Episodes

Gareth Cummings: The Pre-Sales Dividend

Gareth Cummings, CEO of eDesk, on how the AI-led customer support platform built generative AI four years before ChatGPT, why pre-sales is the single most under-answered category in e-commerce, and what agentic commerce will do to brand trust signals once consumer AI agents start transacting on their owner’s behalf.

Apple Podcasts Spotify
 ·  The Business of Marketing  · S4 E76  · 49 min
Read the interview

"If you are not putting the customer at the heart of your business, there is going to be a challenge in the future."

The Conversation with John Horsley

Why pre-sales is the most under-answered question in e-commerce, and how eDesk built generative AI four years before ChatGPT by working on a dataset narrow enough to train on.

Gareth Cummings runs eDesk, the AI-led customer support platform that processes more than 14 million conversations a month and supports over $25 billion in annual eCommerce transactions across 5,000+ online sellers. He joined eDesk in 2018 as CTO, moved to Managing Director in November 2024, and took over as CEO in October 2025. Before eDesk he spent more than two decades in Irish software, through IBM, Accenture, Globoforce (now Workhuman), Sentenial and Brite:Bill, where his last role was Global Head of Delivery building billing software for Comcast, T-Mobile and AT&T.

In this conversation Cummings explains how eDesk shipped its first generative AI layer in October 2018 by taking advantage of a narrow e-commerce dataset where roughly ten query categories cover 70 percent of inbound volume, why pre-sales queries (30 to 60 percent of inbound support volume at a typical online retailer) are routinely unanswered because support teams are trained for post-sale problem solving, and what agentic commerce will mean for brand trust once consumer AI agents start transacting on behalf of their owners.

Key Takeaways

  • eDesk shipped generative AI in October 2018, four years before ChatGPT, by training on a narrow e-commerce dataset where ten query categories cover 70 percent of inbound volume.
  • Pre-sales queries typically account for 30 to 60 percent of inbound support at an e-commerce business. Most go unanswered because support teams are trained for post-sale problem solving.
  • Answering a pre-sales question within 15 minutes lifts the chance of conversion measurably. The gap is behavioural rather than technological.
  • Trust signals like response time and review scores are about to become machine-readable filters as consumer AI agents transact on their owner’s behalf.
  • eDesk’s AI automation is growing more than 100 percent year on year. The next frontier is resolution rate and post-automation customer satisfaction.

Topics Covered

  1. 01 How eDesk shipped generative AI in 2018 on a narrow e-commerce dataset, four years before ChatGPT
  2. 02 Why pre-sales is the most under-answered category of customer support in e-commerce
  3. 03 The fifteen-minute conversion window and why most support teams miss it
  4. 04 Customer obsession as a structural capability, not a brand slogan
  5. 05 Agentic commerce and the rise of machine-readable trust signals

Key Exchanges

05
01 How did you make the transition from CTO to Managing Director to CEO?

Even though the genesis was technical, I had commercial exposure all along the way. So when it came to transitioning from CTO to CEO, it was not a leap into the unknown.

Cummings treats the trajectory as unremarkable only in retrospect. Starting as a web developer who sat in customer meetings as a junior, he layered commercial muscle onto the technical craft at every subsequent stage, through Globoforce, Sentenial, Brite:Bill and seven years as eDesk CTO. By the time the CEO role opened up, the commercial dimension was already in place.

02 What does eDesk do for online sellers?

We bring every message and every order across every channel into a single inbox. Where’s my order? The answer is right there, in one view.

eDesk is purpose-built for e-commerce, not general customer support. Messages from Amazon seller central, Shopify, Meta, TikTok shop and email flow into a single workbench alongside order data and product catalogue. That integration is what lets an agent resolve where-is-my-order queries in one view rather than six logins.

03 Why is pre-sales the biggest growth opportunity in e-commerce customer support?

Answer the questions within 15 minutes and your chance of a conversion goes up by a significant amount. We see it time and time again.

Pre-sales queries typically account for 30 to 60 percent of inbound support at an e-commerce business. They are the highest-conversion moment in the customer lifecycle and the most consistently ignored, because most support organisations are trained and measured on post-sale problem resolution. Cummings frames pre-sales responsiveness as the single highest-return behavioural change any online retailer can make.

04 How did eDesk build generative AI before ChatGPT?

We built essentially our own LLM, on a much smaller dataset. Because we were only focused on e-commerce, that made it easier for us.

eDesk shipped its first generative AI layer in October 2018, four years before ChatGPT. The team had a structural advantage that most customer support platforms did not: a narrow e-commerce dataset where roughly ten query categories cover 70 percent of inbound volume. That constraint gave them enough signal density to train a useful language model on modest compute, at a time when general-purpose LLMs were out of reach for most product teams.

05 What will agentic commerce do to brand trust signals?

You will give the agent certain parameters. Only engage with things that have a review score of X. Only with brands who respond quick.

As consumer AI agents start transacting on behalf of their owners, trust signals that have historically been soft brand attributes, response time, review scores, support quality, become machine-readable filters. The brands that win in this world are the ones whose support behaviour is already measurable, consistent and public today. Cummings sees this as the logical extension of everything eDesk has been building since 2018.

S4 E76Season & Episode
49 minDuration
$25B+ eCommerce Transactions Supported Annually Through eDesk
2018 Year eDesk Shipped Generative AI, Four Years Before ChatGPT

"Answer within fifteen minutes. Conversion follows."

Hear Gareth on
The Business of Marketing
Season 4 Episode 76 49 min
Apple Podcasts→ Spotify→ Gareth's full profile→ Gareth on LinkedIn→
Read the full transcript
Lightly edited for readability.

John Horsley Tell us about the transition from CTO to CEO at eDesk. How did it happen?

Gareth Cummings My career genesis was technical, but customer exposure ran alongside from the very first job as a web developer. I was in customer meetings as a junior, listening to how the commercial side of the room thought. So when the CTO to Managing Director to CEO progression happened at eDesk, the commercial dimension was already in place. It was a continuation, not a new discipline.

John Horsley What does eDesk solve for online sellers?

Gareth Cummings eDesk is built specifically for e-commerce. We bring messages from marketplaces, social channels and direct storefronts, plus the order data and the product catalogue, into a single inbox. So when a customer asks where their order is, your agent answers in one view rather than six logins. That integration is the point.

John Horsley Why do pre-sales questions matter so much?

Gareth Cummings Pre-sales is often 30 to 60 percent of inbound volume. Most support teams are trained for post-sale problem solving, so pre-sales questions get neglected. We see it again and again. Answer within 15 minutes and conversion lifts by a significant amount. It is the highest-return behavioural change an e-commerce business can make.

John Horsley How did you build AI so early?

Gareth Cummings We shipped our first AI layer in October 2018. The advantage was narrow scope. E-commerce customer support has maybe ten query categories that cover 70 percent of volume. That made it feasible to train a language model on a much smaller dataset than a general customer support tool would need. We built essentially our own LLM before the term was in common use.

John Horsley Where does agentic commerce go from here?

Gareth Cummings Consumers will give their agent parameters. Only engage with brands above a review score. Only with brands who respond quickly. The trust signals that today are soft brand attributes become machine-readable filters. The brands with measurable, consistent response behaviour win. That is the logical next step of what we have been building since 2018.

Keep Reading