Top Agentic AI for FCA-Regulated Firms

Introduction

Agentic AI asks more of a regulated firm than earlier generations of software did. A chatbot that answered from a script could be reviewed line by line. An agent that reasons, retrieves information and takes actions in other systems cannot be reviewed that way, and yet the firm remains answerable for every outcome it produces for a customer. For a bank, an insurer, a wealth manager or a payments business supervised by the FCA, the question is therefore broader than which platform builds the cleverest agent. It is which platform lets the firm show, at any later date, how an agent was designed, what it was permitted to do, why it acted as it did and who was accountable.

This article assesses seven platforms against that question: Nexus, Hebbia, Microsoft Copilot Studio, Google Gemini Enterprise, Glean, WRITER and Kore.ai. It is written for the people who carry the decision inside a regulated firm: heads of operations and customer service, chief data and AI officers, compliance and risk leads, information security and procurement.

Every statement about a vendor below carries a numbered source. Almost all lead to the vendor's own public pages, checked on 8 November 2026; a small set of Nexus delivery and commercial facts come from information confirmed by Nexus¹. When a vendor's pages say nothing on a point, this article records it as not publicly stated on the pages we opened. The assessment does not summarise or interpret FCA rules; it frames the regulatory angle as the practical questions a compliance function is likely to ask, and the firm's own compliance team remains the authority on what its obligations require.

How to use the assessment

Bring a defined customer journey and its approval rules to this comparison. The profiles separate the work of building an agent from the evidence a firm needs to review its use: permissions, escalation, release controls and contract terms. A vendor's certification list does not establish that a particular deployment meets the firm's obligations. Treat unanswered questions as diligence items, and have the firm's compliance function decide what must be demonstrated before release.

The order reflects this guide's assessment of fit for an FCA-regulated firm that wants accountable help to get its first agents into production under controls its compliance function can sign off. A firm with different priorities, such as full on premises hosting or the published entry price, may reasonably rank the platforms differently, and the criteria section gives it the material to do so.

The seven, in order

1. Nexus. Experienced consultants build alongside the firm, with a first agent running 4 weeks into a 3 month proof of concept, and no agent is released until its data flow, access model and audit trail have been approved^1,2. 2. Hebbia. Built for finance, with traceability back to every finding, SOC 2 Type 2 and ISO 42001, and US and EU regional processing^3,4,5,6,7. 3. Microsoft Copilot Studio. The deepest published governance toolkit, with approval gates, Purview audit and customer managed keys, built by the firm itself^8,9,10. 4. Google Gemini Enterprise. A unique identity for each agent, Model Armor in every edition and data residency in the US or EU^11,12,13. 5. Glean. Permission enforcement on every read and write and human approval before agent actions, with several agent governance controls still coming soon^14,15,16. 6. WRITER. Strong security and zero data retention, an embedded delivery team, and a focus on marketing and revenue work^17,18,19. 7. Kore.ai. Compliance rules enforced outside the model and an on premises option, with pricing and product limits not publicly stated^20,21.

What changes when the buyer is regulated

Three operating questions shape this comparison for an FCA-regulated firm: who controls the agent, what evidence remains after an action, and how customers reach a person when needed.

The first is accountability. A regulated firm cannot hand responsibility for a customer outcome to a supplier. It can, however, choose a supplier that takes responsibility for getting the design right with it, and that leaves a record a reviewer can follow. A platform that supplies only tools leaves the firm to fill that role with its own people, who are often already committed to regulatory change.

The second is evidence. A firm will want to show that an agent was tested before release, that its effect was measured against an agreed aim, that its behaviour can be reconstructed and that a human could step in. Each of those needs a control in the product and a habit in the delivery process.

The third is the treatment of customers, which is where this site's Consumer Duty angle comes in. For an evaluation, consider how an agent that sits in front of retail customers becomes part of how the firm delivers outcomes to them, and the practical questions for the platform follow: whether the agent answers only from approved material, whether it escalates when uncertain, whether a person can approve an action before it touches a customer's account, and whether the firm can review interactions at scale.

The seven criteria later in this article follow from these three points.

1. Nexus

Nexus sells procedures that operate themselves²², capturing the know how of a company's strongest staff in procedures that then run without them²³. It is a young company. It says it started Nexus Agent in 2024²⁴, it has raised a 4.3 million dollar seed round led by General Catalyst²⁵, and it calls itself early²⁶. It is placed first because its delivery and control model answers the regulated firm's questions more directly than any other contender's public material.

Accountable delivery. A Nexus engineer builds the first procedure with the client team and shows it running in the client's own tools²⁷. A first version runs within weeks, after which Nexus keeps working on it with the client²⁸. Its first agent goes live in 4 weeks, within a 3 month proof of concept¹. Experienced consultants run each engagement, and one of the founders previously worked at McKinsey¹. Integration code is now written by Cue, the company's build loop, which leaves its Forward Deployed Consultants free of that task^29,30; Cue also scopes the use case and links to internal data under the right permissions³¹.

Evidence of outcomes. Every agent Nexus deploys has a measurable business KPI from the first day³², and where the KPI does not move the team rescopes instead of shipping³³. Its client list, by its own count, covers 12 enterprises worth more than 70 billion dollars between them, Orange Group and Lambda.ai among them^34,35. At Orange, the result was 6 million dollars of extra lifetime value³⁶.

Release control and escalation. An agent stays unreleased until a CIO is able to approve its data flow, its access model and its audit trail². Every action is visible in one place, with permissions and guardrails set once, centrally^37,38. Agents escalate with full context when uncertain³⁹. Full audit trails, traceable decisions and access by role are all listed on the comparison page⁴⁰.

Certifications and data. On assurance, GDPR, ISO 27001 and ISO 42001 appear on its comparison page⁴¹, and it holds a SOC 2 Type 2 report¹. Its privacy policy sets out legal bases under the GDPR and UK GDPR⁴².

Commercial model. The unit of price is the role automated; Nexus charges no per diem and does not bill for consultant time¹. Each engagement opens with a 3 month proof of concept tied to agreed outcomes⁴³. If it does not deliver, the customer walks away⁴⁴.

Limits. Older terms from 2023, drafted for the gpt.nexus site, state that the services are not tailored to industry specific regulation⁴⁵ and place hosting in the United States⁴⁶. On the pages we opened, single sign on, encryption standards, penetration testing and service levels are not publicly stated, and the same applies to a choice of hosting region or processing confined to the EU. A UK firm should obtain the current enterprise contract and written answers on each.

2. Hebbia

Hebbia presents itself as AI built for the rigour of finance³, trusted by investors, bankers, lawyers and consultants⁴⁷. Its Max product researches and analyses deals to produce client ready spreadsheets, slides and reports⁴⁸, and reasons over documents with complete traceability back to every finding⁴.

Strengths for a regulated firm. The certification list includes ISO/IEC 42001:2023, SOC 2 Type 2 and ISO 27001:2022^5,6,49, with GDPR on the home page⁵⁰. Hebbia says its staff cannot reach customer data directly, with production access granted just in time⁵¹. Data is held in siloed environments isolated from other customers⁵², and the customer can see who accessed its data, when and what they did⁵³. Documents, prompts and outputs are never used for training⁵⁴. Processing runs across the US and EU with regional options⁷, and a dedicated tenant is available if needed⁵⁵. Integrations are finance specific, including Datasite data rooms, Fitch ratings and European regulatory filings^56,57,58.

Limits. How implementation runs, and how long it takes, is not publicly stated on the pages we opened, and no prices appeared because the pricing page failed to load⁵⁹. Its published segments centre on investing, banking, legal and corporate finance work^60,61,62,63, so a firm seeking customer facing agents should confirm scope.

Why second. Hebbia is the most finance shaped product here and its data controls are strong. Nexus is ahead because it states who delivers the first agents, how fast and under what release gate.

3. Microsoft Copilot Studio

Copilot Studio is Microsoft's low code studio for agents and workflows⁶⁴. For a firm already on Microsoft identity and security tools, it extends controls the firm already uses.

Strengths for a regulated firm. Administrators can turn off agent publishing for the tenant⁶⁵ and control which channels agents use⁶⁶. Release pipelines, approval gates and environment promotion support change control⁸. Maker audit logs are visible in Microsoft Purview⁹, and agent activity alerts flow through Microsoft Sentinel⁶⁷. Customer managed keys¹⁰ and Customer Lockbox⁶⁸ are supported. With ungrounded responses switched off, the agent blocks any answer produced without a knowledge source or tool⁶⁹. The Customer Conversation harness can escalate to a human representative through a defined path⁷⁰. The price is published: Copilot Credit packs at 200 dollars per pack per month⁷¹. Microsoft's customer stories include Virgin Money⁷².

Limits. Certifications are not named on the security page, which points to a separate compliance offerings page⁷³. Files carrying confidential or highly confidential sensitivity labels are excluded from indexing⁷⁴. Each user is capped at 10 sessions a day across every agent in the tenant⁷⁵. Building is done by the firm's own makers and developers⁷⁶.

Why third. Copilot Studio's governance is detailed and familiar to Microsoft estates. It sits behind Nexus and Hebbia because the firm carries the design and evidence work, and because its content limits bear on regulated document sets.

4. Google Gemini Enterprise

Gemini Enterprise combines intranet search, an assistant and an agentic platform⁷⁷, with the Agent Platform sold separately⁷⁸.

Strengths for a regulated firm. Each agent receives a unique SPIFFE identity¹¹, secured by default with Context Aware Access⁷⁹. An Agent Gateway handles agent traffic⁸⁰, plain language rules govern tool use⁸¹, and Model Armor, included in every edition, scans prompts and tool responses^82,12. Data residency is available in the US, EU and global locations¹³, with Access Transparency and Cloud Audit Logs on enterprise editions⁸³. Customer managed and external keys are supported^84,85. Google's foundation models are not trained on a customer's prompts, answers or enterprise data⁸⁶.

Limits. Choosing the global region rules out both customer managed keys and Access Transparency⁸⁷. What Standard and Plus cost per seat is not publicly stated on the pages we opened⁸⁸. Sensitive Data Protection does not apply content policies to older data stores⁸⁹. Third party connectors use public endpoints outside Google's network⁹⁰, and customers cannot add custom network controls⁹¹.

Why fourth. Google documents agent identity and data location precisely. The firm still builds and evidences its own agents, and several protections depend on region and store type.

5. Glean

Glean grew from enterprise search into an enterprise AI platform built on company context⁹².

Strengths for a regulated firm. Source system permissions are enforced on every read and write¹⁴ and changes are reflected immediately⁹³, which supports information barriers. Human approval can be required before agent actions execute¹⁵, and agent actions are checked against the user's intent and authority⁹⁴. An always on audit trail can stream to a SIEM⁹⁵. The security and compliance list includes ISO 42001, SOC 2 Type II, ISO 27001 and GDPR^96,97,98,99. DBS Bank's Chief Transformation Officer is quoted on the home page¹⁰⁰.

Limits. Agent governance policy enforcement and policies for select agents or users are marked coming soon^16,101, and browser use control is in beta¹⁰². Data residency options and price are not publicly stated on the pages we opened¹⁰³. Connector counts differ between pages^104,105.

Why fifth. Glean's permission model is the strongest answer to internal leakage risk among the contenders. For customer facing agents, its governance controls are still being completed.

6. WRITER

WRITER positions itself as an end to end agentic AI platform¹⁰⁶, built for marketing and revenue teams¹⁹.

Strengths for a regulated firm. A SOC 2 Type II evaluation takes place every year¹⁰⁷, and ISO/IEC 27001, 27701 and 42001 are all held¹⁰⁸. It takes a zero data retention approach¹⁷ and does not train on customer data¹⁰⁹. Guardrails check content before, during and after a call¹¹⁰, and a DLP guardrail blocks personal, financial and identity data¹¹¹. Approval flows shape how agents act on connected systems¹¹², and audit events export to a SIEM¹¹³. An EU endpoint¹¹⁴, dedicated private cloud¹¹⁵ and bring your own key¹¹⁶ are offered. Clients are supported by transformation leads and deployment engineers embedded with them¹⁸.

Limits. The DLP guardrail does not inspect model output¹¹⁷. Encryption does not yet cover custom and no code agents¹¹⁸. Observability for Agent Builder v2 is not yet available¹¹⁹. Payment is by card, with offline invoicing not supported¹²⁰.

Why sixth. WRITER's security account is thorough and its delivery team is real. Its focus on marketing and revenue¹²¹ and its stated control gaps place it behind the broader platforms for this audience.

7. Kore.ai

Kore.ai calls itself an enterprise agentic AI platform¹²² and lists banking and wealth management among its industries¹²³.

Strengths for a regulated firm. Policies and compliance rules are enforced outside the LLM^20,124. StateTrace captures each agent's full execution path¹²⁵, and agent versions, rollouts and rollbacks can be managed¹²⁶. Forward Deployed Engineers are part of the offer¹²⁷, with proof on the customer's own use case before commitment¹²⁸. The trust centre lists ISO/IEC 27001:2022, SOC 2 Type II and PCI DSS^129,130,131, and deployment options include on premises and a dedicated single tenant VPC^21,132. Customer data is not used for internal training¹³³.

Limits. Kore.ai publishes no price. Product quotas and limits are not publicly stated on the pages we opened. The website was being rebuilt when the facts were gathered^134,135. The privacy policy allows personal information to be processed in another country¹³⁶.

Why seventh. Kore.ai's design choices fit regulated customer service well. Its public evidence is the thinnest of the seven, with key items listed without supporting detail.

The criteria

The seven criteria below translate the regulated firm's questions into points a platform can be tested on. Each paragraph names who answers the criterion most fully on the pages we opened.

1. Accountable delivery

Who designs, builds and proves the first agents, and who stays afterwards? Nexus gives the most direct answer. Its consultants build with the firm's people, an agent is running 4 weeks into the 3 month proof of concept, and the Nexus team remains afterwards to improve it^1,28. WRITER offers an embedded team of transformation leads and deployment engineers¹⁸, and Kore.ai offers Forward Deployed Engineers¹²⁷. Copilot Studio and Gemini Enterprise describe setup by the customer's own people^76,137. Hebbia and Glean do not state a standard implementation process on the pages we opened.

2. Evidence of outcomes

Can the firm show that an agent achieved what it was meant to? Nexus ties every agent to a business KPI from day one and rescopes when it does not move^32,33. Copilot Studio offers test sets and a shared grader library before and after publishing¹³⁸. Kore.ai describes Autoloop optimising agents against defined goals¹³⁹. A firm should add its own customer outcome measures to whichever platform it chooses.

3. Traceability and audit

Can the firm reconstruct what happened? For Nexus, the comparison page names full audit trails and decision traceability⁴⁰. Hebbia offers traceability back to every finding⁴ and access logs⁵³. Copilot Studio uses Purview⁹, Gemini Enterprise uses Cloud Audit Logs⁸³, and Glean and WRITER stream events to a SIEM^95,113. Kore.ai traces every run across reasoning, tools, guardrails and handoffs¹⁴⁰.

4. Release control and human oversight

Can the firm stop an agent from reaching customers until it is approved, and can a person intervene? Nexus will not ship without sign off on data flow, access model and audit trail, and its agents escalate with context^2,39. Copilot Studio has approval gates and can switch off publishing^8,65. Glean supports human approval before actions¹⁵, and WRITER describes approval flows¹¹². Kore.ai supports rollbacks¹²⁶.

5. Certifications

Which assurance reports exist? Nexus lists GDPR, ISO 27001 and ISO 42001 and holds SOC 2 Type 2^41,1. Hebbia lists ISO 42001, SOC 2 Type 2 and ISO 27001^5,6,49. Glean lists six items including ISO 42001¹⁴¹. WRITER holds ISO 27001, 27701 and 42001 with SOC 2 Type II^108,107. Google's list for Gemini Enterprise includes SOC 1, 2 and 3, ISO 27001 and PCI DSS¹⁴². Kore.ai lists ISO 27001, SOC 2 Type II and PCI DSS^129,130,131. Copilot Studio refers readers to the Microsoft Trust Center for its compliance evidence¹⁴³.

6. Data location, training and retention

Where is data processed, and is it used for training? Gemini Enterprise offers US and EU residency¹³. Hebbia processes across the US and EU⁷. WRITER has an EU endpoint¹¹⁴. Microsoft says Copilot Studio aligns with data residency requirements¹⁴⁴. Kore.ai offers on premises deployment²¹. For Glean, residency options are not publicly stated on the pages we opened. For Nexus, hosting choices are also not publicly stated there, and the only location given is the United States in the 2023 terms⁴⁶. Nexus, Hebbia, Gemini Enterprise, WRITER and Kore.ai each state a form of no training commitment^{145,54,86,109,133}, and Copilot Studio is governed by Microsoft's Product Terms and Data Protection Addendum¹⁴⁶. A firm should check the exact scope of each commitment, since they are worded differently.

7. Commercial terms and exit

Can the firm try before committing, and leave if it fails? Nexus starts with a 3 month proof of concept with a walk away option and prices per role automated^43,44,1. Kore.ai proves on the customer's use case before commitment¹²⁸. Copilot Studio and Gemini Enterprise publish prices^71,147. WRITER publishes Starter terms and model prices^148,149. Hebbia, Glean and Kore.ai do not publish prices on the pages we opened^59,103.

Buyer's guide by type of firm

Retail banks and building societies

The priority is customer facing service with tight control over what the agent says and when it hands over. A bank should test grounding controls such as Copilot Studio's block on ungrounded answers⁶⁹, escalation behaviour such as Nexus's escalation with context³⁹, and approval before actions that change an account. Nexus suits a bank that wants a delivery team to own the first deployment; Copilot Studio suits one with a strong Microsoft platform team.

Insurers

Claims and policy servicing combine customer contact with document heavy back office work. An insurer should weigh Nexus's delivery model and KPI discipline³² against Kore.ai's industry packages¹⁵⁰ and on premises option²¹. Hebbia is relevant for underwriting research where traceability to source documents matters⁴.

Wealth and asset managers

Research, client reporting and suitability support dominate. Hebbia's finance focus and traceability^3,4 are directly relevant, as are its expert call and filings sources^151,58. Glean's permission enforcement¹⁴ helps where research and client teams must be kept apart. Nexus fits a firm that wants to automate a defined role end to end with a proof of concept it can exit⁴⁴.

Payments firms and fintechs

Speed matters, and so does card data. Kore.ai and Gemini Enterprise both list PCI DSS^131,142, and WRITER lists PCI among the frameworks it addresses¹⁵². A fintech with its own engineers may favour a self service build on Gemini Enterprise or Copilot Studio. One without spare engineers should look at Nexus, which states a first agent live in 4 weeks¹.

Brokers and advisory firms

Smaller regulated firms need a clear price and limited internal effort. Gemini Enterprise Business starts at 21 dollars per seat per month for teams up to 300^147,153. WRITER Starter has a 14 day trial¹⁴⁸. For agents that act on customer records, a smaller firm should still ask each vendor how human approval works before choosing.

Due diligence checklist

Use this list in the request for proposal and the proof of concept. Each item should be answered in writing and demonstrated where possible.

1. Name the people who will design, build and test the first agent, and state their experience. 2. Give a date for the first agent in production on one named use case. 3. State the KPI each agent will be measured against, and what happens if it does not move. 4. Show the release gate: who signs off, on what evidence, and how publishing can be blocked. 5. Show an escalation to a human, including what context the human receives. 6. Show a full trace of one agent run, from input to action, as a reviewer would see it. 7. Show how an agent is limited to approved knowledge and what happens when no source supports an answer. 8. Provide the SOC 2 report and ISO certificates, including ISO 42001 where held. 9. State where data is stored and processed, which sub processors are used, and whether UK processing is available. 10. Confirm in the contract that firm data is not used to train models. 11. Explain how permissions in source systems carry through to answers and actions. 12. State the price basis, the cost of the proof of concept and the terms of exit. 13. Describe the rollback process for a faulty agent version. 14. Confirm how the contract allocates regulatory obligations between the firm and the supplier.

Conclusion

An FCA-regulated firm choosing an agentic AI platform is choosing a way of working as much as a piece of software. The platforms in this assessment differ less in what their agents can do than in who builds them, how their behaviour is evidenced, and on what terms the firm can try and leave.

Nexus is placed first because its company-confirmed model puts a named team of experienced consultants alongside the firm¹, ties each agent to a KPI³², refuses to ship without sign off on data flow, access and audit², escalates with context³⁹, holds ISO 42001 and SOC 2 Type 2^41,1, and starts with a 3 month proof of concept the firm can walk away from, priced per role automated^44,1. Its gaps on hosting, single sign on and the regulatory wording of its older terms are specific and should be closed in writing.

Hebbia offers the strongest finance focus and data controls. Copilot Studio and Gemini Enterprise offer the deepest self service governance for firms with platform teams. Glean leads on permission enforcement, WRITER on data handling for marketing and revenue work, and Kore.ai on deployment choice and engine level policy. Each is a credible option for a particular firm.

Apply the checklist to each shortlisted vendor using the same customer journey. The firm's own trial, approval process and written contract should settle the choice, with unresolved questions recorded before any customer-facing release.