Almost every bank and insurer has put its AI in the same place: a chat bubble on the help screen, behind a link marked Support. It answers questions about the product, and it does not open the account, chase the document that is holding the account up, or reach out when a payment is about to fail.

That placement was a decision, and most of the market made it without noticing. Support had the clearest cost per ticket, so support is where AI went first, and what got built is an agent that only exists after something has gone wrong, on the one channel the company picked for it.

Buyers are now scoping 2027 budgets. The question in those rooms has moved past whether AI can close a ticket and on to whether the thing they buy this year can hold a customer relationship.

Where the widget stops

A help center agent is scoped to a moment: a customer arrives with a question, the agent answers, the session closes and the context dies with it.

Most of what is expensive about a financial services relationship sits outside that moment. An application stalls because a passport photo failed an authenticity check, or a card expires two weeks before a renewal, or nobody notices a failed direct debit until the policy has lapsed. A claim gets filed on the web, chased by phone, documented by photo over SMS and settled by email.

None of those are support tickets, and every one of them decides whether the customer is still a customer next year.

The lifecycle is the unit

The arc is longer than one moment: acquisition, onboarding and verification, in-life servicing, collections and hardship, renewal, win-back.

A concierge covers that arc, working before the customer is onboarded and after they have gone quiet, initiating as often as it responds, and it is judged on whether the customer got where they were going.

That changes what you are buying. Deflection measures the old scope, and a product designed around it will keep optimizing for the moment instead of the arc.

One concierge, every channel

Channel is where most lifecycle work actually fails. A customer who ignores three onboarding emails will answer an SMS, or the person who will not upload a document will read the number off it on a call, or a policyholder at the roadside calls, sends damage photos by text an hour later, then checks status in chat the next day.

In most deployments each of those is a separate system with a separate memory, so the customer restates their date of birth at every handover and the company ends up holding three fragments of one story.

Native omnichannel means the concierge is trained once and deployed into chat, SMS and voice, with shared memory keeping the conversation in sync as the customer moves between them. The customer picks the channel and the context follows them.

From tickets to goals

A ticket-scoped agent waits for contact. Give a concierge a goal (get this customer verified, get this account current, keep this policy in force) and it works inbound and outbound until it gets there, or until a human has to rule on risk.

KYC is the clearest case. Verification rarely fails on judgment, it fails on chasing: the utility bill that never arrived, the photo that failed the authenticity check, the customer who went quiet on day three. That work has a defined goal, no fixed number of steps, and no reason to wait for the customer to open a chat window.

This is the direction Lorikeet is building in, and it is worth being precise about what ships today. Omnichannel deployment and shared memory across channels are live now. Handing a concierge a goal and letting it work a case independently is what we are building toward.

Enough AI, no more

Widening the scope raises the stakes. A concierge that can move money, ship a card, cancel an account or state a coverage position holds actions no company can hand to a model on trust.

The obvious response is to narrow the concierge back to answering questions, but the more useful one is to be deliberate about which parts of the work are allowed to be probabilistic.

Reading an unstructured message, recognizing distress, choosing words, adapting a question, holding context across a three-week claim: that is what AI does that deterministic software cannot. Disclosure text, eligibility rules, escalation triggers, contact-time windows and the boundary of what the concierge may decide should be rules the model cannot reason its way around, because those are the parts a regulator scores.

Use AI where flexibility helps the customer, and deterministic mechanisms where correctness is legally load-bearing. The first half on its own is pleasant and non-compliant, and the second on its own is the workflow engine the industry already has.

What focus buys you

None of this generalizes cleanly. The lifecycle of a lender is not the lifecycle of a SaaS vendor, and the deterministic boundary in a claims workflow is not the one in an ecommerce return.

A vendor serving banks, insurers and healthcare companies alongside businesses selling apparel is building a product that averages those requirements, and I think averages are the wrong shape for a regulated arc, where the expensive failures are specific ones.

The other half is rate of change: this market resets roughly every two quarters, so the useful evaluation question is what each vendor shipped in the last six months, who asked for it, and what that says about where they are heading. Trajectory outlasts a feature matrix, because the matrix expires.

Where Lorikeet fits

We build AI concierges for complex, regulated businesses. A concierge is trained once and deployed into chat, SMS and voice, with memory keeping conversations in sync as a customer moves between them. Team of Agents lets a concierge dispatch other agents outbound to third parties, calling a merchant, emailing a vendor, texting a delivery partner, while it keeps the customer conversation live.

The deterministic half is a product decision rather than a prompt. Guardrails are enforced at the platform level, so what a concierge may decide and what it may do are bounded by configuration.

Carmoola, a UK car finance lender regulated by the FCA, runs an agent called Katie across WhatsApp, chat and email. It resolves 60% of inbound conversations end to end, and 90% of outbound ones. The second number is the lifecycle number, because that is work the customer never asked anyone to do.

The market spent 2025 arguing about whether an AI agent could resolve a ticket, and that argument is over. The 2027 budgets are being written against the whole arc.