The AI assembles the file. Approving, returning or rejecting is a person’s call.
Every request arriving from a clinic enters a queue with coverage checked, the deductible calculated, exclusions and pre-existing conditions reviewed, medical necessity contrasted and the amount placed against the fee schedule agreed with that provider. The proposal arrives written and with its reason; the medical auditor decides, and the decision lands in the audit log with their name and the time.
The validity of each authorization is set by the agreement with that provider.
The AI checks coverage, deductible, exclusions, pre-existing conditions, necessity and the fee schedule.
Approving, returning or rejecting is a person’s call. Always.
0
lines in one system
major medical · motor · life · property · surety
0
checks before it proposes
coverage · deductible · exclusions · pre-existing · necessity · fee schedule
0
possible outcomes for a request
approve · return with an observation · reject
0
authorizations the AI approves alone
the medical auditor signs the decision
UNDER THE HOOD
It does not guess whether the case is covered. It checks it against the contract and the fee schedule.
Three requests from the same day in the authorization queue. In all three the AI checks the same things in the same order, writes down what it found and proposes an outcome. In all three the final call belongs to the medical auditor, and the reason travels with it.
Sample lead
Raw lead comes in
Request SOL-2291 from provider Clínica Almendra (preferred network, active agreement). Insured L.P., group major-medical policy on the Integral plan, 26 months of tenure. Signed medical report, two-day stay with a minor procedure. Budgeted amount: $1,860.
Coverage
policy active, premium current, the insured appears on the group roster and the procedure is covered by the plan
Deductible and co-pay
the annual deductible is already met for the period; the plan’s 10 % co-pay on pharmacy, calculated line by line
Exclusions and pre-existing
nothing declared matches the plan’s exclusions; tenure exceeds the waiting period in the contract
Necessity
the diagnosis in the report supports the procedure requested; the studies behind it come attached
Fee schedule
all seven lines sit inside the schedule agreed with Almendra for that plan; the budgeted total stays under the agreement’s cap
Score
0/100
Verdict from Claude
The AI proposes approving at the schedule amount with ten business days of validity. The medical auditor reviews the file on one screen, presses approve and the authorization goes out with its number. Who approved, when and with what reason is written down.
WHAT THE AI DOES HERE
Five concrete functions inside the insurer’s operation
None of them approves, rejects or pays. They assemble the file, cross the contract with the fee schedule, calculate the reserve, draft the adjustment and project the loss ratio; the decision that commits money belongs to a person with a name and a role.
seguros.armarExpediente()
The request arrives from the clinic and comes out ready to decide
It takes the request that came in through the clinic portal, by email or as a PDF, matches it to the right policy and assembles the file the auditor needs on a single screen: active coverage, deductible and co-pay calculated, plan exclusions, declared pre-existing conditions, the medical report and the budgeted amount. Whatever is missing is flagged; it is never filled in by assumption.
› input
SOL-2291 · provider Clínica Almendra · insured L.P. · group major medical, Integral plan · signed medical report
› claude →
policy identified: active, premium current, insured on the group roster
deductible for the period and plan co-pay, calculated per line
contract exclusions and waiting periods, contrasted with what was declared
missing documents flagged by name, with no invented content
the full file on one screen, with a link to each original document
LIVE SYSTEM · Seguros Latitud
This is what it looks like inside an AI-operated system — and how it is governed.
Seguros Latitud receives the authorization requests from its clinic network every morning. Its technical desk is a file: the AI assembles and annotates it, and approving, returning or rejecting is decided by a person, with a written reason.
System modules
- 01Authorizations
- 02Claims
- 03Adjustments
- 04Loss ratio
- 05Network and fee schedules
- 06Underwriting
- 07Policies and terms
- 08Premium collection
- 09Decision files
Boosty Standard for Operating with AISimulated AI · demo data
The AI assembles the file. A person presses the button.
Every request coming in from a clinic reaches this queue with six checks done and a written proposal. Pick a request, read what the AI found and decide: the decision is recorded with who made it, at what time and for what reason.
Request queue
Pick a request and decide
what the AI had in view
group contract · plan roster · signed medical report · Almendra schedule v7 · the insured’s history
Assembled file
SOL-229110:42Clínica Almendra · Group major medical · Integral
L.P. · 26 months of tenure · Two-day stay with a minor procedure
- Coveragecheckedpolicy active, premium current and the insured is on the group roster
- Deductible and co-paycheckeddeductible for the period already met; 10 % co-pay on pharmacy, calculated per line
- Plan exclusionscheckednothing declared matches the contract exclusions
- Pre-existing conditionscheckedtenure exceeds the waiting period set by the contract
- Medical necessitycheckedthe diagnosis in the report supports the procedure; the studies come attached
- Amount against the schedulecheckedthe budgeted total stays under the agreement cap
Budgeted amount against the fee schedule
Budgeted by the provider
$1,860.00
Schedule agreed with that provider
$1,860.00
all seven lines are coded in the agreement with that provider
What the AI proposes
Approve at the schedule amountCase covered by the plan, with no applicable exclusions and an amount inside the schedule agreed with Almendra. Proposal: approve with ten business days of validity.
It is a proposal. It changes no state and issues nothing.
The medical auditor decides
Audit entry
No decision yet: the request stays in the queue and the response clock for the provider keeps running.
No request changes state without a person pressing the button. That is the rule of the system, not a preference.
IN THE SYSTEM · AUTHORIZATIONS DESK
The AI assembles the file. Deciding takes a name and a reason.
The queue with each response deadline, the file with the AI reading annotated beside the paragraph that supports it, and the decision: approve, return or reject, with a written reason without which the button stays disabled.
Real screenshot of the system · demonstration data
- 1Queue by line with the response deadline to the provider.
- 2Coverage, deductible, exclusions and fee schedule, beside the document.
- 3The AI proposes · the medical auditor decides.
Five lines, one portfolio, and a matrix you can read
Each line has its own underwriting rules and its own way of measuring risk, but the portfolio is one. The AI pre-assesses every policy under review and leaves the matrix explained: which factor moves it, in which direction and with what weight. Issuing, renewing or not renewing is signed by underwriting.
Company lines
Switch line
Line portfolio · HCM
3,412
active policies
186
renewals within 30 days
24
under technical review
By segment
| Segment | Policies | Risk |
|---|---|---|
| Corporate group | 2,108 | Medium |
| Small-business group | 844 | High |
| Individual and family | 460 | Medium |
Risk matrix of the policy under review
POL-HCM-1174Small-business group of 38 insured · Esencial plan · third renewal
Expected severity ↑ · Expected frequency →
What moves the matrix
- Use concentrated in two providersseven of every ten authorizations in the group come from the same two centers, both on a high schedule
- Average age of the groupthe group’s age pyramid rose over the last two renewals as no new staff joined
- No long staysseverity stays contained: no case in the period exceeded 72 hours of stay
- Per-event deductible in forcethe plan deductible filters minor events before they reach the company
Proposal waiting for a signature
Renew with a review of the provider panel and the per-event deductible sustained. Written alternative: renew unchanged and review at six months. Underwriting signs.
The matrix neither issues nor cancels. Underwriting signs the issuance, the renewal or the non-renewal, and the reason stays with the policy.
IN THE SYSTEM · UNDERWRITING
The risk matrix is read, not guessed.
Every renewal and quote arrives with its explained risk matrix and the AI proposal. The underwriter decides, and whatever exceeds their authority goes to the technical director.
Real screenshot of the system · demonstration data
A number on its own says nothing. An explained drift does.
The company compares itself against its own history, not against the market. The AI follows the series by line and by provider and raises a flag when one departs from what it had been doing, with the month it started and the item that grew. What to do with the portfolio is the chief underwriting officer’s call.
Pick a series
| Series | Loss ratio | Frequency | Severity | Against its history |
|---|---|---|---|---|
| 74 % | rising | flat | above its band | |
| 62 % | flat | falling | inside its band | |
| 31 % | flat | flat | inside its band | |
| 43 % | falling | one event | below its band |
Last eight months · Major medical
64.1 % · the series’ own averageNo market benchmarks: the reference is the company itself.
The drift, explained
74 % · above its band- What changed
- Short outpatient events grow, not inpatient stays: frequency rises while average severity stays where it was.
- Where
- Concentrated in the small-business group segment and, inside it, in two preferred-network providers.
- Since when
- The series leaves its band from June, three consecutive months above it.
Year-end close
With the frequency of the last three months sustained, the close projects into a range; the assumption is that the provider panel does not change.
It is projected as a range, never as a single figure, and with the assumption written beside it.
The AI changes no rate and removes no provider from the network. It leaves the series, the explanation and the assumption; deciding belongs to the chief underwriting officer with network management.
IN THE SYSTEM · LOSS RATIO
The deviation arrives with its explanation.
By line, plan, provider and portfolio, with frequency and severity. The AI projection says what drives it, and no figure is a market figure: everything comes from the company’s own portfolio.
Real screenshot of the system · demonstration data
SAME ENGINE, A DIFFERENT LINE
What is insured changes. The circuit of the decision does not.
Underwriting, authorizing, reserving, auditing and settling exist in all five lines. What changes is who decides, which file is needed and which table it gets compared against. The system is the same; the rules of each line are not.
Signals specific to the industry
The same item, five agreed prices
A fee schedule is the price the company agreed with each provider for each item. Pick an item and compare: the reference is the network’s own average, not a market table. Drift is flagged with its explanation; renegotiating an agreement is a conversation for network management.
Fee schedule item
Switch item
Inpatient day · standard room
Network average: $180.50 · per day| Provider | Agreed price | Against the average | Response |
|---|---|---|---|
| Clínica Almendrapreferred network · agreement v7 | $180.00 | −0.50 | same day |
| C. Quirúrgico Sotaventosurgical network · agreement v4 | $205.00 | +24.50 | 1 business day |
| Policlínico Barloventogeneral network · agreement v3 | $172.00 | −8.50 | 2 business days |
| Centro Mediterráneogeneral network · agreement v5 | $165.00 | −15.50 | same day |
| Unidad Litoraloutpatient network · agreement v2 | not in this agreement | — | 3 business days |
average time to answer a request, per the company’s own history
What the AI writes down
Sotavento sits above the network average on this item, consistent with its profile: a surgical center with short post-operative stays. Unidad Litoral has no inpatient care in its agreement, so the line does not exist: it is not compared against zero.
The AI changes no schedule and removes nobody from the network: it flags the drift, explains it and puts it on network management’s agenda. An agreement is signed by two parties.
IN THE SYSTEM · NETWORK AND FEE SCHEDULES
Every provider against the network average.
The fee schedule agreed with each clinic compared line by line, and the price deviations the AI finds in billing, with the case that shows them.
Real screenshot of the system · demonstration data
What commits money or health is signed by a person
In an insurer, governance is not a document: it is four things you can see on the screen. Who decides each act, how a rejection is explained to whoever receives it, who publishes the risk matrix and what stays on record when the regulator asks.
Four pieces
Pick a piece
What the AI never does
- Approve, reject or return a request on its own.
- Issue, renew or cancel a policy.
- Change a rate, a fee schedule or a reserve already signed.
- Order a payment to a provider or to an insured.
- Declare fraud or remove a provider from the network.
Mandatory human decision
Every act that commits health or money has an owner with a name and a role. The AI goes as far as the written proposal and stops there; the system lets no flow move on without the signature.
| Assembles the request file and proposes approving, returning or rejecting with a reason. | Medical audit |
| Pre-assesses the policy risk and leaves the matrix explained with its factors. | Underwriting |
| Proposes the claim reserve with a breakdown of what makes it up. | Claims analysis |
| Drafts the adjustment to the provider’s billing line by line. | Claims with network management |
| Flags the overdue receipt and prepares the notice to the insured and the broker. | Premium collections |
The boundary
If a flow tries to change the state of a request, a policy, a reserve or a payment without a signature, it stops and the attempt lands in the audit log with its timestamp.
IN THE SYSTEM · DECISION FILES
What was decided, ready for the regulator.
Who decided, when, with what reason, what the AI read and whether the person departed from its proposal. With filters and a printable version.
Real screenshot of the system · demonstration data
THE SAME FACT · TWO COMPANIES
One authorization: the clinic requests it, the insurer decides it.
It is one fact with two numbers. The clinic owns the clinical report behind it; the insurer owns coverage and the decision. Neither rewrites the other’s part, and both see the same timestamp.
K-9018 · A-9050Clínicas Latitud · requests it
K-9018 · A-9050Seguros Latitud · decides it
Real screenshot of the system · demonstration data
WITH WHAT YOU ALREADY USE
The system connects to the company’s reality
Almost no insurer starts from zero: there is a core holding policies and accounting, and there is a clinic network that works through portals, email and PDFs. The system leans on both: it takes from the core what the core does well and puts the queue where decisions happen on top.
Claude · Anthropic
EngineAssembles the file, contrasts necessity against the fee schedule, calculates the reserve, drafts the adjustment and explains the drift. It does not approve, reject or pay.
Insurance core
Policies, receipts and accounting can stay where they are: they integrate by API and the system writes back what a person decided, with its reference.
Provider portal
Clinics request the authorization, check its status, see the adjustment with its reason and reply. All of it keyed by the request number.
WhatsApp Business
A note to the insured when their authorization comes out or when a document is missing, and a reminder for a receipt about to expire. The conversation stays attached to the case.
Supabase · Postgres
Data lives in your database, with per-line and per-role permissions at row level: the motor appraiser does not see the health file.
Make · n8n
An approved authorization notifies the provider and starts its validity; an overdue receipt enters the collections queue. No policy is ever cancelled on its own.
Frequently asked questions about systems for insurers
No. The AI assembles the file, checks coverage, deductible, exclusions, pre-existing conditions and medical necessity, compares against the provider’s fee schedule and proposes an outcome with its reason drafted. Approving, returning with an observation or rejecting is pressed by a person with a name and a role, and that decision lands in the audit log with the time and the reason. It is a rule written into the system, not good intentions: no flow can change the state of a request without a human signature.
With the reason, the contract clause it rests on and what would be needed to review it, written in plain language rather than internal codes. The provider sees it in their portal and the insured receives it through their channel. What the AI saw before proposing and what the person decided are both recorded: if someone appeals tomorrow, the explanation is reconstructed in full instead of depending on the memory of whoever signed.
In your own database, with per-line and per-role permissions at row level: medical audit sees the clinical file for its line, underwriting sees the health declaration relevant to it, collections sees the receipt and not the diagnosis. Access is recorded — who opened what and why — and the AI receives the fragment needed for the task, not the whole file. The applicable legal framework depends on the country the company operates in and is reviewed when the scope is defined.
Not necessarily, and almost never on day one. The usual path is to leave in the core what it does well — issuance, receipts, accounting — and integrate it by API, then put on top what today lives in emails and spreadsheets: the authorization queue, the fee schedule per provider, the file behind every decision and the loss ratio explained. What stays, what gets integrated and what gets absorbed comes out of the assessment, and the transition plan is written before anything is touched.
It is treated as a pattern to be documented, not as an automatic accusation. The system flags inconsistencies between documents, short repeats of a procedure already paid, amounts outside the agreement and series that depart from that provider’s own history. Every signal arrives with the file behind it and goes to audit, where a person decides whether to ask for clarification, renegotiate the agreement or pay. No request is rejected on a suspicion without someone reviewing it.
The system keeps what those reports need: every decision with its author, its reason, its date and what the AI had in view when it proposed, plus the trail of changes to policies, reserves and payments. The specific formats and deadlines are set by each country’s regulator and they change: that is why they are configured at implementation, against the rules in force where the company operates, instead of being promised out of the box.
It is defined in the assessment. The model has three parts: setup, a monthly fee and a variable cost per processed volume. The scope — which line comes first, how many providers, whether the core is integrated or absorbed — comes from what we see in your operation. Book 30 minutes and we give you the range in writing.
A WORD FROM THE FOUNDER
“In an insurer the question is never whether the AI can decide. It is how long a person takes to have in front of them everything they need to decide well.”
When you sit down with a medical auditor, the bottleneck shows up in five minutes: the request is in an email, the policy is in the core, that provider’s fee schedule is in a spreadsheet, the insured’s history is in another system and the medical report is a PDF you have to open separately. Deciding well means assembling all of that by hand, request after request, while the clinic calls to ask.
That is exactly what a system with AI does well: read, cross-check, calculate and leave the file assembled on a single screen, with what is missing flagged and a proposal written with its reason. And there it stops. Approving, rejecting or returning touches health and touches money: a person decides, and the decision stays explained for the insured, for the provider and for whoever audits it later.
If you want to see your own circuit — from the request that comes in to the authorization that goes out, and from the provider’s invoice to the adjustment — book 30 minutes with me. We take a real case of yours and walk it end to end.
Gabriel Montiel
CEO · Boosty Digital
START
How many clinic requests do you receive per day?
Book a 30-minute assessment. We walk the full circuit — request, file, decision, authorization, provider invoice and adjustment — and tell you what we would build first.