Grit · 9 Feb 2026 · From the week of 9 February
The Truth Behind Automation Claims in Customer Support | Cresta CEO Ping Wu
These are notes on the conversation, checked against its transcript. The episode itself has the full discussion.
In brief
Joubin Mirzadegan, a partner at Kleiner Perkins, hosts Grit and interviews Ping Wu, CEO of Cresta, which provides AI for contact centers, including AI agents and AI assistance for human agents. Ping describes three factors that limit how much of a contact center can be automated: conversation complexity, IT infrastructure readiness and customer demographics. The episode also covers Cresta's work with United Airlines, the company's funding and lean-operating approach, and Ping's path from Google to CEO. The central argument is that automation alone does not transform customer experience, and that different conversations need different AI roles: eliminate the ones caused by broken processes, automate the simple ones, and augment human agents on high-emotion calls.
For founders
- Ping Wu argued that many contact center conversations exist because of failed processes or broken products, so root-cause analysis can eliminate them.
- He said the share of conversations that can be automated depends on conversation complexity, IT infrastructure readiness and customer demographics.
- On fundraising, he said Cresta weighs new capital by how much value it can create with existing funding versus the cost of dilution, and that a high valuation is not always the right goal.
- He described running Cresta lean as an operating principle, agreeing with the host's point that extra money can inflate salaries and burn, and saying that once a company is not lean other trouble will come.
- He said the most important question for an engineering leader considering the CEO role is whether they are interested in the business side and in working with many people.
For revenue leaders
- Ping Wu said AI can handle parts of many calls, such as authentication, after-call data entry and summaries, while human agents keep the complex or high-emotion calls.
- He said QA teams typically listen to only one or two percent of calls manually, and that AI can cover far more calls in real time.
- Host Joubin Mirzadegan said vendor claims of automating 90% of support often refer only to the easier initial use cases; Ping Wu responded that automation alone is not the right solution to transform customer experience.
- He described AI agents and human-agent assistance as one system, where human interactions feed back into improving the AI agents.
- He said contact center impact is easy to measure through containment, agent attrition, customer satisfaction and revenue per conversation.
What was said 25, most useful first
Some contact center conversations should not exist and are best removed by fixing their root cause rather than by automating or augmenting them. Listen
Ping Wu said these conversations arise from failed processes, broken products or confusing statements. He said AI's role here is to give full observability across channels, find the root cause and fix it, and gave United Airlines as an example of doing this with Cresta.
“The first bucket are the conversation that shouldn't even be automated or augmented. They should not even exist.”
Listen to the episode Retention & customer success Link to this Report a problem
Human quality assurance teams often review only one or two percent of calls manually, a coverage gap that Ping Wu says AI can close. Listen
He described large contact centers with hundreds of QA staff listening to calls by hand, with results that are not real time. He said AI can cover far more calls and check whether agents follow playbooks and compliance requirements.
“they can only cover one or two percent of calls.”
Listen to the episode Retention & customer success Link to this Report a problem
Cresta is deployed on the 9,000 agents of United Airlines under a multi-year partnership. Listen
Ping Wu said Cresta provides assistance across chat and voice, including note-taking and workflow automation, as well as QA and insights. He said United has an insight-to-action team that uses the platform to find why customers call so the root causes can be fixed and calls eliminated.
“every 9 ,000 agents is deployed on Cresta.”
Listen to the episode Retention & customer success Link to this Report a problem
Ping Wu predicts that in some verticals it will take five to ten years for the benefit of current models to be fully realised, because of workflow, data and infrastructure work. Listen
He said enterprise transformation is slower than consumer adoption because it takes time to find the right use cases, redo workflows, and clean data and modernise APIs. He said the challenges are not mainly from the models.
“Probably in some verticals, it'll probably take five to 10 years to really even saturate the benefit that can bring by the current models capabilities.”
A high valuation can make it harder to hire, because candidates may question whether they are being asked to pay for perfect future execution. Listen
He said if a company already prices in flawless execution for the next few years, prospective employees may judge the equity as overpriced and prefer to bring their own stock. He framed this as a fairness question for people joining the company.
“The other one is when you hire, if you're already pricing all the hard work and flawless execution for the next few years and bring that mortgage that to today, is that fair for the people that joining?”
Listen to the episode Hiring & team building Link to this Report a problem
Cresta expanded into a multi-product platform earlier than most would, because the products share data and intelligence and make go-to-market easier. Listen
He described the products as sharing the same data and intelligence substrate, so the AI assistant and AI agent draw on the same workflows and knowledge base. When asked about timing, he placed it in the range of 10 to 30 million dollars of ARR.
“So there's just a lot of synergy and compounding value to build a multi -product platform little early.”
Listen to the episode Strategy & market Link to this Report a problem
Low-emotion conversations such as password resets or shipment tracking are ones neither the customer nor the business wants a human to handle, so they are good candidates for automation. Listen
Ping Wu described this as the second bucket of conversations, where customers just want the task completed. He said an AI agent or a website that automates the interaction is the right solution here.
“In the second bucket of the conversations are the one that neither party want to even a human, they just want to get it done like password reset or tracking my shipment.”
Listen to the episode Retention & customer success Link to this Report a problem
For high-emotion conversations, Ping Wu says AI should augment expert human agents rather than replace them. Listen
He gave examples such as a flooded house or lost property on a flight, where customers want to be heard by a person. He said AI should free agents' hands so they can be more emotionally available to the customer, and that for businesses this builds loyalty.
“So for those, we believe that the role AI should play is to really augmenting the expert humans and make them really do a good job and free up their hands so that they can be more be more emotional available to the customer.”
Listen to the episode Retention & customer success Link to this Report a problem
The share of contact center conversations that can be automated depends on three factors: conversation complexity, IT infrastructure and customer demographics. Listen
On conversation complexity, he contrasted a simple e-commerce toy store with airlines and healthcare, where some conversations last hours with older customers. On infrastructure, he said modern APIs make automation possible while systems built for humans are hard for AI to use. On demographics, he said some customers simply prefer humans.
“One is the complexity of the conversations or the complexity of the business.”
Listen to the episode Strategy & market Link to this Report a problem
For a simple e-commerce business with simple infrastructure, Ping Wu said it would probably be possible to automate 100% of conversations. Listen
He used a Shopify store selling toys online as the example of a very simple conversation and infrastructure. He said this is at the simple end of the spectrum, while airlines operating global routes sit at the complex end.
“You and I start an e -commerce website, our Shopify, and we buy toys and sale online. Very simple conversation and very simple infrastructure. We can probably automate 100 % to be very frank.”
Large companies often have agents working across many separate systems built for humans, which makes automation harder than a set of modern APIs would. Listen
He said that, as far as he understands, agents in most Fortune 500 contact centers deal with multiple systems, sometimes nine to ten. He described AI as able to use APIs for actions like shipments, refunds and payments, but finding it hard to operate systems that are only accessible through graphical user interfaces.
“if you, most of the Fortune 500 from as far as I understand, in contact centers, agents, humans have to deal with multiple systems, sometimes like nine to 10 different systems. So the rail is not there yet.”
Older customer demographics are more likely to prefer human or in-person interaction, while younger customers are more open to chatbots. Listen
He presented customer demographics as the third factor limiting automation. He said this affects how much conversation can be contained, because some customers in airline and healthcare insurance segments simply prefer humans.
“if you're dealing with the young demographics, they're probably more open to digital channel interaction, chatbot, talking to chatbot, versus much older demographics, they will potentially will really prefer in -person interactions.”
Listen to the episode Retention & customer success Link to this Report a problem
Current models still have limits in following complex instructions, handling multimodal inputs, and acting on a screen when no API is available. Listen
He said these areas need improvement before automation becomes much easier. He added that model improvements alone will not solve the problem, because tacit knowledge and tools built for humans also limit automation.
“there are still a lot of areas that need to be improved around instruction following, especially complex instructions around multimodality that understand things that just not only text.”
Context engineering is still a manual step: the right information has to be selected from the knowledge base and fed to the model at each interaction. Listen
Ping Wu said that for simpler businesses the service procedures are already in a clean knowledge base, but for others the knowledge is scattered across multiple sources and must be cleaned up first. He said voice is harder still, and tribal knowledge that is not in the knowledge base is another limit.
“still, you have to do a lot of context engineering to bring the right information at particular interaction to feeding to the model.”
AI can perhaps automate the first 20% of a call by handling authentication, and can also automate after-call work such as data entry and summaries. Listen
He gave this as a way to create value on the complex and high-emotion calls that stay with humans. He also listed answering knowledge questions directly for agents and drafting emails as ways to support the human agent.
“Maybe you can automate the first 20 % by handling authentication. All the after -call works by automatically doing the data entry and entry and then take away the summaries”
Listen to the episode Retention & customer success Link to this Report a problem
Cresta builds AI agents and human-agent assistance on one platform, where human interactions help improve the AI agents. Listen
He said the same data and AI substrate powers both co-pilot assistance and automation. Observing where AI agents fail and how humans reach resolution gives a feedback loop that he says helps with discovering workflows, simulating callers and improving the agents.
“That's the same AI that's feeding both. That has huge opportunity because it kind of formed the feedback loop that you can see where AI agent fail and then you can see how human get to resolution, right?”
The claim that AI automates 90% of support often refers only to the easier early use cases. Listen
He said that looking under the hood of vendor marketing shows the 90% figure covers the initial, easier use cases, and asked Ping Wu how far automation could go if models stopped advancing. Ping Wu responded that automation alone is not the right solution for transforming customer experience.
“But if you really dig in under the hood, it's like 90 % of the initial earlier easier use cases to solve.”
Listen to the episode Positioning & marketing Link to this Report a problem
Nearly all contact center results can be measured through containment, agent efficiency, agent attrition, customer satisfaction and revenue per conversation for sales use cases. Listen
He said this makes the contact center an easy place in the enterprise to show measurable, quantifiable impact.
“everything, almost everything's measured efficiency and the core containment and the agent efficiency attrition rate.”
Listen to the episode Metrics & finance Link to this Report a problem
Ping Wu expects enterprise AI pilots to converge on use cases with clear ROI, naming coding and contact center transformation as likely examples. Listen
He said pilots will narrow to use cases with concrete value, such as automation and AI that helps humans in sales and support. Asked whether the music will stop, he said Cresta cares less about that and focuses on what drives long-term, undeniable value.
“I think as time goes by, and people start to converge on use cases that will drive very concrete, undeniable ROI, coding is probably one of them. And then Contact Center, CX Transformations, definitely the other one.”
Listen to the episode Strategy & market Link to this Report a problem
One factor in weighing new funding is the value Cresta can create with existing funding versus the cost of capital and dilution at its valuation. Listen
He said Cresta raised a Series D of 125 million dollars the previous year, still holds much of it, and is constantly evaluating opportunities. He said multiple factors matter, one being value created with existing funding and equity versus the cost of capital.
“One is how much value we can create with the existing funding and equity value that we can create versus the cost of capital.”
Listen to the episode Fundraising & investors Link to this Report a problem
Cresta treats being lean as an operating principle, because once a company is not lean other trouble follows. Listen
Responding to host Joubin's point that a bigger round makes it easy to pay new hires more and raises burn, Ping said Cresta is very careful about that. He said Cresta removed stay from its operating principles so it reads simply lean.
“We're very careful about that and I think that one of our operating principle is lean.”
Listen to the episode Leadership & culture Link to this Report a problem
Ping Wu questions whether driving a higher valuation is always the right goal for a company. Listen
He said he is not sure a high valuation is always right and deferred to others on the trade-off calculations. He noted that a higher round means the company must grow into the price, and that a public-market comparison helps test whether the valuation is justified.
“And also, I'm not sure driving high valuation is always the right thing to do.”
Listen to the episode Fundraising & investors Link to this Report a problem
He served as interim CEO for about a quarter before being appointed, which let the board evaluate him in action. Listen
He said there was no strict interview, rather a multi-month process, which the host likened to try before you buy and a long CEO work trial. He said he had already worked with the team as interim CEO when the board decided.
“It was a try before you buy.”
Listen to the episode Leadership & culture Link to this Report a problem
People who join Cresta self-select for being excited about hard technical problems tied to real customer value, rather than technology for its own sake. Listen
Host Joubin observed that engineers often first work on interesting technical problems and later realise these do not solve a real customer problem. Ping said those who join Cresta self-select to want both hard technical problems and measurable customer impact.
“It's not just we have a hammer and looking for nails and that type of thing.”
Listen to the episode Hiring & team building Link to this Report a problem
An engineering leader considering the CEO role should ask whether they are interested in the business side and in working with many people. Listen
He said selling, seeing value end to end, dealing with partners and customers and persuading people are personality and interest questions. He said skills like these can be learned with help from mentors, advisors and board members.
“So I think the most important thing is whether you just feel you're interested in the business side business side or interacting with a lot of humans, that aspect of the business.”
Listen to the episode Leadership & culture Link to this Report a problem