Ruleguard

The agents arrived after a decade of platform building, layered on top of a working platform rather than native to it. The company was founded in 2013, built its first product in 2015 for the incoming client assets rules, and assembled roughly twenty compliance modules on a shared evidence layer before the agentic positioning arrived. Take the models away and all of that remains: registers, attestations and breach management, along with monitoring plans, policy libraries and task lists.

The vendor states the relationship plainly, arguing that regulators are not asking for AI but for evidence, transparency and accountability. That frames the models as a means to an established end. Still, the agentic layer is presented as the platform's current identity, with its own product surface, and it does the reading and reasoning across every module rather than sitting as one feature in a corner.

Autonomy is configurable by the customer, but the gates themselves are not described. The vendor publishes that oversight checkpoints are configurable and that the customer, rather than a vendor default, sets the level of autonomy. It pairs this with permanent recording of every agent decision, its data source and its output, plus immutable audit trails and the stated principle that the compliance team stays in control. What is missing is the substance behind the claim: no checkpoint is described, no threshold or confidence measure is published, and no sampling audit of automated decisions is mentioned.

There is no published accuracy, precision or recall figure, but two disclosures carry real weight. Every agent decision is recorded permanently together with the data source it drew on and the output it produced. That makes any determination reconstructable after the fact rather than merely explicable at the time.

An externally audited ISO 42001 management system is the mechanism through which model performance evaluation, risk assessment and lifecycle control are supposed to be documented. That is closer to what a supervisor examining model risk actually asks for than a headline accuracy percentage would be. What remains missing is the number itself, and with it a failure mode that recurs across this category: when an agent never flags an obligation, nothing is recorded, because there is no decision.

Thirteen years of operating history and institutional backing, and not one customer named anywhere. The claim is that tier one banks, global insurers and leading wealth managers use the platform, but no institution is identified and no customer count is given. No case study or analyst evaluation is published. A 3.5 million pound investment from a named growth investor in 2022 is a diligence signal about the company rather than a measurement of the product.

No quantified outcome of any kind appears: no review time saved, no error reduction, no efficiency figure. That is the sharpest contrast in the vendor's profile, since it discloses a great deal about how its technology is governed and little about whether it works.

Nothing published says whether one firm's policies, evidence or agent interactions inform models, retrieval or outputs serving another. The agents are explicitly described as trained on the firm's own policies, which is exactly the material a competitor would not want shared. One real control exists. An on premises deployment option is available, and a customer that takes it keeps the question from arising at all.

That answers it for the firms that choose it and leaves silence for everyone on the cloud path, which will be most of them. The accompanying framing about the customer's data and the customer's rules is a slogan attached to a deployment choice rather than a stated position on tenant separation or model training.

No privacy statement, lawful basis, retention rule or handling commitment is published. An information security certification does not fill that gap, since it covers how data is protected rather than what may be collected, why, and for how long. The scope reaches personal data on two distinct populations. Employee modules hold attestations, personal account dealing records, gifts and hospitality declarations and conflicts disclosures on named staff, which is career relevant material.

Client file review and complaints modules hold records about the firm's own customers, including the substance of grievances. Hosting across several regions is a deployment strength and is not a privacy position either.

The credential is named, current and independently verified: ISO 27001. Naming the standard is what separates a real credential from an assertion of compliance without specifics. It is supported by described infrastructure, with data centers named across four regions rather than a generic cloud claim. No trust center serving live artifacts, SOC report or equivalent second attestation is public, and no penetration testing cadence or audit period is stated. A buyer can confirm that a certification exists without seeing its scope statement or its currency.

The company is positioned as a technology supplier, and its regime mapping is granular, with an origin story that shows it. It built its first platform in 2015 specifically for incoming client asset rules. The module list now enumerates named British obligations one by one. They include client assets, the senior managers and certification regime and Consumer Duty, along with appointed representative oversight, product governance and operational resilience.

That is a vendor organized around the rulebook rather than around a technology. No formal admission process, regulator run program or supervised test of the product is public, and no supervisor has examined the agents themselves.

ISO 42001 compliance, audited by the British Standards Institution, gives this vendor a real external governance credential. That body holds the first national accreditation to certify against the AI management system standard. The vendor pairs the credential with permanent recording of every agent decision and data source. Its own wording draws a distinction, though.

Information security is described as certification, independently verified, while the AI standard is described as compliance being audited, and no certification announcement is published. A firm holding certification says certification. No fairness position or bias testing is published separately.

What the product produces raises the stakes. The agents generate audit ready evidence that a firm presents to its supervisor. An incomplete or incorrect evidence pack does more than inconvenience the compliance team: it becomes the evidence on which an examination turns, and the finding lands on the regulated firm. No accuracy warranty, service commitment or remedy is published. A second population sits inside the platform without a contract.

Employees have their attestations, personal trading and hospitality declarations assessed, and customers have their complaints and file reviews processed. Neither group has a described route to contest a determination made about them.

The class of model is disclosed, along with the deployment choice around it: the platform runs on frontier cloud models or on premises. That tells a buyer that third party foundation models are in use rather than proprietary ones, and that an alternative exists for a firm whose risk appetite or jurisdiction rules that out. Combined with an externally audited AI management system under ISO 42001, it gives a procurement team something to work with.

No provider, base model or version is named, so a firm cannot document which specific model its compliance controls depend on. A platform built on frontier models also inherits whatever changes beneath it when a provider revises one.

The distinguishing fact is architectural rather than a connector list. Roughly twenty modules run on one shared evidence and task layer, so a breach recorded in one place, an attestation gathered in another and a monitoring finding in a third all resolve against the same evidence base. That is the integration most firms are trying to buy when they consolidate point solutions.

An on premises deployment path adds real estate depth, since running inside a bank's own environment is a harder engineering commitment than a hosted tenancy. The outward facing layer is missing. No named integration to a core banking platform, human resources system, customer relationship system or document repository is published, and there is no public API documentation or connector catalog.

Residency is answered in detail. Hosting is offered across four named regions covering the United States, United Kingdom, European Union and Asia Pacific, which is region choice rather than a single stated location. An on premises deployment path exists for firms that cannot accept a hosted model at all. The choice also extends to where the intelligence runs, with the vendor offering either frontier cloud models or an on premises alternative, a layer that is often left unaddressed.

A firm with a data localization obligation can therefore determine from public material that a compliant configuration exists. Two gaps remain: no tenancy model is described and no subprocessor list is published.

The closest thing to a pricing signal is a promise of compliance without spiraling costs, which describes the direction of a number rather than the number. No rate card, tier ladder or billing basis is published. The gap matters more for a modular platform than for a single product. Roughly twenty modules can be bought in many combinations, and public material gives no indication whether pricing runs per module, per user, per entity or per regulated permission. A firm trying to compare this against a point solution for one obligation cannot build the comparison from anything published.

Functional breadth is the strongest part of the claim, and it is real: roughly twenty modules covering prudential, conduct, employee and third party obligations on one evidence layer, a wider surface than a pure point solution. Infrastructure breadth supports it, with hosting offered across four regions spanning the United States, United Kingdom, European Union and Asia Pacific, which implies customers in each. Two limits matter.

The encoded regimes are overwhelmingly British, built around client assets rules, the senior managers regime, Consumer Duty and appointed representatives, so the depth is national even where the hosting is not. And the buyer claim of tier one banks and global insurers rests on no named institution and no customer count.