: 25
- Dimension 9: 45
- Dimension 10: 70
- Dimension 11: 55
- Dimension 12: 68
- Dimension 13: 72
- Dimension 14: 40
- Dimension 15: 40
- Dimension 16: 50
An identity that holds still
Sixteen dimensions, drawn from the latest Theory of Mind research, give each persona their own personality and their own way of communicating with the person in front of them. They do not drift with the conversation or with a model release; only an Atlis administrator changes them.
Four colleagues, four shapes
The same sixteen dimensions, in the same order, as on the previous sheet. Four different shapes: four personalities, and four ways of talking with people.
A mood that moves, and settles
- The person’s mood, fourteen dimensions from valence to openness, read from how they write. It moves at most 25 points in one turn.
- The persona’s own mood toward that person, which moves at most 15 points in a turn.
- Both settle on their own. Left alone, a change halves every three days, back toward rest.
Each is kept per person, inside that customer’s partition.
A memory of the person
- Memories in three depths. A one-line gist, a summary and the full note, so the persona carries the gist and reads the rest when it matters.
- The people they mention. A spouse, a child, a doctor, kept so they never have to be explained twice.
- Past conversations. Searchable, and readable back word for word when an earlier detail is needed.
- Follow-ups. When a persona says they will pick something up later, the note comes back the next time the topic does.
- The last twenty exchanges are always in view; anything older says it is on record rather than vanishing.
The employee’s own record, and nobody else’s
The persona reads the employee’s figures, documents, the text of their plans and their history through tools bound to the signed-in person. None of those tools has a place to name anyone else.
Documents arrive already resolved: if the employee’s plan replaces the company default, the persona never sees the one it replaced.
- read
- document_text
- document
- the one the documents list just showed them
- query
- “speech therapy”
- employee
- no such field: it is the person signed in
A paraphrase of a real call; the persona can only cite a document they have been shown.
A change waits for a yes
When a persona wants to change something on the employee’s record, they cannot write it. They show a card with exactly what would be written, and stops. Nothing is stored until the employee approves that card.
- Balances the employee holds, such as what is left in their HSA.
- Plan documents they pasted, kept on their record so they can be read all year.
- The persona is told what is still waiting, so they never ask twice for something already decided.
- Document
- BlueShield PPO Gold, Summary of Benefits and Coverage, 2027
- Kind
- Medical plan
- Starts with
- “Summary of Benefits and Coverage: What this Plan Covers”
- Ends with
- “…see the next section for more information.”
- Kept as
- a plan document on your record, which your benefits administrators can see, as they see the plan documents they publish
An illustrative plan; the five fields and the last one’s wording are the card’s own.
What they know
Knowledge comes in named modules. One shapes every answer; the rest are found when a question needs them, and read in full.
| Atlis’s modules | Loaded |
|---|---|
| Tone of an informed advocate | every turn |
| HSA and FSA contribution limits, 2026 | when needed |
| HSA and HDHP limits, 2027 | when needed |
| ACA marketplace eligibility, 2026 | when needed |
Your administrators can publish modules of your own, for your organization alone, and scope them to particular personas.
A program to follow
Each benefits persona works through a program of phases, recording what they learn as they go. Atlis’s is the open-enrollment and claims walk-through below; the program picks up where it left off in the next conversation.
Katnis runs a wellness program; Clara keeps a case record for a disputed claim; Amylia works through community support.
Programs live in the console’s catalog, where Atlis edits them and they reach every persona without a release.
A phase moves on only along routes the program allows, and some routes wait for a recorded fact.
Arithmetic is a tool’s, not a guess
A persona never works a deductible through in their head. The numbers that decide a plan come from tools that compute them and show their breakdown on screen.
| A year under each plan | premiums and what the household would pay, service by service, with whose deductible is met when |
|---|---|
| HSA qualification | whether a plan qualifies for the year being chosen, against that year’s IRS limits |
| An EOB, line by line | checks the arithmetic, the plan’s rate and the totals, and names each finding as a dispute, a question or in their favor |
| Dates and single figures | an appeal deadline, a monthly figure from a yearly one, computed rather than typed |
The rules come last
Atlis’s standing rules are placed after everything else the persona reads, every turn, where nothing earlier can talk past them. The persona engine measured the difference when it introduced this position (engine release 0.223.0, 2026-09-18): the same rule held 8 times of 8 placed last, against 3 of 10 inside the persona’s own text.
- “Whether a plan qualifies for an HSA comes only from check_hsa_qualification … never take the label on trust.”
- “Call end_conversation only when the employee’s message is purely a goodbye.”
- “A card not listed there has already been approved or declined: never say it is waiting.”
A team, not one assistant
AtlisBenefits NavigatorKatnisWellness MotivatorClaraMedical Claims & Benefits Resolution AdvocateAmyliaCommunity Support Navigator
Four benefits colleagues, and fourteen clinical ones for health questions. Each organization chooses which its people can reach, and a colleague offers another when a question belongs elsewhere, with the reason; the employee decides.
They speak, and listen
- A voice of their own, with backups at other vendors already chosen, so they can still be heard if one is down.
- Listening through Deepgram or Amazon Transcribe.
- No dead air. Measured 2026-09-24 over a five-minute conversation: the longest silence was about four and a half seconds, always the opening, while short fixed acknowledgements are served from a cache.
- Never stuck. If speech fails, the answer still arrives as text.
The model is a setting
Each persona is set to a model family in the console, and the engine pins the version. Claude runs on Amazon Bedrock, in US regions only. Atlis answers on Claude Sonnet.
- Claude: Opus and Sonnet, through Bedrock.
- OpenAI, Google Gemini and xAI Grok, available per persona.
- A change takes effect within a minute, with no release.
- Every engine upgrade is measured before and after, on the same full benefits conversation.
Private by construction
- The employer does not see the conversation. Every persona is told so, and nothing in the console reads conversations.
- Each customer’s conversations and memories live in its own partition, reached with credentials that cannot open another’s.
- The persona engine runs only in licensed AWS accounts.
Tested like a colleague
A simulated employee works through a full open-enrollment conversation with Atlis: their household, two plan summaries, an EOB, a cost comparison and a summary to take home. Another model grades it on accuracy, tool use and how it reads.
Atlis administrators run these from the console and see every grade and transcript.
Colleagues your people can trust
- A settled identity, the same on Monday as on Friday.
- Their own record, and nobody else’s, behind every answer.
- Arithmetic computed, sources named, changes approved.
- A team that hands a question to the right colleague.