AX Field Notes

All casesNo. 075

Kogan.com's Care Team Builds Its Own AI Agent

A forms-and-email help desk added an agent one use case at a time, in order of volume

Case no.
075
Primary source
Salesforce blog, trade press, and ASX filing
Source published
2026-07-10

1. Why this case

Kogan.com is one of Australia’s largest online-only retailers. It sells televisions, appliances, mobile plans, and insurance on competitive pricing and high volume, so customer questions “arrive in the thousands,” and Black Friday and Boxing Day push volume higher still. Before the agent described here, according to Salesforce, its customer service ran through web forms and email, with no live chat and no inbound phone line.

In March 2026 Kogan.com put an AI agent built on Salesforce’s Agentforce in front of customers. Three things make the case worth studying.

  1. The roadmap was a ranking of contact reasons. The first use case was order tracking because “where’s my order?” questions were about 60% of inbound requests. Each later use case was picked the same way.
  2. The customer care team builds and runs the agent. CTO Goran Stefkovski says the care team “build and own the agent.” Customer care analysts, not the central engineering team, build the flows, while engineering looks after security, authentication, and scale.
  3. Scripts only where a missed step has consequences. A seven-step TV return runs as a fixed script. Order questions, FAQs, and product recommendations are left to the language model.

The results come from Salesforce and from Kogan.com’s CTO, speaking at and around Salesforce’s Dreamforce conference. According to Salesforce’s July 2026 blog post, the agent handles about 50% of total customer inquiry volume, and Kogan.com’s own “true resolution” metric rose from 6.2% at launch to about 20%. Over the past year, the post says, Kogan.com absorbed a 20% increase in sales volume with a 10% drop in customer servicing costs. These are vendor-published and company-reported figures, not audited results.


2. Problem

Slow channels. The Salesforce post says: “Before Agentforce, Kogan.com’s entire customer service operation ran through web forms and email, with no live chat or inbound phone line. This resulted in significant delays in customer response.” Marketing-Interactive adds self-service to the list and says these processes “worked, but required customers to find information, submit requests and wait for a response.”

Predictable questions at peak volume. Most inquiries were predictable, such as order status or how to make a return. In the post’s words, “predictable doesn’t mean easy to scale, especially when Black Friday and Boxing Day send volume through the roof.”

Peaks were met by hiring. Marketing-Interactive reports that meeting seasonal demand traditionally meant recruiting more staff, often offshore, and spending weeks on training. Stefkovski: “You can scale agents over Christmas in 30 seconds, but to scale people, your call centre team, you need weeks of training, lead time to hire people, training time to get them to understand the processes.”

The stated goal was “to handle that volume without adding headcount every time sales spiked.”

A weak first bot. The trade press reports that a bot was not new for Kogan. iTnews writes that a deterministic bot was deployed first, built with “a technology that was available to us,” and resolved about six percent of inbound inquiries. Marketing-Interactive describes “the company’s initial chatbot” as handling “little more than” order-status questions, with a resolution rate of 6.2%. Stefkovski did not treat this as a reason to stop: “You’ve got to be comfortable with a starting point.”

Confirmed facts: service via web forms and email only, no live chat or inbound phone (Salesforce); order-status questions about 60% of inbound requests (Salesforce); peak demand traditionally met with more staff, often offshore, and weeks of training (Marketing-Interactive); goal of not adding headcount at every sales spike (Salesforce); an earlier bot resolved about 6% (iTnews) or 6.2% (Marketing-Interactive).


3. Solution

3.1 From a hack day to production

Kogan.com rebuilt the bot on Agentforce with Salesforce’s forward-deployed engineers. According to Marketing-Interactive, the customer care team and those engineers held a hack day, had a working prototype by the end of it, and went live six weeks later. iTnews reports the same six weeks and lists four Agentforce services in use: NextGen Authoring, Agent Script, Observability, and Testing Center. Kogan.com’s FY26 results announcement to the ASX dates the deployment: “AI (artificial intelligence) was deployed across Kogan.com’s Customer Care operations in March 2026”.

3.2 Use cases in order of contact volume

The first use case was order management, because order-status questions were roughly 60% of inbound requests. It covers status lookups, cancellations, and delivery address changes. iTnews says the order status comes “directly from SAP.” Customers who are not logged in are sent to an authentication subagent. In the post’s words: “That subagent has exactly one job: get the customer logged in.” Afterward the customer returns to where they were.

Returns came next. Membership, product recommendations, and TV troubleshooting were also added. Salesforce calls the roadmap “essentially a ranked list of contact frequency.” Product recommendations use Kogan.com’s own search and product detail APIs; iTnews describes a product concierge giving recommendations “across 16 million SKUs.” The TV troubleshooting subagent draws on Salesforce Knowledge, “using the same knowledge articles the customer care team already maintains.”

By the October 2026 iTnews report, the agent handled 10 customer service use cases. The ones iTnews names include order status, the product concierge, membership, mobile voucher redemption, unknown charge investigations, troubleshooting, identity verification, and returns and warranty.

3.3 A fixed script for returns, the model for the rest

The TV return flow “runs through seven enforced steps: order lookup, marketplace eligibility check, duplicate case detection, a troubleshooting interlude if the product issue is unconfirmed, photo upload, return eligibility determination, and case creation.” Agent Script, Salesforce’s scripting language for deterministic agent control, “enforces each one in sequence, holds state across the full session, and if a customer abandons mid-flow and comes back, they pick up where they left off.” If a customer drops the return to ask something else, an exit protocol clears the return’s state and sends them back to the main support topic without ending the session.

Other flows do not use a script: “Conversational flows — order management, FAQ, product recommendations — lean on the LLM’s reasoning.” The team learned this by overdoing it. Its earliest Agent Script builds forced if/else branching onto flows that did not need it, and the rigid trees were harder to maintain and broke when customers went off-script. The rule it settled on: “use Agent Script where the sequence is non-negotiable and a missed step has real consequences. Leave everything else to the LLM. Product returns need determinism. Order status lookups don’t.”

3.4 Photos as evidence in returns

For damage in transit, Kogan.com asks customers for photos. iTnews describes an emerging use case in which a generative model reads those images instead of a trained image-recognition model. Stefkovski said this also catches customers who submit a picture of a damaged product found online: “[The AI] can detect that multiple people have submitted [the same] image for this case or for different cases, and so we’ve actually been able to eliminate that type of aspect of nefarious behaviour that sometimes happens.”

3.5 Handoff to people

When the agent cannot resolve an inquiry, it routes the customer to a human representative. “The system checks live queue availability and if the wait exceeds 15 minutes, it offers customers the option to create a case and receive a callback. On every handoff, the representative receives a conversation summary before picking up.” Marketing-Interactive reports that the local team continues to handle complex enquiries.

3.6 Who builds the agent

Marketing-Interactive reports that “Kogan’s customer care analysts, rather than its central engineering team, have been instrumental in building the system.” They had already been trained on Salesforce and now work directly with Salesforce’s engineers. Engineering teams remain “responsible for security, authentication and scale.” Stefkovski, in the Salesforce post: “What matters most to me is that the customer care team can own this channel. We’ve been able to upskill our customer care team and they’ve been able to deploy this use case by use case, grow its capability and steer it inline with our customer’s needs.”

Changes are tested before they go out. Salesforce credits the rise in resolution to “the A/B testing discipline the team applied at every release, the shift to deterministic flows for complex use cases, and continuous tuning anchored to real transcript data.”

3.7 Paying for it step by step

The investment was not a large upfront program. Stefkovski told Marketing-Interactive: “The investment was incremental. When we did our shift of platforms, we wanted a bot, so we did an RFP and we looked at all different solutions, and the bot was already included in that.” The company tracks token spending, agent actions, and operating costs as usage grows. No total investment or dollar return was disclosed.

Confirmed facts: hack day prototype, live about six weeks later (Marketing-Interactive, iTnews); AI deployed in Customer Care in March 2026 (ASX); four Agentforce services (iTnews); first use case order management, order status from SAP; authentication subagent; seven-step scripted TV return with exit protocol and resume; LLM for order management, FAQ, recommendations; Knowledge articles maintained by the care team; 10 use cases by October 2026 (iTnews); image check for transit damage and reused photos (iTnews); human handoff with a 15-minute callback threshold and a conversation summary; care analysts build, engineering keeps security, authentication, and scale; A/B testing at every release; incremental investment.


4. Impact

Figures come from three sources with different dates.

Salesforce blog, 2026-07-10

MetricFigureBasis
Share of total customer inquiry volume handled by the agentabout 50%At time of post
“True resolution” (question fully answered, no follow-up contact)6.2% → about 20%Launch in March 2026 → time of post
Cases processed automaticallyabout 5,000 per monthAt time of post
Agent actions36,850 per weekAt time of post
Customer preference for the agent50% → 78%Since deployment began
CSAT, agent-handled vs. human-handledwithin 4 percentage pointsAt time of post
Sales volume / customer servicing costs+20% / −10%“Over the past year”

Kogan.com validates the true resolution rate monthly: it “runs its own extraction and cross-checks the figure against Salesforce’s number.” Salesforce also calls Kogan.com “the highest-consuming Agentforce customer in the ANZ region,” which is the vendor’s own claim. The post’s headline says Kogan.com “Automated 67% of Customer Inquiries”; the body gives about 50% handled and about 20% true resolution.

Marketing-Interactive, 2026-09-22 (Stefkovski’s Dreamforce talk and interview)

  • Figures presented at Dreamforce: 67% of conversations resolved with AI tools, “compared with the original bot’s 6.2% resolution rate”; 78% of customers say they prefer interacting with an agent; more than 44,000 agent actions during a peak week. The definition of “resolved” is not given; it is a different measure from true resolution.
  • In the same article, 81% of AI-handled customer conversations are not escalated to humans at all.
  • Stefkovski said customer satisfaction scores are seven percentage points higher among customers using the bot. This is a different comparison from the Salesforce post’s four-point figure.

Kogan.com FY26 results announcement to the ASX, 2026-08-24

The company says the March 2026 deployment in Customer Care is “automating a significant proportion of customer enquiries, considerably improving resolution times, and achieving reduced serving costs.” It gives no figures for customer care specifically.

Staffing. Stefkovski told Marketing-Interactive: “We’re seeing an increase in sales volumes, but it’s actually getting cheaper for us to serve that because we don’t need to scale the headcount.” The publication reports that Kogan.com is “already reconsidering how much seasonal recruitment it needs,” and that “employees previously responsible for repetitive work are increasingly involved in developing and improving the agents.” No headcount figures were published.

Speed and reach. Marketing-Interactive reports that the time to deploy agents has fallen to two weeks when adapting the same workflows to other brands, systems, and loyalty programs, and that the architecture is being taken to Mighty Ape, the group’s New Zealand retailer. The Salesforce post says Agentforce Voice is in testing as “Kogan.com’s first-ever inbound voice channel,” and that automated returns are expanding beyond TVs to robot vacuums and portable air conditioners, which “account for roughly 7% of remaining case volume.”


5. Insight

The contact log chose the roadmap. Kogan.com did not start from what an agent could do. It started from what customers asked most, built that, and moved down the list. Order status at about 60% of volume came first; returns for robot vacuums and portable air conditioners are next because they are about 7% of what remains. Each step is sized by the questions it takes off the queue.

Determinism is spent only where errors have consequences. A return involves eligibility, duplicate claims, photos, and a case that a person will act on, so every step is forced in order. An order-status question is not, so the model handles it freely. The team’s first builds scripted too much, and the lesson was that over-scripting breaks when customers go off-script. The split follows the cost of an error: scripts where a missed step has real consequences, the model wherever a wrong turn can be corrected in the next message.

The work of the care team changed from answering to building. The people who know which questions come in and where answers fall short now write the flows, maintain the knowledge articles, and decide the next use case. Engineering stays behind them on security, authentication, and scale. Stefkovski described the shift as moving “from a repetitive style of work to a more innovative style of work of okay, let’s build flows and let’s identify where new customer challenges are coming through.”

A strict metric and a low start. Kogan.com launched while its own measure stood at 6.2%, and it chose a measure that counts a follow-up contact as failure. The bigger numbers in circulation, such as 67% of AI-handled conversations resolved, measure something else. A number that counts repeat contacts, checked monthly against the vendor’s figure, gives each release a test it can fail.

Peak capacity moves from hiring to the agent. Before, a sales spike meant recruiting and weeks of training. Now the company reports higher sales volume served at lower cost without scaling headcount, and is reconsidering seasonal recruitment.

Limits. The detailed account is a Salesforce blog post written by the head of Salesforce’s Forward Deployed Engineering team for its ANZ Commercial segment, and the other figures are the company’s own. The post’s headline (67%) and body (about 50% handled, about 20% true resolution) do not match. The sources describe the 6% starting point differently: Salesforce dates it to the March 2026 launch, while the trade press attributes it to an earlier bot that was then rebuilt. The two CSAT comparisons use different baselines and give different gaps: within four points of human-handled, and seven points higher among bot users. The 10% cost reduction covers a year that began before the March 2026 launch, so not all of it can be credited to the agent. No headcount, investment, or per-contact cost figures were published. The image check is described as emerging, and voice is still in testing.

One-line takeaway: Kogan.com built its service agent by ranking contact reasons, scripting only the steps where a mistake costs money, and handing the build to the care team that used to answer the questions.


6. Implementation hypotheses (not in the source)

6.1 The TV return as a stored sequence of checks

Basis: The seven enforced steps, Agent Script that “holds state across the full session,” an exit protocol that “clears the sub-topic’s state,” and an image check that detects the same photo submitted “for this case or for different cases.”

Each of the seven steps likely writes one field to the session: order ID, marketplace flag, any existing case ID, troubleshooting outcome, photo references, and eligibility result. The script would advance only when the previous field is set, which is what makes the order enforceable rather than suggested. The marketplace eligibility check probably separates items Kogan.com sells itself from items sold by third-party sellers on its marketplace, which would follow a different return path. Duplicate case detection would query open cases for the same order and product before anything is created, so a customer who comes back to the chat does not open a second claim. The troubleshooting interlude likely calls the same Knowledge-grounded troubleshooting subagent and ends the return without a case if the fix works. At the photo step, a multimodal model would describe the damage and compare the image against photos already attached to other cases. The eligibility decision itself is more likely a rule over the order date, warranty, and damage type than a model judgment, since this is the flow where “a missed step has real consequences.” The case created at the end would carry every collected field and photo, so the person who handles it does not have to ask the customer again.

6.2 A release loop the care team can run

Basis: iTnews lists NextGen Authoring, Agent Script, Observability, and Testing Center. Salesforce credits gains to “A/B testing discipline the team applied at every release” and “continuous tuning anchored to real transcript data.” Marketing-Interactive says engineering teams are “responsible for security, authentication and scale.” The time to adapt workflows to another brand is two weeks.

The likely division is that engineering exposes a fixed set of actions, such as SAP order status, cancellation, address change, search and product detail, case creation, and login, with permission checks inside each action. Care analysts would compose topics, instructions, and scripts on top of those actions in the authoring tool, without touching credentials or system connections. A release probably runs in four steps. Analysts read transcripts in the observability view to find where conversations fail or lead to a follow-up contact. They change an instruction, a script step, or a knowledge article. They run the change against stored test conversations in Testing Center. Then they route part of live traffic to the new version and compare its resolution and escalation rates with the old one before switching over. The two-week adaptation for other brands suggests the topics and scripts are reused, and only the action layer is pointed at the other brand’s order, membership, and loyalty systems.

6.3 Counting “true resolution” across channels

Basis: True resolution is “whether a customer’s question was fully answered without a follow-up contact,” and “Kogan.com runs its own extraction and cross-checks the figure against Salesforce’s number.” Separately, 67% of AI-handled conversations were reported as resolved.

To find a follow-up contact, Kogan.com has to link each agent session to later contacts from the same customer through every channel it runs: chat, web form, email, and callback cases. The monthly extraction likely pulls the month’s agent sessions with customer and order IDs, then searches case and email records for any new contact from that customer, probably on the same order, within a fixed window. A session followed by such a contact would count as unresolved even if the conversation itself ended cleanly. Kogan.com’s own records would include contacts that never touch the agent platform, such as emails, which is a likely reason to run its own extraction rather than rely only on the vendor’s figure. The same difference would explain why a resolved-conversation rate measured at the end of a chat can sit far above a rate that also waits to see whether the customer comes back.


7. Sources

  1. Salesforce (Tomasz Oczapowski) — How Kogan.com Automated 67% of Customer Inquiries and Tripled Resolution
    https://www.salesforce.com/blog/kogan-agentforce/
    2026-07-10 (accessed 2026-11-17)
  2. Marketing-Interactive (Matthew Eaton) — ‘90% of questions don’t need a human’: Kogan’s six-week sprint to agentic AI
    https://www.marketing-interactive.com/not-all-questions-need-human-kogan-six-week-sprint-agentic-ai
    2026-09-22 (accessed 2026-11-17)
  3. iTnews (Ry Crozier) — Kogan taps AI agent to handle customer service
    https://www.itnews.com.au/news/kogan-taps-ai-agent-to-handle-customer-service-629602
    2026-10-14 (accessed 2026-11-17)
  4. Kogan.com Ltd, ASX announcement — Kogan Group exceeds $1 billion in Gross Sales as growth accelerates and margins expand (FY26 results)
    https://announcements.asx.com.au/asxpdf/20260824/pdf/0733sjhpgwc4p8.pdf
    2026-08-24 (accessed 2026-11-17)