Take a disputed card charge from a merchant the customer does not recognize, opened in chat late on a Sunday night. Closing it means pulling the transaction history, cross-referencing the merchant ID against known DBA names, identifying it as a subscription the customer signed up for months ago, surfacing the original authorization, and then either canceling the subscription or filing a formal dispute. Every one of those steps writes to a different system, and the customer's problem is only solved when all of them are done inside the same session.

Resolution is a closed loop, where the customer's problem goes from open to solved in a single session with no ticket number standing in for an answer.

Most AI in fintech summarizes policies and routes tickets instead. The gap between what vendors claim and what actually gets resolved end to end is wide, and it matters because these are customers dealing with their own money.

What resolution actually means

Resolution means the customer does not need to come back. Applied properly, that test covers the whole customer lifecycle and every channel. It includes the onboarding someone abandoned, the payment about to fail, and the renewal nobody has raised yet, handled in chat or over SMS or on a call with the same context carried across all three.

Most CX metrics do not separate deflection from resolution. A chatbot that says "your dispute has been submitted" and closes the conversation counts as a deflection win, while the customer who follows up three days later asking what happened counts as a new conversation. The numbers improve while the experience gets worse.

True resolution in fintech means the concierge completed every step a human would have completed, pulling data from core banking or payment systems, applying business logic, executing the action, confirming the outcome, and updating the record. If any of those steps gets skipped or handed off, it is partial automation at best.

The use cases

Dispute resolution

Dispute resolution is where AI earns its keep or exposes its limitations. The dispute arrives wherever the customer already is, in chat, in email or in the app. A capable AI agent receives the dispute, pulls the transaction from the payment processor, checks the merchant category code and authorization details, determines if it qualifies under the card network's reason codes, and either resolves it immediately (clear DBA mismatch, duplicate charge) or files the chargeback with the correct documentation attached.

It touches the core ledger, the payment processor API, the card network's dispute portal, and the CRM. A successful resolution means the customer has a provisional credit, the dispute is filed correctly and the audit trail is complete, with no human review on straightforward cases.

Payment processing failures

Payment processing failures follow a similar pattern. When a customer says a payment did not go through, the AI concierge checks the transaction status in the payment gateway and identifies the failure reason (insufficient funds, expired card, processor timeout, fraud hold). It then retries the timeout, prompts the customer to update the expired card and resubmit, or escalates the fraud hold to the fraud team with the context already attached. The concierge fixes the failure, or routes it so the next step happens fast.

Identity verification

Identity verification is more nuanced. The concierge handles the mechanical parts well, checking document uploads against OCR output, validating that the name matches the account, flagging expired documents, and requesting resubmission with specific instructions ("your passport photo is too blurry, please retake in direct light"). Ambiguous cases still need human judgment, such as a name transliterated differently across documents, or an address that does not match because the customer moved last week.

The concierge still does the mechanical preparation, so the reviewer opens a case that is already assembled.

Account onboarding

Account onboarding is one of the highest-value automation targets in fintech because every drop-off is lost revenue. The concierge walks the customer through KYC requirements, collects and validates documents in real time, runs sanctions and PEP screening via API, and either approves the account or explains exactly what is missing. A good concierge finishes onboarding in the same session rather than sending the customer away with a checklist and hoping they come back.

Proactive outreach

Not every case starts with the customer opening a conversation. A card that expires before the next billing date, an onboarding abandoned at the document upload step and a direct debit heading for a Friday failure are all visible in the data days before they turn into support tickets. All three can be worked outbound over SMS or email.

A standing outcome to pursue is where this is heading rather than something to buy today. The goal would be getting an account current or getting a customer through KYC, worked inbound and outbound until it is reached. For now, the practical version is a defined outbound workflow that opens a real conversation instead of firing a templated message, with the context carried across channels so the customer can answer in chat, over SMS or on a call without starting again.

Hardship assessment

Hardship assessment starts with collecting financial information, asking the right questions without being intrusive, calculating repayment scenarios and documenting everything for compliance. Hardship criteria differ by jurisdiction, so the concierge has to apply the rules for the market the customer sits in rather than one global policy. In Australia that means ASIC's hardship obligations.

The concierge can handle the intake and the initial assessment. Final approval of a hardship arrangement usually needs human sign-off, both for regulatory reasons and because these are sensitive conversations.

Payment arrangements

Payment arrangements automate cleanly. When a customer cannot pay their full balance, the concierge checks account status, calculates the arrangement options the business rules allow (days overdue, previous arrangement history, balance thresholds), presents them, and locks in the one the customer picks. It then updates the billing system, adjusts the collections workflow and sends confirmation, all of it logged.

Where AI still falls short

Regulatory judgment calls are the clearest boundary. When a transaction triggers a suspicious activity report, a human needs to make that call. AI can flag it, gather the evidence, and pre-populate the report, but the decision carries legal liability that requires human accountability.

A customer alleging they were mis-sold a financial product needs a human who can investigate, exercise judgment, and take responsibility for the outcome. Complex complaints that span multiple products or involve potential misconduct don't resolve well through automation.

Emotional escalations are their own category. A customer whose account was frozen the day before rent is due wants someone who understands the stakes. The concierge can detect the register and escalate fast, but trying to close these cases end to end creates more damage than it prevents.

A single transaction that touches AML rules in three jurisdictions requires expertise that current AI systems don't reliably have, and the cost of getting it wrong is too high to automate.

Deflection in disguise

One test separates resolution from deflection. Pull 100 conversations your AI marked as resolved and count how many of those customers came back within seven days about the same issue. If a material share of them did, the AI was closing conversations rather than solving problems.

An agent that cannot tell you which backend systems it wrote to during a conversation did not resolve anything. Answering a question without executing an action is self-service search, and collecting information to open a ticket for a human is triage. Both are useful and neither is resolution.

The handoff is the other place to look. When the concierge cannot finish a case, the customer should not have to repeat anything they have already said on any channel, because the context travels with the case rather than living in the conversation that started it.

If a customer could have gotten the same outcome from a well-organized FAQ, the AI is an expensive search bar.

Automation readiness

Rules-based, API-driven work automates cleanly, and work carrying judgment, empathy or regulatory liability needs a person on it. The middle of that range is where deterministic configuration earns its place, because a fixed workflow can hold the rule that must never bend while the concierge handles everything around it.

Where this lands

The fintechs getting real value from AI in CX are deploying AI agents that finish the case inside the session. Carmoola resolves 60% of inbound conversations and 90% of outbound conversations end to end, in UK consumer credit regulated by the Financial Conduct Authority, where an answer about affordability or a missed repayment is a decision, not a lookup.

That depth is the part a generalist vendor cannot copy quickly, because it means learning one industry's systems and rules well enough to write into them safely.

Lorikeet only charges for successfully resolved tickets. If you're unhappy with how Lorikeet handled a ticket, you don't pay for that ticket. That commercial model only works if the concierge closes the loop.

Full audit trails are a compliance requirement in regulated industries. Every action the concierge takes, every system it queries and every decision point is logged, and we support audit trail exports for regulatory reporting.

Certifications are table stakes in this market. Lorikeet is SOC 2 Type II audited, ISO 27001:2022 certified and GDPR attested, all published on our public trust center.

The gap between current AI hype and actual resolution capability is closing, but it's not closed. The vendors who are honest about what falls in the "human required" column are the ones building systems you can trust with your customers' money.