Wilhelm Jansson Stiernblad

Founder and CEO of Regulatory Knowledge - Building BARK, the future of banking with agents

[Wilhelm Jansson Stiernblad @Wilhelm_BARK]

I think the idea that building AI into the tools we have been using is the wrong approach. Instead the future will be one interface that brings the tools to you. Like CODEX has the web browser built in.

However for that to work well we will need a standard on agent communication so UI/UX = AX (agent experience). I am not a believer in AI using the UI, there is so much waste that the agent does not need.

Give the agents structured data that represents the same thing the user can see, then the agents can interact with it efficiently and navigate in the app.

Looking forward to see if @AnthropicAI, @OpenAI or @GeminiApp will lead the way here.

[Wilhelm Jansson Stiernblad @Wilhelm_BARK]

I think specialized agents are deeply misleading of the actual purpose . The agents essentially have a finite token amount, by managing the prompt around the task, we can steer what areas of the token space gets activated.

It is not about giving the agent a role, it is steering what neurons should fire!

[Wilhelm Jansson Stiernblad @Wilhelm_BARK]

Things that had not changed with Opus 4.8 versus Sonnet 3.7, the LLM tends to prefer code analysis and coming up with complex answers to issues instead of doing simple testing and reading the errors, e.g. an index missing when loading data into a DB.

[Wilhelm Jansson Stiernblad @Wilhelm_BARK]

Non-Maturity Deposit (NMD) modeling is one of the hardest problems in banking right now. Few real experts, genuinely complex requirements — stable/non-stable, core/non-core, pass-through rates — all calibrated on at least 10 years of data.

Work that used to take weeks or months: done in an afternoon. Every requirement captured and cited, fully backtested, model documentation written, output mapped to the J-templates — ready for human review and sign-off.

[Wilhelm Jansson Stiernblad @Wilhelm_BARK]

What are the criteria to apply 100% risk weight on an ADC loan?" Hours of CRR, Q&A and EBA guidelines for a bank's risk team, then defending your reading to the regulator.

BARK Research scopes the question, investigates, and returns a fully cited brief. Every claim is backed by and connected to the regulatory source text. Click through to the exact paragraph to validate yourself.
Interpretation that used to take hours, ready to defend and reduce your capital requirements.

Banking made easy.

[Wilhelm Jansson Stiernblad @Wilhelm_BARK]

I think @jamdac and @a16z are spot on here, compliance is becoming software. However you can identify when someone has not the full picture of how a bank actually works, they mistake regulations for compliance. Inside a bank, the regulation that matters most is not compliance, it is the business itself constrained under dense regulations.

Onboarding, KYC, AML, monitoring etc. those are all compliance issues. AI will reduce costs, improve accuracy and speed and as a result improve the customer experience. The software opportunities are real, and I do not think incumbents can just go headless, the systems were not natively built for agent experience, AX. However that is all distribution, smoothing the funnel. It does not change the name of the game, and it does not move the underlying economics.

Then there is the part of regulations that is the actual business, the balance sheet and how you run it. The part of the business that lives in Treasury, ALM, capital and liquidity management, model validation, second line. The part the CFO and CRO are responsible for. Here the rules are not binary gates, they are the playing field. Designing an F1 car, if you are not compliant with the rule book you are disqualified, but compliant with the rulebook is not what makes you win the race.

A clear example: UniCredit. In a Bloomberg interview, the head of its balance sheet management, Stefano Chiarlone, said the bank now folds the capital benefit of significant risk transfer (SRT) straight into how it grants and prices new loans — evolving SRTs from a pure capital-efficiency tool into something that "directly improves the competitiveness of our bankers when they originate loans." It is targeting SRTs on EUR 14-16bn of loans this year, potentially EUR 20bn. That is regulation as a pricing weapon: a CRR/SecReg construct wired into the front book to win business — and notably, it is run out of Treasury, not the compliance function.

But to actually run an SRT programme you need, simplified: policies, instructions and routines to govern it; the data — not once, but continuously; systems to manage the transactions; a core ledger that can mark and track them; regulatory reporting (COREP/FINREP/ESMA); investor reporting; and so on.

And from the “compliance” aspect, in an EU context, touches at least the following:

Tier 1:
- CRR (2013/575): capital treatment and the SRT tests (Art. 244/245)
- CRD (2013/36): supervision, Pillar 2, the SREP overlay
- Securitisation Regulation (2017/2402): retention, due diligence, transparency, STS

Tier 2:
- (2023/2175) RTS on risk retention
- (2024/920) RTS on performance triggers
- (2017/2402) RTS on homogeneity of the pool of credits
- (2020/1224) RTS & (2020/1225) ITS on disclosure and reporting (loan tape etc.)
- (2020/1226) RTS & (2020/1227) ITS on STS notification (to achieve preferential treatment)
- ITS on reporting (COREP/FINREP)

Tier 3:
- (EBA/GL/2014/05) EBA GL on SRT
- EBA report on SRT
- (EBA/GL/2018/09, EBA/GL/2018/08, EBA/GL/2024/05) EBA GLs on STS criteria
- Several EBA Q&A’s for clarifications etc.

Tier 4:
- Local rules and requirements

As anyone can see, that is a very large corpus for one specific type of transaction - though one that can fundamentally move the competitiveness of a bank, or transform it from capital heavy to capital light.

The idea that any single individual can master this is of course wrong, and even if they did, this is not covering the data and modeling side, investor relations and the rest.

A few years ago I had dinner with the structuring desk at Citi in London, and they told me they had stopped structuring synthetic SRT deals, the complexity and support needed for the originator, simply did not make sense. That has probably changed with the uptick in the SRT market. But it captures the reality, the complexity is enormous, and not just in interpreting regulations, turning that into computation and pairing it with actual portfolio data is a major task.

This is where the a16z piece comes in. To use software to make banking more efficient, the current software simply is not up to the task. You need the right foundation — one built with both agents and humans in mind, which in my view are fundamentally the same requirement. Humans need a UI/UX layer, but that is just a representation of well-structured data which is what the agent needs. Well-structured data, in proper hierarchies that let the user or the agent zoom in and out and that map how regulation flows into internal process and into the data — that is not, in itself, a new problem. What is new is what is possible today that wasn't six months ago, and what becomes possible in the years ahead.

That is why we are building BARK: a platform that enables agents and human experts move quicker, smarter and fundamentally different than before. It is built on three pillars:

1. Regulatory knowledge - a knowledge graph of the rules and how they interconnect. The playing field.
2. Institutional knowledge - how a given bank actually operates. Fully connected to the regulations, the playing field.
3. Structured financial data - the balance sheet, at contract and exposure level.

With those three in place, you can put agents on top and make durable decisions that scale and compound. This is not about building agents to improve a specific process. It is about enabling agents to build whatever process you need.

[a16z @a16z]

Compliance officers are one of the fastest growing occupations in America.

Compliance is a bigger business than you'd think. Every dollar that leaves or enters a business: paying employees, reporting revenue, and moving capital are subject to compliance.

As AI clears the "good enough to trust" bar and sales cycles speed up, there may finally be an opening for startups.

Full piece from a16z's @jamdac and @astrange: https://t.co/niRB3jPioN

[Wilhelm Jansson Stiernblad @Wilhelm_BARK]

I think @jamdac and @a16z are spot on here, compliance is becoming software. However you can identify when someone has not the full picture of how a bank actually works, they mistake regulations for compliance. Inside a bank, the regulation that matters most is not compliance, it is the business itself constrained under dense regulations.

See my full comment here https://t.co/Oi9ITDMUHz

[Wilhelm Jansson Stiernblad @Wilhelm_BARK]

https://t.co/ErgLadVfp0

[Wilhelm Jansson Stiernblad @Wilhelm_BARK]

https://t.co/ErgLadVfp0