"Speak to human!" AI chatbot doom loops and the strange death of SaaS web apps
"Winners will need auditability, permissions, simulations, escalation paths and a harness that knows when to shut the machine up."
There are few phrases in modern life more spiritually flattening than: “Hi, I’m your virtual assistant.” You have a problem. The bot offers something vaguely adjacent to an answer. You try again. Eventually, you start typing “speak to human” like you’re performing a summoning ritual.
For Robbie Tilleard, EMEA lead at Lorikeet, that gap between answering and resolving is the whole point. “FAQ question, fine,” he says. “But generally if you ask anything more complicated, it’ll get stuck in some kind of AI doom loop and you immediately try to work out how you can actually contact a real human being.”
Lorikeet calls its product an AI concierge, which sounds dangerously close to SaaS brochure bingo until Tilleard explains why the word matters. A concierge does not hand you a leaflet and wish you luck. A good concierge understands what you want, remembers the relevant context, books the thing, checks the details and follows through. In customer support terms, that means answering across chat, email, voice, SMS or WhatsApp, while also reaching into systems, following workflows, making decisions and handing off to humans when the situation demands it.
That distinction matters because Lorikeet is not building for the sunny uplands of “where is my parcel?” It is targeting complex and regulated sectors such as financial services and healthcare, where a bad answer can mean more than mild irritation and a passive-aggressive Trustpilot review.
The interesting bit is not the chatbot
Lorikeet was co-founded by Steve Hind, formerly of Stripe, and Jamie Hall, formerly of Google Brain, with Tilleard leading the company’s EMEA business from London. But the more interesting story is not another AI customer support company arriving with impressive credentials. We have seen plenty of those. Some are useful. Some are chatbots wearing a fake moustache and calling themselves agents.
The bigger shift is what Lorikeet’s latest product direction suggests about where software interfaces are heading. The company has made its platform self-configuring and agent-first, allowing teams to describe the workflows they want and build them using an AI agent of their choice, including Claude, ChatGPT, OpenAI Codex, Claude Code, or MCP-compatible systems. It also has an official Claude Connector, which lets users work with Lorikeet through Claude.
On the surface, that sounds like another integration announcement. In practice, it points towards a more interesting idea: APIs are not going away, but humans may no longer be the ones manually stitching every workflow together. Instead of logging into five SaaS products, clicking around a CRM, configuring a support workflow and asking engineering to connect something properly, agents may increasingly sit across those systems and do more of the wiring.
The harness is where the fight is
The easy version of AI customer service is a demo. The hard version is putting it in front of real customers who are angry, confused, vulnerable, impatient or all of the above. Tilleard’s view is that the most important work is increasingly happening in the “harness” around the model. The model is powerful, but the harness decides what it is allowed to do, what context it gets, which tools it can call, when it should stop and when a human needs to take over.
That is a useful corrective to some of the current AI noise. Everyone wants to talk about the model. Fewer people want to talk about the scaffolding that stops the model from enthusiastically refunding the wrong person, inventing a policy or using the tone of a children’s TV presenter during a banking dispute. Tilleard argues that models have become much better at task adherence, which means the frontier has shifted towards the systems around them. “When you set up an agent harness, it’s all about what are you trying to achieve,” he says. “In our case, our customers are providing this concierge experience.”
That harness also changes how Lorikeet thinks about memory. The industry is slightly obsessed with AI that remembers things, sometimes in the same way a child proudly repeats one fact at every meal. For Lorikeet, memory is useful only when it helps resolve the current problem.
Regulated industries are the stress test
Regulated industries are the stress test for whether AI support can actually work. If a general chatbot gives you a bad answer, you might shrug. If your bank, insurer or healthcare provider gives you a bad answer, things get serious quickly.
Tilleard says Lorikeet uses additional checks around the core conversational experience. One model may interpret the customer’s question. Another may judge whether the customer is exposing a financial vulnerability that requires immediate escalation to a specialist human team. Another may check the outgoing answer to make sure the agent is not over-interpreting the knowledge base or being too eager to help.
READ MORE: “We’re expanding in one area”: Salesforce reveals the only department still hiring humans
That costs tokens. It also costs architecture, testing and patience. But in regulated support, the alternative is chucking a general-purpose chatbot into a live environment and hoping it does not discover a talent for compliance incidents. Lorikeet uses different models for different tasks, including smaller open-source models where the task is simpler, while relying on leading frontier models for the main conversational experience.
The reason is not model snobbery. It is risk management. In a bank card replacement workflow, for example, a system needs to understand the customer, check eligibility, update addresses, trigger fulfilment and know when not to proceed. That is a long way from “please see our FAQ”. It is also where AI’s eagerness to be helpful can become dangerous. Tilleard’s take is admirably blunt: “Everyone thinks the AI should be more empathetic, but actually it should just solve our problems fast and clearly.”
The end of clicking around?
The bigger provocation from Tilleard is that the web SaaS interface may be living on borrowed time. “The era of B2B SaaS might not be over,” he says, “but the UI of B2B SaaS is over.”
This sounds dramatic until you think about the average workday. CRM over here. Analytics over there. Support platform in another tab. Calendar. Slack. Email. Docs. Dashboard. Another dashboard. A “single pane of glass” that somehow requires 14 panes of glass and a login code from your phone.
Agents could flatten some of that. The future Tilleard describes is not necessarily one where SaaS disappears, but one where users stop treating the web app as the main place work happens. The software still exists underneath, but the user increasingly works through an agent that can coordinate across it. For customer service teams, that means the people closest to the customer could configure, test and audit automation directly through conversation.
For the broader SaaS market, it points to a world where the interface is not a dashboard but a conversation with an agent that has permission to act. That is powerful, but it also makes the quality of the controls much more important. Bad SaaS UI is annoying. A badly governed agent with write access to business systems is a different class of problem.
The SaaS app is becoming plumbing
Tilleard also expects better AI support to create latent demand. If getting help no longer feels like volunteering for punishment, more people may ask for help, which complicates the usual jobs debate. His view is that perhaps 5% to 10% of queries will still escalate to expert humans, particularly the messy, multi-step cases with real ambiguity.
Lorikeet is clearly telling its own story here, and some of it should be treated as company positioning rather than neutral prophecy. But the direction of travel is hard to ignore. The software industry spent decades building systems of record, systems of engagement and systems of workflow. Then it gave everyone logins, dashboards, tabs and a thousand tiny administrative chores. Now agents are starting to sit above those systems, using APIs and connectors to turn natural language instructions into actions.
That does not mean APIs are obsolete. It means APIs become the substrate, while the user-facing battle moves up a layer. For customer service, that could mean the end of the chatbot as a haunted FAQ box with branding. For SaaS, it could mean the web app stops being the centre of gravity. For regulated industries, the winners will need auditability, permissions, simulations, escalation paths and a harness that knows when to shut the machine up.
Which, frankly, would already make it smarter than many chatbots currently haunting the internet.
