FRANK Raises €2.9M for an Insurer-Portal AI Operator

Portuguese insurtech FRANK has raised €2.9 million to train an AI operator that logs into insurer portals and types the way a broker would. Armilar and Start Ventures led the round. The Lisbon firm was founded in December 2024 by António Castro, Ricardo Vallejo and Alexandre Silva.

Most of the money will go into continuous training of that operator, the company said, before a planned push outside Portugal in 2027. Brokers get hours back. The websites they already use stay as they are.

The Operator That Types Into Insurer Portals

FRANK sells an AI operator for insurance advisors that copies human work across brokerage systems. A quote for home cover, a claim chase, or a client-file update still means email, WhatsApp, PDFs and a stack of insurer sites that do not share a process. The operator sits on those channels, reads the papers, names the job, and tells the broker what is missing. Then it signs into the right portal, enters the data, and sends back a quote or a claim status.

Castro, the co-founder and chief executive, put the time drain in blunt terms.

Most brokers spend 80% of their time on back-office tasks instead of being with clients.

António Castro, co-founder and CEO of FRANK

He wants that time moved into risk advice, client work and basic insurance literacy. The product is built as three specialised agents already running in live agencies: an ERP and CRM copilot, a document agent, and a portal agent that does the typing on insurer sites. A higher-level controller decides when to answer and when to hand a file back to a person.

WHAT THE OPERATOR DOES IN SEQUENCE

• Inbox first: It watches email and WhatsApp, the channels brokers already use with clients.
• Paper next: It reads the attached documents and names the process, from a new quote to a claim update.
• Gaps flagged: It tells the broker which fields are still empty before anyone opens a portal.
• Portal login: It signs into the relevant insurer or brokerage system with the agency’s own access.
• Result back: It returns a commercial proposal or a claim status instead of a half-filled screen.

When business rules and data are written down, the company said a new workflow can go live in less than a day. The slow work is pulling logic out of manuals, contracts and people’s heads. FRANK calls the rollout method Spiral Kaizen: a pilot of five to 15 agencies, then the tool’s own usage data picks the next job to automate.

Portugal’s 10,024 Intermediaries Still Carry Non-Life

That design fits the way insurance is still sold in Portugal. The Autoridade de Supervisão de Seguros e Fundos de Pensões, in its first dedicated mediation report, counted 10,024 insurance intermediaries at year-end 2025. Traditional mediators took 75.4% of non-life sales. Banks took 72.5% of life. Direct counters, phone and web were 3.2% of life and 7.4% of non-life. Mediators handled more than 92% of premiums issued in 2025.

WHERE PORTUGAL’S POLICIES ARE SOLD

The register is almost all agents. ASF counted 9,936 insurance agents, 66 brokers, six accessory intermediaries and 16 reinsurance intermediaries. Pay for the activity reached €1,475 million in 2025, up 9.7% on the year before, with €390 million in life and €1,085 million in non-life. A firm that can take even a slice of that admin load is not selling a toy for a handful of digital brokers. It is selling time to the channel that still moves most of the non-life book.

How Fast Brokers Put the Bot to Work

FRANK began selling the platform less than 12 months ago. It now says it is used by more than 350 insurance agencies and brokerages, representing more than 1,500 users across Portugal. The company puts monthly back-office volume at 70,000 tasks. Agencies using it report saving about ten hours of admin per broker per week.

Those figures have moved quickly. At Portugal Fintech’s FinAI Playbook on 15 June 2026, FRANK said it had already passed 30,000 automated tasks a month and cut more than 70% of the manual time on the processes it had targeted. The 70% cut is a measure of those targeted jobs, not of a whole broker week. Castro’s 80% figure is his estimate of how much of the day still goes on back office before the tool is applied.

FROM LAB TO 70,000 TASKS

1. December 2024: Castro, Vallejo and Silva found FRANK in Lisbon.
2. June 2025: The company is still in the lab with a small engineering group, showing the product to sole traders and agencies of about 15 people, with a public launch planned for September and sales due at the end of that year.
3. 15 June 2026: Three agents run in production; monthly automated tasks pass 30,000.
4. 1 October 2026: Armilar and Start Ventures lead a €2.9 million round; monthly tasks are 70,000, with more than 350 agencies and more than 1,500 users.
5. 2027: FRANK plans to start selling outside Portugal after it has tightened its hold at home.

Carlos Siqueira, the commercial lead, said several client firms greet the operator and thank it as if it were a colleague. Usage, he said, shows up every day of the week, in office hours and after them. That after-hours pattern is the tell. The bot is being used as night staff on portals that were built for a person at a desk, not as a dashboard brokers open once a morning and close.

Capital From a €600 Million Lisbon Firm

Armilar is the larger name on the ticket. In a 5 November 2025 note on the first close of its fourth flagship fund, the Lisbon firm described itself as managing over €600 million in assets, with more than 80 companies backed since 2000, 22 exits, three IPOs and three unicorns. That fund’s first close was over €120 million, with the European Investment Fund among the backers, and a target of about 20 companies over the fund’s life. Inês de Macedo Santos, a principal at Armilar, said FRANK was built to pull mediators out of manual back-office work so they can protect clients more fully.

Start Ventures went in even earlier. Nuno Marques Afonso, an associate at the firm, said it invested before FRANK had a single client, then watched one of the most conservative trades it knows take the product up at a pace it rarely sees in insurtech. The company now has a team of 20, with careers that run through Lovys, JPMorgan, McKinsey, SAP, Accenture, FiveAI, Indie Campers and Bolt. Castro himself was listed as chief financial officer of digital insurer Lovys, the Portuguese subscription insurer founded by João Cardoso. The new company is a turn from selling policies on a phone to selling labour inside the agencies that still sit between those policies and the public.

Brokers Average 54, and Banks Own Life

The people who would use that labour are getting older. ASF put the average age of individual insurance mediators at 54.1 years in 2025, a figure that has drifted up. The regulator recorded 702 access-to-the-activity training sessions in 2025, down from 768 in 2024, though enrolments rose from 5,119 to 5,446. Gabriel Bernardino, chair of ASF, said the task in a fast technology shift is to combine tech with proximity, and that the firms which last will be technical in method and human in the advice they give.

Life business is already a bank product in Portugal. Non-life is still a mediator product. FRANK is not trying to replace the person who sits with a household and talks through a policy. It is trying to keep that person in the room by taking the portal-hopping that has made the job look like data entry. For a 54-year-old agent running a shop of one to 15 people, ten hours a week is not a productivity slide. It is the difference between seeing another client and spending the afternoon re-keying a claim.

Banks do not need that help on life, where they already hold 72.5% of sales. The independent channel does, because its share of non-life still depends on people who work across many carrier sites. If those people leave and are not replaced, the 75.4% figure will not stay with small agencies. It will drift toward whoever can still staff a back office, including larger brokers and the same banks.

The Portal Layer the Industry Never Replaced

The second effect of the round is quieter than the time saving. FRANK is not waiting for every Portuguese insurer to publish a clean quoting API. It is training a model to use the pages those insurers already run. That is the same workaround now sold to US agencies, where analysis of AI web agents that log into carrier portals starts from a blunt count: 98% of US property and casualty carriers still have no quoting API, so the bot uses the website. Portugal is a different market, but the constraint is the same shape. Comparative raters that need an API only ever reach the carriers that built one. A portal operator can, in principle, reach any carrier that still has a login box.

The cost of that reach is maintenance. Portals change. Field names move. A model that is “hard to copy,” in the company’s own phrase for where the new capital is going, is a model that has seen thousands of those screens. Each extra agency, each extra insurer site, each extra language is training data. The round is a bet that this library, not a set of one-off RPA scripts, is the thing rivals cannot cheaply rebuild. It is also a bet that carriers will keep the portals. If they opened proper interfaces tomorrow, a lot of the operator’s edge would shrink to document reading and inbox triage, which is a crowded job.

So the product helps the broker and, in the same motion, reduces the daily pain that would have forced those carriers to modernise the login layer. Agencies get the hours. Insurers keep their extranets. The messy middle becomes FRANK’s training set.

2027 Starts as Homework in Lisbon

The company said the new capital will strengthen support for brokerage clients, fund research and development, and prepare an international launch. Most of it, it said, will pay for continuous training of the operator so the system becomes scalable, hard to replicate, and hard for a mediator to drop. 2027 is the date given for sales outside Portugal. The work between now and then is Portugal: more agencies on the same portals, more of the 10,024-strong register, more of the 70,000-task load turned into labelled examples.

Other firms are already selling computer-use agents that drive insurer sites as a person would, so FRANK does not have the idea to itself. What it does have is a live book of Portuguese agencies, a 20-person team that already knows this trade, and a year of production data on three agents. Exporting that means teaching the same operator a new country’s portals, rules and languages, which is exactly why the cheque is weighted toward training rather than a bigger sales floor.

Siqueira’s clients are already saying good morning to the bot on weeknights. The next test is whether that night clerk can learn a foreign extranet as cheaply as it learned a Portuguese one, while the login boxes it was built to fill remain right where they are.