AI practice & rule checks for insurance agents
New agents rehearse real customer conversations with an AI customer. Every answer is scored and checked against your own rules, so managers can see who’s ready, and why.
Now opening four-week pilots with US agencies, IMOs and FMOs. See what runs today
Answer an AI customer three times. Risky lines are stopped before the customer hears them. Then you get a scored review and make the call: approve the agent, or assign drills. One minute, no sign-up.
Daniel 44 · comparing two term life quotes
Turn 1 of 3
Practice round · Daniel is an AI customer
Why are you recommending this plan? Is it really the best fit for me?
INCE is watching your draft
Manager review · prepared by the agents
–/ 4 criteria
Your call
The agents prepare the evidence. A person makes the call.
Team readiness · sample team · planned feature
| Agent | Conversation | Rubric | Rule check | Status |
|---|---|---|---|---|
| You, this round | ||||
| J. Alvarez | Medicare · first call | 4/4 | Clean | Ready for live calls |
| P. Shah | Life · suitability | 2/4 | 2 stopped | Needs another round |
| S. Kim | Annuities · best interest | 3/4 | 1 flagged | 1 drill left |
One workflow, 5–10 agents, four weeks. We build the scenarios, scoring and rules from your approved material, and measure against your own baseline.
30-minute call. No customer data needed to start.
Simulation: a scripted customer and a simplified INCE check. Nothing you type leaves your browser. Rules are examples, not legal advice. Review and coaching agents are design prototypes.3
Our work with financial groups in Asia
The problem
Managers answer the same questions over and over, coach from memory, and find out who’s struggling only after the numbers drop.
01
Plenty of candidates, but many skip the first meeting. Volume isn’t the bottleneck.
02
After the licensing exam, training depends on one manager’s memory and spare time.
03
Managers can’t see how conversations go. Notes are scattered across spreadsheets and chat apps.
04
Results lag, the warm market runs dry, and coaching comes too late.
05
Agents leave before their hiring and training costs are paid back.
What agency owners told us
I want to know in advance who will stay.
They asked for: aptitude insights and earlier coaching signals.
Record and analyze calls without manual entry.
They asked for: call transcription, analysis and automatic CRM entry.
Keep evidence of how the sales process was explained.
They asked for: rule checks and an evidence log.
From interviews and operating data at our 23 paying agencies in South Korea; quotes translated from Korean.2
Product
One loop, built on your own approved material. AI handles the repetition. Your managers keep the judgment.
One customer situation at a time, built from the products and rules you’ve approved.
Every answer is scored on a fixed rubric, with the evidence attached, so a manager can judge in minutes.
The weak topic, a corrected explanation and another round, until it sticks.
For agents and prospective agents
Recruiting platform in South Korea. Free for candidates.
For agency managers and admins
For agencies in South Korea and enterprise teams.
For agency home offices and insurer compliance teams
Enterprise: scoped PoC, then a separately agreed license.
Live Pilot Proof of concept Planned Status as of October 2026.3
How the checks work
INCE, our rule-check engine, keeps reading and judging apart. A language model reads what was said. Your rules, encoded for a logic solver, decide whether it holds. Anything uncertain goes to a person.
Input
Policy terms, internal rules and what was said.
Stage 1 · Understand
A language model turns every sentence into explicit claims.
Stage 2 · Rule check
The Z3 logic solver tests each claim against your encoded (SMT) rules.
Stage 3 · Block
Exaggeration, misleading claims and banned phrases are stopped. Unsure cases go to a person.
Output
The verdict, the clause it cites and the rule version, logged together.
Each verdict carries its cited clause and rule version, ready for internal review and audit prep.
Rules can be updated without retraining the language model. Changes still get expert review and regression tests.
2 granted 7 pending 1 PCT
A patented engine, with domain rules and training data on top.4
What it catches depends on the rules, inputs and test set, so pilot targets are agreed with each customer and a person stays in the loop. Watch a simplified version stop a line in the Medicare demo.
Pilot
One workflow, measured against your own baseline. If it misses the bar we agreed on, we narrow the scope or stop.
What’s live, what’s still in pilot or proof of concept, and what we build for your team. Status as of October 2026.3
| Part | Status today | In a US pilot |
|---|---|---|
| AI customer role-play | Pilot Product UI in Korean today | English scenarios built from your scripts, product guides and rules |
| Scoring and manager review | Design prototype | A fixed rubric agreed up front, with scores reviewed by people |
| INCE rule checks | Proof of concept | Rules encoded from your material; anything uncertain goes to a person |
| Team readiness dashboard | Planned | Not live yet |
| Agent recruiting platform | Live South Korea | Not offered in the US |
Move the sliders to match your team. A pilot measures the real number.
Manager time spent onboarding
1,200 hours / year
About $72,000 a year, or 30 working weeks of a manager.
Freed up each year at 20% less manager time, an example pilot target
$14,400
Manager time only. It leaves out time to first sale and agents who leave early, which a pilot also measures. Your inputs, simple math. Not a forecast or a guarantee.
| Term | What we agree |
|---|---|
| Scope and owner | One workflow, 5–10 agents, one training manager and a named budget sponsor. |
| Baseline | Time today’s task and score it on a fixed, human-reviewed rubric. |
| Evaluation | New scenarios of equal difficulty at the end. We record completion, rubric scores and manager review time. |
| Decision criteria | Set before we start. For example: 20% less manager time, an 80% rubric score and no critical rule errors in the tested sample. |
| Your side | A sponsor, access for the evaluation, agreed terms and a date for the paid-license decision. |
| Stop or narrow | No sponsor, unusable data or a missed bar means we re-scope or stop. |
Four weeks, then a decision. Not an open-ended trial.5
We’ll come back with a draft pilot plan: owner, baseline, how we test, and what counts as success.
or email biz@intalk.im
We use what you share only to reply. Privacy notice
Traction
Since 2022, we’ve run agent recruiting, onboarding and sales training for insurance agencies in South Korea. That’s where our product, our data and our next markets come from.
$330K+
Total company revenue, January–August 20261 KRW 500M+, incl. services & projects · not ARR
23
Agencies paying for agent recruiting2 September 2026
80
Agencies onboarded2 September 2026
Company-reported, unaudited. Sources
Proofs of concept (PoCs), pilots and program selections in South Korea and Vietnam, not commercial contracts. Names show project context, not endorsement.
Operating base
Paying agencies, product and data. Seoul HQ.
Training PoC
Agent training with Shinhan Life Vietnam.
Southeast Asia hub
Partners and PoC opportunities across Southeast Asia.
Now entering
Agencies with a named training or QA owner.
Insurance first. Then other sales teams where every explanation has to be accurate, such as finance and pharma.
| Path | Buyer | Commercial basis |
|---|---|---|
| Agent recruiting | Independent agencies in South Korea | Candidate-matching packages |
| Annual license | Agency seats in South Korea | Per seat, per year, prepaid |
| Enterprise | Insurers and sales teams | Scoped pilot, then a separately agreed license |
The 12-months-free launch offer is separate from paid enterprise work.
Every number here is company-reported. The investor overview shows the evidence behind each one, how we define unit economics, and what we’re testing next.
The button opens your email app. No email app? Write to biz@intalk.im.
Team
Founder & CEO
AI engineering
Strategy
From the founder
“Intalk isn’t here to replace salespeople. We turn managers’ experience into a system, so the people in the field can help customers more accurately and more responsibly.”
Ian Yoo, Founder & CEOMeet us
Agency leaders, training and QA teams, investors: grab 30 minutes with us in person.
Oct 19–23, 2026 Washington, D.C.
With the Global Launchpad program of Korea’s four science and technology institutes, including KAIST.
Oct 27–29, 2026 Singapore
Singapore Week of Innovation and Technology.
Oct 30, 2026 Singapore
Selected for the Singapore Deep Tech PoC program.
Global launch
Intalk AISO, our AI sales operating system, opens beyond pilots: practice, review and coaching for more sales teams. Apply now and we’ll suggest the right starting point for yours.
12 months free
after launch
Planned launch November 30, 2026
Applications are reviewed; applying doesn’t guarantee approval, seats or immediate access. Need it built on your own scripts and rules? That’s a pilot.
A few details about your team and role. We reply with a clear next step.
Apply for 12 months freePrefer a conversation first? biz@intalk.im
We use what you share only to review your application. Privacy notice
FAQ
Something else? Email biz@intalk.im.
Intalk is an AI sales operating system for insurance distribution, built by Intalk Inc. in Seoul, South Korea, and founded in November 2022. It gets new agents ready before the customer conversation: agents rehearse real customer situations with AI role-play, managers review scored answers with the evidence attached, and the INCE engine checks explanations against the organization’s approved material and rules.
Insurance agencies, IMOs and FMOs, telesales teams and insurers’ agency channels that train many new agents and answer for what those agents tell customers. A pilot is usually led by an agency owner or sales-ops lead, with a training, QA or compliance owner.
INCE is Intalk’s rule-check engine. A language model turns each sentence of a sales conversation or an AI-generated answer into explicit claims; encoded rules and the Z3 SMT solver then test those claims for violations such as exaggeration, misrepresentation or banned phrases. Each verdict is logged with the cited clause and the rule version, and low-confidence cases go to a person. INCE is built and in proof-of-concept (PoC) testing.
Three things. Scenarios, scoring and rules come from your own approved material, not a generic script. Answers are checked by a rule engine that cites the clause it relied on, not just rated by a language model. And every pilot is measured against your own baseline, with the success bar agreed up front.
The product UI is in Korean today. For US pilots we build the English workflow and scenarios from your material, and agree up front what “ready” means and how it’s measured.
One workflow, 5–10 agents, a training manager who owns the pilot, a named budget sponsor, and the approved material the scenarios should follow: scripts, product guides and compliance checklists.
Data permissions are agreed in writing before delivery. Practice scenarios can be built from synthetic or de-identified examples, so a pilot does not need access to live customer records or protected health information (PHI).
We encode the rules from your material, such as required disclosures (a TPMO disclaimer on Medicare calls, for example), claims agents must avoid and suitability steps, and INCE checks practice answers and AI-generated answers against them. Coverage depends on the rules and the evaluation set, so uncertain cases go to a person, and “no critical rule errors in the evaluated sample” is one of the decision criteria.
A general AI tool can play the customer. It doesn’t give you the training system around the conversation: scenarios built from your approved scripts and product guides, answers scored on a fixed rubric with evidence a manager can review, and rule checks where encoded rules and a solver, not the language model, decide whether an explanation holds, with the cited clause logged. Anything uncertain goes to a person.
No. Intalk turns your managers’ experience into practice and review, so agents get more repetitions and managers stop repeating themselves. People make the final call on anything a customer will hear.
Use of Intalk AISO for 12 months from the general launch, planned for November 30, 2026. Which features, how many seats and any services outside the offer are confirmed with your team when we review your application. The product UI is in Korean today, so ask us about English when you apply. It is separate from paid pilots, which we build in English on your own scripts and rules.
No. It’s a simulation that runs in your browser: the customer follows a branching script, and the rule check is a simplified, pattern-based version of INCE. It shows how the workflow fits together (practice, rule check, review, the manager’s call), not how accurate the real engine is. Nothing you type is sent to us. The product UI is in Korean today; the Product section screen is a design prototype with sample data, shown in English.
Scope, price and start date are agreed before delivery, together with a dated decision on a paid license. If the pilot clears the agreed bar, it converts to an annual license and expands team by team. If it doesn’t, we narrow the scope or stop.
Intalk is headquartered in Seoul, South Korea, where its paying agency customers are. It runs an agent-training PoC with Shinhan Life Vietnam, works with partners across Southeast Asia from Singapore, and is opening pilots with US agencies, IMOs and FMOs.
Contact
Insurance & sales teams
A 30-minute call about one workflow. Then a draft plan with a baseline and a success bar.
Investors & partners
Investor materials, with the evidence behind every number and what we’re testing next.
Press
Interviews, fact checks and logos. We answer in English and Korean.
