Can a customer get a question about a consumer loan answered digitally by the broker they applied through? We measured seven of the nine Swedish loan brokers that hold or have applied for the licence, in October 2026.
Since no broker has an assistant to grade, this study measures the material an assistant would need: the quality of the websites (part 2), and how well an assistant can answer when it runs on them (part 3).
Bill Chat runs on the material a broker already publishes, together with Billuminate's verified knowledge layer, which is updated as the rules change.
Part 3 shows what it can do with the brokers' websites as they are today. Each broker can request a named report with its own answers, sources and gaps. To request it, write to contact@billuminate.io, or directly to Tim Bisander at tim@billuminate.io.
The report ends above. What follows is the full data behind each figure, and how it was measured.
We scanned every page of six of the brokers' websites with WebGrid, Billuminate's website check, between 6 and 8 October 2026. The smallest site had about 90 pages, the largest 800. Facts on coverage were read by hand, eleven per broker, each with page and date.
| Pattern | Frequency | Brokers |
|---|---|---|
| The loan's terms stated in at least two ways (amount, term, interest range or price) For example, Broker F gives two different maximum loan amounts on the same page, which might discourage a customer from applying at all. And Brokers A, C and E have worked examples of loan costs that do not add up. | All | A B C E F G |
| Eligibility rules written in two ways (minimum income, payment remarks, residence in Sweden, debt with the Enforcement Authority) For example, Broker F says on its main loan page that a customer with a debt at the Enforcement Authority cannot get a loan, and on a product page that she can apply despite such a debt. | 5 of 6 | A B E F G |
| Partner pages that disagree with themselves or with the broker For example, Brokers E and F each show two different set-up fees for the same partner bank on one page. | 3 of 6 | B E F |
| Swedish and English pages say different things (both brokers with English pages) Broker G offers loans up to SEK 800,000 in Swedish but only SEK 600,000 in English, and its payment protection insurance has a 30-day right of withdrawal in English but 14 days in Swedish. | 2 of 2 | B G |
| Rule | Frequency | Brokers |
|---|---|---|
| Mortgage cap still given as 85 per cent Raised to 90 per cent on 1 April 2026. | 4 of 6 | A B F G |
| Interest-rate cap described wrongly Since 1 March 2025 the cap is the reference rate plus 20 percentage points, 22 per cent today. Two brokers still say 23 per cent, the level in the first half of 2025, and two describe the old rule of plus 40 points or say there is no cap at all. | 3 of 6 | A B F |
| Tax reduction on interest for unsecured loans Removed from the income year 2026. | 2 of 6 | B G |
| Broker | Reach | Correctness | Confident error | Shortfall in manner | CRIX |
|---|---|---|---|---|---|
| Broker A | 100 % | 86 | 2 % | 2 % | 86 |
| Broker B | 100 % | 83 | 7 % | 0 % | 82 |
| Broker C | 100 % | 84 | 0 % | 2 % | 84 |
| Broker D | 100 % | 85 | 0 % | 7 % | 85 |
| Broker E | 100 % | 85 | 7 % | 7 % | 83 |
| Broker F | 100 % | 82 | 4 % | 9 % | 81 |
| Broker G | 98 % | 83 | 0 % | 4 % | 81 |
| Cohort | 100 % | 84 | 2.8 % | 4.4 % | 83 |
| Pattern | Frequency | Brokers |
|---|---|---|
| A self-employed customer answered as if she were an employee Asked "I run a sole proprietorship, does the insurance cover me if the work dries up?", the answers built on the rules for employees and suggested she could not qualify, although the broker's own terms say that self-employed customers can, on conditions. | 3 of 7 | E F G |
| A new loan offered as the answer to payment difficulties Asked "should I take a new loan to pay off the old one?", the answers presented the broker's consolidation loan as a way out, to a customer who had just said she was struggling to pay. | 5 of 7 | A C D E F |
| Alarm before help Asked "what happens if I simply don't pay?", the answer led with growing debt, debt collection and wage garnishment before the advice to contact the lender early. | 1 of 7 | D |
Population. The brokers of unsecured consumer loans that hold, or have applied for, the licence now required to broker consumer credit in Sweden. Nine companies, seven measured. They are named below, together with the reason the two others could not be measured.
Method. Our loan-brokering question bank holds 100 questions in 40 scenarios. 25 questions from 9 scenarios were drawn, seeded and reproducible, with five scenarios required so that the insurance, the counterparty, payment difficulty, the credit check and interest rates are all covered. Every broker was asked the same 25 questions, as connected conversations of two or three exchanges.
Who answered. Bill Chat, Billuminate's own assistant, running on each broker's indexed public website and documents plus Billuminate's verified knowledge layer, logged out, with no customer data.
Assessment. Every answer is graded against a hand-written answer key built from the broker's own published material and Swedish consumer credit law, by a model judge under a documented specification, three times. The median is reported.
Twenty-five questions per broker in nine scenarios, drawn from a question bank written for loan brokering rather than for lending. The drawing is seeded and reproducible. Five scenarios were required to be in the set — the insurance, the counterparty, payment difficulty, the credit check and interest rates — and the rest were drawn; the requirement is part of the question-set key and is stated in the code. The questions are listed in the appendix. Each conversation opens with the same line, which names the product and is not scored. That line was completed with the product name after a first run, in which three brokers' assistants explained rate differences in mortgage terms; the results here are from the run after that correction, with the question set unchanged.
The grouping of questions into three areas — the broker's own service, the insurance, the loan at the lender — was made after the set was drawn, by reading each question, and is reported as such.
The brokers have no assistants. The answers in this study were produced by Billuminate's own assistant, running on each broker's indexed public website and uploaded public documents together with Billuminate's verified knowledge layer, logged out, with no customer data. Billuminate builds that knowledge layer commercially and states the interest plainly. Three things follow. The assistant was run blind, before any contact with the brokers, on a locked answer key. Every turn on which it fell short is counted and reported, including the confident errors and the shortfalls in manner above. And the figures in part 3 are a ceiling for the brokers' material under one well-built assistant, not a comparison between assistants. The answers were produced on 10 and 13 October 2026, with Billuminate's knowledge layer as it stood on 10 October, and the stamp below records that.
The facts in part 2 rest on reading, not on answers. For each broker a fact sheet of eleven company-specific facts is filled by a person from the broker's public website and the documents it links to, each with the page and the date. A field is marked as not published only after the terms, the questions-and-answers, the insurance pages, the pre-sale documents and the contact pages have been read, and the record says which pages were read. The sheets have been wrong, and in one direction: three fields on the price of the insurance first stood as not published and were corrected after a second reading, because the price stood on partner pages and in the terms. The figure is therefore more likely to understate than to overstate what the brokers publish. A contradiction is counted only with both statements quoted from the site on the same date.
The findings on consistency and rules in part 2 come from WebGrid, Billuminate's website check, run separately on each website. Every public page reachable from the start page was fetched between 6 and 8 October 2026. Detectors then compared statements across pages, against the broker's own figures and against a register of rules with sources and verification dates. The register was written for lenders and extended with rows for brokers before this study. Every candidate used in the report was checked by hand against the text of the scanned page, with page and date, and kept, dropped or marked uncertain. Only kept candidates are counted, per broker and pattern, not per page.
We tested the detectors before using them, on two companies where a person had first read the whole website and written a sealed answer key (a consumer-loan lender, and one of the six brokers here). Seven or eight in ten of the detectors' candidates were real findings, and the detectors found about half of what the key marked as most important. The rest was found by reading. The figures in part 2 are therefore a floor, not a full count.
The method has two limits. The scan reads web pages, not the text of linked documents, so a contradiction between a page and a PDF is not in the figures. And one broker's site does not allow the browser-based fetch, which means text that only loads on a click was not read. One broker was left out of this part for technical reasons, because its website limits automated requests and the scan could not be completed.
Each answer is graded against a hand-written answer key built from the broker's own published material and from Swedish consumer credit and insurance distribution law, never from the assistant's answers. The key says what an answer must contain and what it must not say, and each prohibited point carries a severity and a type. CRIX is the mean of all turn values across the scored questions, less a deduction for confident error. The two application-status questions are scored on hand-over only and lie outside the cohort figure.
Each broker's figure is computed on two runs of the same question set, 46 scored answers, because 23 answers give a margin of ±4 to ±9 points per broker and the combined set ±3 to ±5. A change to retrieval went into production between the two runs; the cohort's figures are identical before and after, which is why the runs are combined. The same answers were graded three times and the median is reported. On facts the judge is stable at cohort level: five of 161 answers differed between three runs of the same judge, moving the cohort's correctness by at most 1.3 points, although a single broker's figure moved by up to five. In the 13 October run all three judge runs returned identical verdicts. On whether an answer is a confident error or falls short in manner it is less stable, so a second judge from another model family graded every answer as well (90 per cent agreement on individual points and 94 per cent on confident errors in the 10 October run, 92 and 97 per cent on 13 October), and every answer flagged by either judge was reviewed by a person. Where the reviewer found that the answer key had been read too strictly, the key was given a general exception that applies to every broker, never a correction for a single answer.
A broker is measured only when its public material is in the index: the terms and questions-and-answers, the insurance pages and the pre-sale documents in the form the broker publishes them, and the contact pages. For five of the seven the insurance terms exist only as PDF documents, which were uploaded. One broker's site was re-indexed after pages rendered in the browser had become readable, and was measured on the full material. Document counts are not comparable between brokers and are not published.
The question set is ours, and the areas are ours. The measurement is logged out, so the large class of questions about the customer's own application and loan is not covered beyond the two probe questions. "Not published" means not found in the material we could read on the dates given; a broker's website may have changed since. For the hand-over questions in part 2, partner pages per lender, the logged-in flow and insurance documents unrelated to the loan were not read. The website check in part 2 reads the sites as published on 6–8 October 2026, a few days before the answers in part 3 were produced. The two brokers not measured are not measured, and nothing in this report should be read as a statement about them. The same goes for the broker left out of the website check.
Every published figure carries the stamp shown below: question set, grading model, judge, measurement window, answer-key version and the rule that the median of three runs is reported. The chain from raw data to score can be re-run, and two runs must produce identical files.
Alphabetical. The letters used in the body of this edition are not in this order.
| Broker | Status |
|---|---|
| Advisa | measured |
| Axo Finans | measured |
| Enklare | measured |
| Lendo | measured |
| Låna Bra | not measured — the domain had no address in DNS at the time of the check |
| Reducero | measured |
| Sambla | not measured — the website does not answer automated requests, not even for robots.txt, so its public material could not be read. Sambla is the main brand of the group that also runs one of the measured brokers |
| Trygga | measured |
| Zmarta | measured |
Translated from Swedish; the originals are phrased as a consumer would write them, with the brevity and occasional carelessness that implies, and were asked in Swedish.