Built for the industry that can’t take answers on faith.
Banks and insurers don’t reject AI because they lack ambition. They reject it because most of it can’t answer the questions their risk, security and audit teams are paid to ask. Kallidin was built to answer them.
The problem, in your world
In financial services every vendor passes through the same gate: risk, security, procurement, and increasingly the regulator’s implied questions behind them. If you cannot show where the data lives, the conversation is over. If you cannot show how an answer was reached, the answer cannot be used, because the person who owns the results answers for them. Most AI tools treat that bar as friction. It is not friction. It is the job.
Start with the report you already have to write.
Under Consumer Duty you have to evidence that outcomes do not differ unfairly between groups of customers, including customers in vulnerable circumstances, and you have to put that evidence in front of a board. Most firms produce it the same way: a working group, a handful of analysts, a shared drive, and about six weeks.
It is also the hardest possible test of whether an AI tool is any use in a regulated business. An answer nobody can defend is worse than no answer at all.
Did outcomes differ for customers in financial difficulty last quarter, and was there an early warning signal you could have acted on?
KAL builds the cohort from the conditions you actually use: IFRS9 stage, days past due, collections contact, vulnerability flags, churn in the period. It compares outcomes across every segment. It names the segment that fared worst and by how much. It separates the early warning signals that genuinely pre-dated the difficulty from the ones that do not survive scrutiny. And it tells you where its own analysis is thin.
Every number arrives bound to a proof your second line can re-run, with the definitions and assumptions written next to it. Which is the entire point. The report is only worth writing if it holds when somebody pushes back on it.
What an answer looks like.
The data below is synthetic. We built it to test KAL, not to flatter it. We are showing you this run so you can see the shape of an answer, not so you can learn something about banking. If you want to see it on your own estate, that is what a discovery call is for.
558 customers meeting at least one of eight financial difficulty conditions in the quarter, compared against a control group of 3,442.
Mass Market customers bore the worst composite outcomes. Churn at 45.4%, more than ten percentage points above the cohort average, accounting for 93% of every churned customer in the cohort.
Premier and Private customers had worse per-loan distress rates, but close to zero churn, which suggests those relationships were being actively managed. It said so, rather than picking the worst number and stopping there.
An NPS detractor score, observable at any point beforehand. And a prior contact with the Collections team, visible up to twelve months before the difficulty window opened.
We seeded the estate with app abandonment as a plausible early warning of financial difficulty. It is not one. In the data the difficulty cohort was in fact less affected than the control group, and KAL reported that as a finding in its own right rather than quietly dropping it.
That Premier and Private were small cohorts, so their rates carried higher sampling uncertainty. Nobody asked it that.
Nothing above is a claim about what KAL might do. It is a specimen of what it produces: the verdict, the nuance it did not skip, the signals that failed, and the limits it declared on itself. Judge the standard, then ask us to meet it on your data.
The same shape of question, elsewhere in your building.
They are all the same request underneath: take a defined population, compare outcomes across groups, evidence the difference, and be able to show your working when someone asks. That is what KAL does, and it does not care which obligation prompted the question.
What you get
KAL, the Autonomous Data Office, is built for that bar, not retrofitted to it. Answers to business questions from your data, each one checked by an independent AI agent before it reaches you, each one carrying an audit trail of what was checked, what was assumed and why it passed. Your data stays in your own environment, in the UK. The consulting that comes with it fixes the foundations underneath, the access, governance and quality problems that stall programmes hardest in regulated businesses.
The trust check, done early
Your security team will have a questionnaire. We’d rather see it before the first call than after the last one. Residency, audit trails, access control, versioning and exit are answered on our trust page, including where our own certifications stand. If your team cannot find an answer there, that is a gap we want to know about.
Why us, for this sector
Our first business, Aquila Insight, spent five years inside UK banks, insurers, retailers and airlines before Merkle bought it in 2017. Sam Riddington has spent more than 25 years in enterprise consulting at IBM, Accenture and Optima. We know what an audit committee asks, because we have sat in the room when it asked.
Questions we hear from financial services
Could KAL produce our Consumer Duty outcomes evidence?
Can this operate inside our regulatory environment?
What happens when the AI gets something wrong?
Would our second line accept an analysis an AI produced?
Start with a conversation.
The first step is a discovery call, and the usual next step is the diagnostic: a short, flat fee piece of work that finds what is broken and designs the fix. Bring your hardest trust question with you.