Chief Product Officer / Europe
Three product choices that each gave something up
- Annual recurring revenue
- €64m
- Calvessin Security, June 2026
- Paying small organisations
- 33,800
- Portugal, Spain and Ireland; five to 250 staff each
- Product, design and research colleagues
- 92
- Engineering and the alert desk report elsewhere
Decisions & consequences
01A flat Micro plan that reduced revenue before it added any
In early 2023 our trial data showed a pattern we had been explaining away. Firms with fewer than ten staff made up 46 per cent of trials but only 21 per cent of new subscriptions. Their trial-to-paid rate was 11 per cent, against 27 per cent for larger firms. Interviews pointed to the price page: Standard then required a minimum of ten users, so a four-person accountancy practice paid €65 a month for six empty seats.
I proposed a flat Micro plan at €39 a month for up to eight users, with Standard features but without the 24-hour alert desk. Our chief financial officer objected, reasonably. About 6,200 existing customers had fewer than ten staff, and many could move down at renewal. The partner team warned that a lower price would cut partner earnings on exactly the customers who make the most support calls. Two large partners threatened to stop recommending us to micro firms.
We accepted the revenue loss rather than hide Micro from existing customers; a small firm should not have to complain to find it. To protect partners, Micro carried a fixed €7 monthly partner fee in place of a percentage margin. We tested it in Portugal for three months before Spain and Ireland.
By December 2024, 7,900 new organisations had joined on Micro, and the conversion rate for firms under ten staff had risen to 19 per cent. Some 2,350 existing customers moved down, reducing annualised revenue by about €0.94 million. Net new annual recurring revenue from the plan reached €3.7 million. Thirty-eight partners stopped selling to micro firms, and one of the two large partners left us.
Every pricing proposal now has to show its effect on existing customers and on partner earnings per support hour. What I would do differently is the support model. Micro customers raised 1.6 times as many tickets per user as Standard customers, and I had assumed parity. Guided self-service should have preceded launch, not followed it by nine months.
Supplier bank-detail checks that warn rather than block
Invoice-redirection fraud was the costliest incident our customers reported. In 2023 our alert desk handled 212 cases in which a supplier's mailbox had been compromised and a payment requested to a new bank account. The median loss was €18,400, often a large share of a small firm's monthly cash. Several EU countries were moving towards mandatory electronic invoicing between businesses, on differing timetables. As more invoices arrived as structured data, comparing bank details automatically against a supplier's history became practical.
Our chief executive and two large partners wanted the feature to block payments to changed accounts through integrations with accounting software. It would have been a stronger sales message. I argued against it. We could not see every payment route, and a block applied only some of the time would teach customers to stop checking for themselves. Our legal counsel also feared that blocking would be read as an assurance we could not give. We agreed to warn, not block, and dropped a proposed 'verified supplier' badge.
The feature compared the bank account on each incoming invoice with earlier invoices from the same supplier, for structured e-invoices and for PDFs arriving in connected mailboxes. When details changed, the customer received a warning and a short call-back checklist using a telephone number already on file, never one printed on the new invoice.
In the twelve months to September 2025, 11,400 customers enabled the check and it flagged 4,860 changes. After calling suppliers, customers confirmed 61 fraud attempts covering €2.7 million of invoices. Nine customers paid despite the warning and lost €212,000 between them. At first, 74 per cent of flags were legitimate changes; letting suppliers give advance notice through the customer reduced that to 52 per cent.
What I would do differently is the launch scope. We included scanned PDFs from the start, and our extraction misread roughly one scanned invoice in fourteen. The first release should have covered structured invoices and digital PDFs only, and said so plainly.
An assistant that explains alerts but cannot close them
By 2024 owners acknowledged only 34 per cent of medium-severity alerts within 24 hours. Many called the alert desk simply to ask what a message meant, absorbing about a fifth of analyst time. Language models could now explain an alert in plain Portuguese, Spanish or English. Several competitors were advertising automated response, in which software investigates and closes alerts without a person.
Pradomar Ventures, our largest investor, asked at a 2024 board meeting whether we should position our assistant the same way. A Spanish distributor offered a contract worth about €1.2 million in annual recurring revenue if we did. I recommended a narrower product: the assistant would explain alerts and draft staff messages, but it could not dismiss, close or alter any alert. The board accepted the recommendation, and we lost the distributor.
Before release, analysts graded 2,400 explanations across all 38 alert types. In the first pilot, 7.9 per cent contained a material error, such as recommending a password reset for the wrong account. Restricting the assistant to the fourteen types with written playbooks reduced that to 2.6 per cent. With our data-protection officer, we limited the model's inputs to alert metadata, ran it in an EU region and labelled every explanation as automated.
By June 2026 the assistant explained 61 per cent of the alerts shown to Standard and Complete customers. Acknowledgement within 24 hours for covered types rose from 34 to 52 per cent, and 'what does this mean' contacts to the desk fell by 29 per cent. We recorded three complaints where the advice was wrong; each was corrected and added to the test set.
The company now has a written standard for automated features, which the audit and risk committee reviews twice a year. What I would do differently is accessibility testing. The assistant's panel failed keyboard navigation at launch, and screen-reader users could not reach the draft message for five weeks. Our release condition should have applied in full, and I let a deadline override it.