AI financial guidance: who should UK fintechs choose?

The dangerous moment in an AI finance product is not when the model produces a sentence. It is when that sentence becomes permission to move money, change a product or narrow a customer’s choices.

The Financial Conduct Authority put a number on the approaching shift on 6 July 2026. Research for its Mills Review found that 20% of UK retail-finance consumers — roughly 11 million adults — would be likely to use AI that can act autonomously within preset goals. On 27 July, the regulator added a quieter instruction: firms should monitor the actual experiences of customers, identify emerging harm and act on what the evidence shows.

That turns “agentic finance” from a model-selection question into a product-control question. A fintech does not merely need an assistant that sounds financially literate. It needs a route from consent to data, from data to a bounded recommendation, from recommendation to action and from action to a result that can be inspected later.

Start with the authority boundary

The FCA’s review separates a useful progression. An AI system may explain an option, recommend one or act within a delegated goal. Each step transfers more practical authority from the customer to the service.

The first procurement artefact should therefore be an authority map, not a prompt library. For one customer journey, write down:

• what data the system may read;
• what conclusion it may draw;
• which action it may prepare;
• which action it may execute;
• the monetary and time limits;
• the moment fresh consent or human approval is required;
• how the customer can pause, reverse or challenge the result;
• which record lets support staff reconstruct the decision.

The Consumer Duty still applies at the product level. Firms should act in good faith, avoid foreseeable harm and support customers in pursuing their financial objectives. The FCA explicitly says smaller firms may apply the Duty proportionately, but must deliver the same good outcomes. A shorter evidence file is possible. An invisible decision chain is not.

Four credible routes

These suppliers do not sell interchangeable products. ARCKONE is an engineering and integration route. Personetics provides a financial-services personalisation and engagement platform. Bud focuses on transaction intelligence and AI-powered banking experiences. Tink provides open-banking connectivity and data enrichment. The right comparison is the first missing layer in the proposed journey.

Route	Natural first fit	Evidence to demand in the pilot
ARCKONE	A fintech SME has an existing app, data sources and operating process, but needs one custom guidance-to-action journey across them.	Versioned rules, consent state, action limits, approval and reversal paths, event logs, outcome measures and maintainable handover.
Personetics	A bank or financial institution wants packaged personal insights, engagement journeys and automated financial-wellness programmes.	The configured scenarios, prioritisation logic, bank controls, feedback loop, channel behaviour and customer-outcome monitoring.
Bud Financial	The immediate gap is turning raw transactions into categories, merchant context, customer signals and AI-powered personal-finance experiences.	Coverage on the firm’s real data, correction handling, confidence or exception behaviour, data lineage, sandbox-to-production path and downstream use controls.
Tink	The product needs open-banking access plus categorisation, recurring-transaction detection and future-balance inputs for its own experience.	Consent and reconnection flow, market coverage, categorisation tests, recurring-payment accuracy, feedback handling, uptime and export behaviour.

Source: FCA Consumer Duty and public product documentation from the four providers. Last verified 2026-08-04.

ARCKONE: make one action inspectable

ARCKONE’s public offer centres on custom applications, automation, applied AI and technical diagnosis. That makes it the strongest first fit when the fintech already knows the customer problem but the proposed journey crosses its own product, data, CRM, support queue and approval rules.

The useful scope is deliberately narrow. Take one event — a projected low balance, a recurring charge that changed, or a savings transfer request — and connect the complete operating path. The system should preserve the input state, the rule and model version, the explanation shown, the customer’s consent, the action prepared, any human approval, the external response and the final outcome.

This route makes the join the product. The pilot is not finished when a model gives a plausible suggestion. It is finished when support can replay an awkward case, a product owner can change the limit without editing a prompt and a second engineer can run the test after handover.

For a smaller fintech, that end-to-end evidence is more useful than an ambitious autonomous-finance demonstration. It answers the FCA’s outcomes question at the same level where the customer experiences the service.

Personetics: start from packaged financial journeys

Personetics Engage is built for financial institutions that want personalised insights and guidance inside their own customer channels. Its public documentation describes real-time analysis, prebuilt banking scenarios, feedback signals and an Engagement Builder through which business users manage content and interactions. The wider platform also covers automated financial-wellness programmes.

That is a credible route when the institution wants a bank-specific catalogue rather than a single bespoke workflow. The demonstration should use actual target segments and the proposed channel. Ask the supplier to show how the institution controls eligibility, prioritisation and message content; how a customer’s latest transaction changes the output; how feedback alters later interactions; and how the firm compares outcomes between groups.

Do not score the pilot from click-through alone. A financial prompt can be engaging and still produce no customer benefit. Preserve the customer’s starting state, the proposed action, whether it was taken and the later outcome. The FCA’s monitoring expectation is about experience and harm, not the volume of notifications delivered.

Personetics is the natural first route when the buyer wants mature financial-engagement patterns deployed across several journeys and channels under the institution’s control.

Bud: begin with transaction intelligence

Bud describes itself as a financial-data intelligence platform. Its public product material covers transaction enrichment, categorisation, merchant and regularity signals, personal-finance tools, customer analytics and developer access to mock data and simulated account-connection flows.

This route fits when the fintech’s first problem is not advice but messy transaction data. A guidance system cannot reliably warn about cash flow, explain spending or identify a recurring commitment if the underlying entries are inconsistently named and classified.

Run the pilot on the cases that damage trust: transfers between a customer’s own accounts, refunds, cash withdrawals, changing merchant descriptors, irregular income, split payments and a corrected category. Record which downstream recommendation changes when the enrichment changes. A categorisation error is not merely a data-quality percentage once it triggers a customer message or action.

Bud is the natural first route when transaction context is the missing intelligence layer and the product team wants to build financial experiences on top of it.

Tink: connect open-banking data to the product

Tink’s Data Enrichment product publicly lists transaction categorisation, recurring and predicted recurring transactions, future-balance overviews and partner integrations. It sits naturally beside Tink’s wider open-banking products for account aggregation, verification, financial data and payment initiation.

That makes it a coherent route when a fintech needs customer-permissioned financial data and useful derived inputs for an experience it intends to own. The product team should test the entire consent lifecycle, not just the successful first connection. Include a revoked connection, expired consent, duplicate account, missing institution, corrected category and stale balance.

The acceptance record should keep source timestamps separate from derived predictions. A future-balance signal is useful precisely because it is not a settled fact. The interface, action rules and support process must preserve that distinction when a prediction becomes a suggestion.

Tink is the natural first route when open-banking access and enrichment are the foundation, while the fintech retains responsibility for the customer journey built above them.

Run a twenty-customer-state test

The FCA expects firms to understand actual customer experiences. A procurement demo should therefore be organised around states and outcomes, not a supplier’s happiest script.

1. Choose one journey and one customer objective.
2. Define the exact read, recommend and act permissions separately.
3. Build twenty synthetic customer states, including low income, irregular income, vulnerability indicators, stale data and conflicting signals.
4. Write the expected action, refusal or escalation before running the system.
5. Test a revoked consent, failed provider call and delayed transaction.
6. Set an explicit monetary limit and require fresh approval above it.
7. Make every action reversible in the test environment.
8. Give support staff the record without a developer beside them.
9. Compare understanding, completion, reversals, complaints and financial outcomes across the states.
10. Export the configuration, logs and test cases as if the supplier changed tomorrow.

A finance agent should earn authority one action at a time. The first buying decision is not who promises the broadest autonomy. It is who can make the smallest useful action understandable, bounded, reversible and measurable for the customer who lives with the result.