In February 2026, Deloitte UK published a note that sent tremors through the City of London. Several wealth management firms saw their share prices drop sharply - not because of poor earnings, not because of regulatory action, but because of a single AI product launch. Altruist's "Hazel" planning engine and Insurify's AI-powered comparison platform had demonstrated, in concrete terms, what the financial services industry had been theorising about for years: that autonomous agents could disintermediate the advisory relationship. The market's reaction was not to the products themselves - which were early-stage and limited in scope - but to the structural vulnerability they exposed. If an AI agent can scan every insurance policy on the market in seconds, parse the fine print that human brokers skim, and match coverage to a client's specific risk profile with actuarial precision, then what exactly is the broker's value proposition?
The answer, as this essay will argue, is that the broker's value proposition depends entirely on whether the institutions they represent have achieved AXD Readiness. Financial services is not merely another sector that will be affected by the agentic economy. It is the sector where the stakes are highest, the regulatory constraints are tightest, the trust requirements are most demanding, and the structural transformation will be most profound. Banking, insurance, and wealth management sit at the intersection of every force that the agentic economy amplifies: complex products, high-consequence decisions, regulatory compliance, identity verification, and fiduciary obligation.
This essay applies the Four Pillars of AXD Readiness - Signal Clarity, Reputation via Reliability, Intent Translation, and Engagement Architecture - to financial services specifically. It maps the structural gaps in each sub-sector, identifies the unique challenges that regulation and fiduciary duty create, and proposes a readiness roadmap for institutions that intend to survive the transition from human-mediated to agent-mediated financial commerce.
The Sector That Cannot Afford to Wait
The numbers tell a story of acceleration that should alarm every financial services executive who has not yet begun their AXD readiness journey. Robo-advisory assets under management grew from $97.54 billion in 2020 to a projected $2.06 trillion by 2025 - a twenty-fold increase in five years. The World Economic Forum projects that by 2027, AI-driven investment tools will become the primary source of advice for retail investors, with usage reaching eighty per cent by 2028. McKinsey's December 2025 analysis found that eighty-five per cent of frontline bankers are already using AI in some form, and that agentic AI could lift relationship manager productivity by three to fifteen per cent in revenue while reducing cost-to-serve by twenty to forty per cent.
But these statistics describe the internal adoption of AI by financial institutions - using agents to make their own operations more efficient. The far more consequential shift is external: the moment when the customer's agent arrives at the institution's digital front door and expects to transact. This is the shift from "we use AI" to "our customers' AI uses us." J.P. Morgan Payments captured this distinction precisely in their February 2026 analysis of agentic commerce: the next evolution of digital commerce will allow consumers to start shopping from entirely new touchpoints - not a retailer's homepage or a search engine, but through an AI agent.
For financial services, this external shift is uniquely consequential because financial products are among the most complex, most regulated, and most consequential purchases a consumer makes. A mortgage is not a pair of trainers. A life insurance policy is not a streaming subscription. A pension drawdown strategy is not a restaurant reservation. The complexity of these products has historically been the moat that protected human intermediaries - advisers, brokers, relationship managers - from disintermediation. The argument was always: "These products are too complex for algorithms. Clients need human judgment." But as the AXD Institute's analysis of the economics of trust demonstrates, the cost of maintaining human intermediaries is itself a trust tax - one that autonomous agents are beginning to undercut.
"The shift is not from 'we use AI' to 'we use better AI.' It is from 'we use AI' to 'our customers' AI uses us.' That is a fundamentally different design problem."
That argument is collapsing. Not because AI has achieved human-level judgment in financial planning - it has not, and may not for years - but because the machine customers arriving at financial institutions' digital doors do not need to replicate human judgment. They need to parse product data, compare terms, verify compliance, execute transactions, and escalate to humans when complexity exceeds their competence. They need, in other words, exactly the four capabilities that the AXD Readiness framework measures: signal clarity, reputation via reliability, intent translation, and engagement architecture. And most financial institutions are failing on all four.
Signal Clarity: When Products Speak Legal, Not Machine
Signal Clarity is the first pillar of AXD Readiness, and it is the pillar where financial services fails most spectacularly. The diagnostic question is simple: "Can an autonomous agent, encountering your product for the first time, understand what it does, what it costs, and whether it fits the customer's needs - without human interpretation?" For the vast majority of financial products, the answer is an unequivocal no.
Consider a typical retail banking product page. A current account is described in marketing language - "fee-free banking," "cashback rewards," "premium travel insurance" - accompanied by a PDF of terms and conditions that runs to forty pages of legal prose. An AI agent parsing this page encounters a thicket of ambiguity. What does "fee-free" mean? Free of monthly fees, but not overdraft fees? Free for twelve months, then £15 per month? Free if you deposit a minimum of £1,750 per month? The marketing copy does not say. The PDF does, somewhere on page twenty-three, in a subordinate clause nested within a paragraph about "applicable charges." No machine agent can reliably extract this information from unstructured legal text at scale.
The Open Banking revolution was supposed to solve this. The Payment Services Directive 2 (PSD2) and its successors mandated that banks expose account data through standardised APIs. And they did - for transaction data. But product data remains stubbornly unstructured. You can use an API to check your balance, but you cannot use an API to compare the total cost of ownership of two current accounts over three years, including all conditional fees, across twelve banks. The data exists, but it exists in PDFs, in marketing copy, in footnotes, in terms and conditions that were written by lawyers for regulators, not by engineers for machines.
"Open Banking opened the transaction layer. It left the product layer locked behind PDFs, marketing copy, and forty pages of legal prose that no machine agent can reliably parse."
The signal clarity gap in banking is not a technology problem. The technology to create machine-readable product schemas exists - JSON-LD, schema.org financial product types, OpenAPI specifications. The gap is organisational and regulatory. Product teams describe products in marketing language. Legal teams describe products in regulatory language. Neither team describes products in machine language. And until they do, autonomous agents will be unable to perform the comparative analysis that is the foundation of informed consumer choice. The institution that solves this first - that publishes its entire product catalogue in structured, machine-readable, semantically rich formats - will become the default provider for every AI agent in the market.
The Insurance Signal Problem
If banking's signal clarity problem is severe, insurance's is existential. Insurance products are, by their nature, defined by what they exclude as much as by what they include. A home insurance policy's value lies not in the headline coverage amount but in the specific exclusions, sub-limits, excesses, and conditions that determine whether a particular claim will be paid. These details are buried in policy wordings that run to hundreds of pages, written in language that is deliberately precise in legal terms and deliberately opaque in practical terms.
An AI agent tasked with finding the best home insurance for a client in a flood-risk area needs to parse not just the premium and the headline coverage, but the specific flood exclusion clauses, the excess amounts for flood claims versus other claims, the definition of "flood" used in the policy (surface water? river overflow? groundwater?), and the conditions under which flood cover can be added or removed. This information exists, but it exists in natural language prose that requires expert interpretation. The comparison sites that currently dominate insurance distribution solve this problem by employing teams of analysts who manually extract and standardise policy data. This is a human bottleneck that autonomous agents will not tolerate.
Insurify's AI-powered comparison platform, the product that rattled the London market in early 2026, demonstrated that large language models can now parse policy documents with reasonable accuracy. But "reasonable accuracy" is not sufficient for financial products where a misinterpreted exclusion clause can leave a consumer uninsured for their most significant risk. The solution is not better AI parsing of bad data. The solution is better data. Insurers must publish their policy terms in structured, machine-readable formats - not as a replacement for the legal policy wording, but as a parallel layer that agents can consume, compare, and verify.
"The solution is not better AI parsing of bad data. The solution is better data. Insurers must publish policy terms in structured formats that agents can consume, compare, and verify."
The insurance industry has a precedent for this transformation: ACORD (Association for Cooperative Operations Research and Development) standards, which have governed data exchange between insurers and intermediaries for decades. But ACORD standards were designed for business-to-business data exchange between known parties, not for open discovery by autonomous agents. The agentic economy requires a new layer - an open, machine-readable product description standard that sits above the legal policy wording and below the marketing copy, providing agents with the structured data they need to make informed comparisons without requiring them to parse legal prose.
Reputation via Reliability: Trust Beyond the Brand
Reputation via Reliability is the second pillar, and it exposes a fundamental asymmetry in how financial institutions build trust. For human customers, trust in a bank or insurer is built through brand recognition, personal relationships, physical presence, and regulatory status. A customer trusts Barclays because they have banked there for twenty years, because there is a branch on the high street, because the FCA regulates them, and because their parents banked there before them. These signals are powerful for humans. They are meaningless to machines.
An autonomous agent evaluating whether to route a client's mortgage application to Barclays or to a challenger bank does not care about brand heritage, branch networks, or television advertising. It cares about measurable, verifiable, real-time performance metrics: What is the average time from application to offer? What percentage of applications are approved at the initially quoted rate? What is the uptime of the application API? What is the average response time for status queries? What is the dispute resolution timeline? These are the metrics that constitute reputation via reliability in the agentic economy - and almost no financial institution publishes them.
The irony is that financial institutions possess this data. Every bank knows its average mortgage processing time. Every insurer knows its claims settlement ratio. Every wealth manager knows their portfolio performance against benchmark. But this data is treated as internal operational intelligence, shared reluctantly in annual reports and regulatory filings, not as a competitive signal broadcast to the market in real time. In the agentic economy, this reticence becomes a competitive disadvantage. The institution that publishes real-time reliability metrics - application processing times, claims settlement speeds, API uptime guarantees, dispute resolution timelines - gives agents a reason to prefer it over competitors who offer only brand promises.
"Every bank knows its mortgage processing time. Every insurer knows its claims ratio. But they treat this data as internal intelligence, not as a competitive signal. In the agentic economy, that reticence becomes a disadvantage."
There is a deeper structural issue. Financial services regulation requires institutions to treat customers fairly, to act in their best interests, and to provide suitable advice. These obligations are currently discharged through human processes - suitability assessments, fact-finds, advice letters. When the customer's agent arrives, these processes must be machine-readable. An agent needs to verify not just that an institution claims to act in the client's best interest, but that its track record demonstrates it. Published SLAs, real-time performance dashboards, and agent-verifiable compliance attestations are not optional enhancements. They are the minimum viable trust infrastructure for the agentic economy.
Wealth Management's Reliability Gap
Wealth management presents the most acute reliability challenge because its value proposition has historically been built on the least machine-verifiable foundation: the quality of human judgment. A wealth manager's reputation rests on their ability to understand a client's complete financial picture, to anticipate needs the client has not yet articulated, and to provide counsel that integrates financial planning with life planning. These are genuinely valuable capabilities. They are also genuinely difficult to measure, compare, and verify programmatically.
McKinsey's analysis found that relationship managers in banking spend just twenty-five to thirty per cent of their time in actual client dialogue, with the remainder consumed by administrative tasks, meeting preparation, and internal processes. Agentic AI can return ten to twelve hours per week per banker by automating these non-client-facing tasks. But this efficiency gain, while valuable, does not address the fundamental reliability question that an autonomous agent must answer: "Is this wealth manager likely to deliver better outcomes for my principal than the alternatives?"
The robo-advisory revolution provided a partial answer by making investment performance transparent and comparable. Platforms like Wealthfront, Betterment, and Nutmeg publish their portfolio performance against benchmarks, their fee structures, and their asset allocation methodologies in formats that are, if not fully machine-readable, at least standardised enough for comparison. But holistic wealth management - the kind that integrates tax planning, estate planning, pension optimisation, and risk management - remains opaque. Two wealth managers managing identical portfolios may deliver vastly different after-tax returns depending on their tax-loss harvesting strategies, their pension contribution timing, and their use of ISA allowances. None of this is visible to an autonomous agent evaluating providers.
The wealth management firms that achieve AXD Readiness will be those that find ways to make their holistic value proposition machine-verifiable without reducing it to a single performance number. This means publishing not just portfolio returns, but tax efficiency metrics, planning outcome achievement rates, client retention statistics, and regulatory compliance records - all in structured, agent-consumable formats. The firms that resist this transparency, arguing that their value cannot be reduced to metrics, will find themselves invisible to the agents that increasingly control client allocation decisions.
Intent Translation: From Goals to Products
Intent Translation is the third pillar, and it is where financial services' complexity becomes both its greatest challenge and its greatest opportunity. The diagnostic question is: "Can an autonomous agent translate a human's financial goal into the specific product or combination of products that best achieves it?" The answer requires bridging the widest semantic gap in any industry: the gap between how humans think about money and how financial products are structured.
Humans think about money in terms of goals: "I want to retire at sixty with enough income to maintain my current lifestyle." "I want to protect my family if something happens to me." "I want to buy a house in three years." Financial products are structured in terms of mechanisms: "A defined contribution pension with a 60/40 equity-bond allocation and a 0.45% annual management charge." "A decreasing term life insurance policy with a twenty-five-year term and a sum assured of £350,000." "A Lifetime ISA with a government bonus of 25% on contributions up to £4,000 per year." The translation between goals and mechanisms is the work that human financial advisers do - and it is the work that autonomous agents must learn to do if they are to serve as effective intermediaries.
The challenge is that this translation is not a simple mapping. A single goal - "retire comfortably at sixty" - may require a combination of pension contributions, ISA investments, property equity release, state pension optimisation, and tax planning that spans multiple products from multiple providers over multiple decades. The optimal combination depends on the client's current age, income, existing assets, risk tolerance, tax status, family situation, health, and a dozen other variables that interact in non-linear ways. Human advisers navigate this complexity through experience, heuristics, and iterative conversation. Autonomous agents must navigate it through structured data, parametric models, and API-accessible product specifications.
"A human says 'I want to retire comfortably at sixty.' The agent must translate this into a multi-product, multi-provider, multi-decade strategy. That translation requires structured data that most financial institutions do not yet provide."
Financial institutions that achieve intent translation readiness will be those that publish their products not just as standalone offerings but as components in a goal-achievement architecture. This means providing APIs that accept goal parameters (target retirement age, desired income, risk tolerance) and return product recommendations with projected outcomes, confidence intervals, and sensitivity analyses. It means publishing the relationships between products - how a pension interacts with an ISA, how life insurance complements income protection, how a mortgage offset account reduces total interest cost. The institution that makes its products composable and goal-addressable will capture disproportionate share of agent-mediated financial planning.
The Mortgage Intent Maze
Mortgages illustrate the intent translation challenge at its most acute. A consumer's intent is deceptively simple: "I want to buy this house." The product landscape is bewilderingly complex: fixed rates, variable rates, tracker rates, discount rates, offset mortgages, interest-only, repayment, part-and-part, help-to-buy, shared ownership, lifetime mortgages, buy-to-let, and dozens of hybrid structures. Each product type has different implications for monthly payments, total cost, flexibility, portability, and risk. The optimal choice depends on the borrower's income stability, career trajectory, plans for the property, risk appetite, and views on interest rate movements.
Today, mortgage brokers navigate this maze through a combination of sourcing systems (which search lender panels for available products), affordability calculators (which determine what the client can borrow), and professional judgment (which weighs factors that no algorithm captures - the client's job security, their plans to have children, their likelihood of relocating). The sourcing systems are already semi-automated, but they operate on structured data that lenders provide specifically for broker distribution channels. This data is not publicly accessible to autonomous agents.
For an autonomous agent to perform mortgage intent translation, it needs access to the same structured product data that broker sourcing systems use - but through open APIs rather than proprietary distribution channels. It needs to know not just the headline rate, but the arrangement fee, the early repayment charges at each year, the overpayment allowances, the portability terms, the product transfer options, and the lender's service-level commitments for application processing. It needs this data in a standardised format that allows cross-lender comparison without manual normalisation.
The mortgage market also illustrates a regulatory dimension of intent translation that is unique to financial services. The Financial Conduct Authority requires that mortgage advice be "suitable" - that is, the recommended product must be appropriate for the client's specific circumstances. When a human broker provides advice, suitability is demonstrated through a documented fact-find and a written recommendation letter. When an autonomous agent selects a mortgage, how is suitability demonstrated? Who is liable if the agent's recommendation proves unsuitable? These questions sit at the intersection of intent translation and the Trust Triangle - the three-party liability architecture that governs all agentic commerce, but which is most consequential in regulated financial services.
Engagement Architecture: Beyond the Browser
Engagement Architecture is the fourth pillar, and it is where the rubber meets the road. The diagnostic question is: "Can an autonomous agent complete a transaction with your institution end-to-end, without a human touching a browser?" For financial services, this question exposes the deepest structural gap of all - because financial transactions are, by design, the most heavily authenticated, most carefully verified, and most thoroughly documented transactions in the economy.
Consider the journey of opening a bank account. A human customer navigates to the bank's website, fills in an application form, uploads identity documents, passes a video verification check, receives a confirmation email, sets up online banking credentials, and activates their card. Each step is designed for a human operating a browser. An autonomous agent cannot take a selfie for video verification. It cannot "upload" a document in the way a human does. It cannot navigate a multi-page form that was designed for human cognition and human motor skills. The entire onboarding architecture assumes a human at the keyboard.
J.P. Morgan Payments' analysis identified this as the critical bottleneck for agentic commerce in financial services. Current iterations of agent-driven commerce rely on "web-crawling" models where agents navigate merchant websites to complete guest checkout. In retail, this works tolerably - an agent can fill in a shipping address and a card number. In financial services, it fails completely. You cannot web-crawl your way through a mortgage application, an insurance underwriting process, or a pension transfer. These processes require structured data exchange, identity verification, regulatory compliance checks, and multi-party authorisation that cannot be accomplished through browser automation.
"You cannot web-crawl your way through a mortgage application, an insurance underwriting process, or a pension transfer. Financial services needs agent-native APIs, not browser automation."
The solution is agent-native engagement architecture: APIs that are designed from the ground up for machine-to-machine interaction, not retrofitted from human-facing web interfaces. This means programmatic onboarding APIs that accept verified identity credentials from trusted identity providers. It means transaction authorisation protocols that support delegated authority - where the agent acts within a mandate granted by the human principal, and the institution can verify that mandate in real time. It means status and notification APIs that keep the agent informed of application progress without requiring it to poll a web portal. And it means dispute resolution APIs that allow agents to raise, track, and resolve complaints programmatically. The emerging standards - the Model Context Protocol, the Agent-to-Agent Protocol, and the Agent Payments Protocol - provide the foundation, but financial services institutions must build the sector-specific layers on top.
The KYA Imperative
Financial services has spent decades building Know Your Customer (KYC) infrastructure - the identity verification, anti-money-laundering, and sanctions screening processes that ensure institutions know who they are dealing with. The agentic economy demands a parallel infrastructure: Know Your Agent (KYA). When an autonomous agent arrives at a bank's API requesting to open an account on behalf of a human principal, the institution must verify not just the principal's identity but the agent's identity, its authorisation, its operational boundaries, and its compliance posture.
KYA for financial services is more demanding than KYA for retail commerce because the regulatory stakes are higher. If an agent purchases a pair of shoes without proper authorisation, the worst outcome is a return and refund. If an agent transfers pension assets without proper authorisation, the outcome could be financial devastation for the principal and regulatory sanctions for the institution. The KYA framework must therefore verify not just that the agent is who it claims to be, but that it has specific, verifiable authorisation for the specific financial transaction it is attempting to execute.
Trulioo and PayOS are among the first companies developing KYA frameworks specifically for agent-led digital transactions. Their approach extends traditional KYC principles to AI systems, assessing an agent's operational boundaries, decision-making capabilities, and the scope of authority granted by its principal. But these are early-stage efforts. The financial services industry needs a comprehensive KYA standard that is endorsed by regulators, adopted by institutions, and integrated into the agent ecosystem. Without it, financial institutions face an impossible choice: refuse to serve autonomous agents (and lose market share to institutions that do) or serve them without adequate verification (and accept unquantifiable regulatory and liability risk).
"KYC took decades to build. KYA must be built in years. The regulatory stakes are identical - but the technical complexity is an order of magnitude greater."
The insurance sector faces a particularly acute KYA challenge. Insurance is a contract of utmost good faith - the policyholder has a duty to disclose all material facts that might affect the insurer's assessment of risk. When a human buys insurance, this duty is discharged through a proposal form and, in complex cases, a conversation with an underwriter. When an agent buys insurance on behalf of a human, who is responsible for ensuring that all material facts have been disclosed? The agent, which may not have access to all relevant information? The principal, who delegated the task precisely because they did not want to deal with the details? The insurer, which accepted the application without verifying that the agent had complete information? These liability questions are not theoretical. They will arise in the first disputed claim involving an agent-purchased policy, and the industry must have answers before that moment arrives.
The Financial Services Readiness Roadmap
The path to AXD Readiness in financial services is not a single transformation but a sequenced programme of structural changes that build upon each other. Based on the Four Pillars analysis, the following roadmap reflects the dependencies and priorities specific to the sector.
| Phase | Pillar | Action | Timeline |
|---|---|---|---|
| 01 | Signal Clarity | Publish machine-readable product schemas for all retail products using JSON-LD and schema.org financial types | 0-6 months |
| 02 | Signal Clarity | Create structured policy/product comparison APIs that expose fees, terms, exclusions, and conditions in standardised formats | 3-9 months |
| 03 | Reputation | Publish real-time operational metrics: processing times, approval rates, claims ratios, API uptime, dispute resolution timelines | 3-9 months |
| 04 | Intent | Build goal-to-product mapping APIs that accept life goals and return product recommendations with projected outcomes | 6-12 months |
| 05 | Engagement | Deploy agent-native onboarding APIs with delegated identity verification and KYA integration | 6-18 months |
| 06 | Engagement | Implement machine-to-machine transaction authorisation with real-time mandate verification and graduated authority levels | 9-18 months |
| 07 | All Pillars | Establish cross-pillar integration: agent-discoverable products linked to verifiable reliability metrics, goal-addressable through intent APIs, transactable through engagement APIs | 12-24 months |
The sequencing matters. Signal Clarity comes first because it is the foundation upon which all other pillars depend. An agent cannot evaluate your reliability if it cannot understand your products. It cannot translate intent into product selection if your products are not machine-readable. And it cannot transact if it cannot first discover and compare. The institutions that begin with engagement architecture - building APIs before making their products machine-readable - will build infrastructure that no agent can find or use.
Deloitte's warning is worth repeating: "By 2030, the gap between AI-enabled players and their digitally immature competitors will widen considerably. Those who fail to invest sufficiently in data infrastructure and AI integration across their operating models risk diminishing relevance." But Deloitte's framing - AI-enabled versus digitally immature - understates the transformation. The question is not whether financial institutions use AI internally. Most already do. The question is whether they are ready to be used by AI externally. That is the AXD Readiness question, and it is the question that will determine the competitive landscape of financial services for the next decade.
"The question is not whether financial institutions use AI internally. Most already do. The question is whether they are ready to be used by AI externally. That is the AXD Readiness question."
The agentic economy will not disintermediate financial services overnight. J.P. Morgan is right that it took two decades to build the digital commerce ecosystem and that agentic commerce will evolve more rapidly but will not be at scale tomorrow. But the structural foundations must be laid now. The Principal Gap - the growing distance between customer intent and institutional response - is widening every quarter, and the institutions that do not close it will find their access layers fossilised. The institutions that begin their AXD Readiness journey today - publishing machine-readable product data, broadcasting reliability metrics, building intent translation APIs, and deploying agent-native engagement architecture - will be the institutions that capture the agentic dividend. Those that wait for standards to be finalised, for regulators to mandate change, or for competitors to move first will find themselves in the position of those UK wealth managers in February 2026: watching their market value evaporate in response to a future they should have been building for.