Home Depot's AI Win Is a Triage System, Not a Chatbot (2026)
Home Depot's AI win is ticket triage, not a smarter chatbot. Klarna cut resolution time 82% and saved $40M/year with the same pattern. A mid-market team can copy it in 90 days for under $2,000/month.
Home Depot's AI Win Isn't a Chatbot. It's a Triage System.
Everyone's Talking About the Wrong Thing
Home Depot named Franziska Bell as CTO on March 31, 2026. She starts April 6. She ran AI at Ford. Before that, Uber, BP, Toyota. PhD from Berkeley.
CEO Ted Decker said it plainly: "Her expertise will be invaluable as we invest to remove friction."
Remove friction. Not "build a cooler chatbot."
Home Depot runs 175+ AI and machine learning projects. They have 15+ petabytes of proprietary data. 80% of their software is built in-house. They partnered with Google Cloud instead of AWS specifically to avoid conflicts with Amazon's retail business.
Their Magic Apron AI suite doesn't just answer questions. It checks real-time inventory. It gives aisle-level product locations. It generates material lists from voice or text. It places orders on behalf of customers.
That's not a chatbot. That's a triage system with teeth.
Jordan Broggi, Home Depot's EVP of customer experience, put it perfectly: "Customers want to be helped in the right moment by the right tool. And if that's a human, that's a human. If that's a chatbot, that's a chatbot."
That one sentence is the entire strategy.
A Quick History Lesson: Phone Trees Tried This in 1990
Interactive Voice Response — the "press 1 for billing" system — was supposed to solve this problem 35 years ago. It didn't. IVR was dumb routing. It sorted you by topic, not by complexity.
You'd press 3 for "order status," then sit on hold for 20 minutes while a human typed your order number into the same screen you could've checked online.
The problem wasn't routing. It was capability. The automated path couldn't actually do anything. It could only sort you into a queue.
AI triage fixes this by asking a different question: not "what topic is this about?" but "can we resolve this without a human, and if not, which human needs it right now?"
That's the shift. Triage based on resolution capability, not topic classification.
The Numbers That Actually Matter
Klarna is the clearest proof this works at scale.
They handle 2.3 million customer conversations per month. AI now takes 1.5 million of those. Humans handle the remaining 800,000. Average resolution dropped from 11 minutes to 2 minutes. That's 82% faster.
Customer satisfaction for AI-handled conversations? Statistically equal to human-handled ones. For simple requests, AI scores higher because there's zero wait time.
Annual savings: $40 million. AI system cost: roughly $5-10 million/year. That's a 4-7x return.
Banco Bradesco posted a 30% cost reduction and a 22-point NPS improvement.
These aren't pilot numbers. This is production data.
The 90-Day Blueprint for a Mid-Market Team
Here's how to build an AI triage/deflection/escalation workflow without a $2.6 billion ARR platform and a CTO from Ford.
Days 1-14: Audit your tickets.
Pull every support ticket from the last 90 days. Categorize them by two criteria: complexity (can AI handle it?) and frequency (how often does it come in?). You'll find that 60-70% of your volume is repetitive and resolvable with data lookups, status checks, or standard responses. That's your deflection layer.
Days 15-30: Build the deflection layer.
Pick one tool. If you're on Zendesk, use their AI agents (they're piloting agentic models with OpenAI right now that aim for 80% automation). If you're on Intercom, use Fin. If you're rolling your own, use n8n with a retrieval-augmented generation setup pulling from your knowledge base and order system.
The AI handles: order status, returns eligibility checks, password resets, FAQ responses, basic troubleshooting.
Cost: $500-$2,000/month depending on volume and tool.
Days 31-60: Build the escalation rules.
This is where most people screw up. They build the AI layer and forget the handoff.
Your escalation rules should trigger on three conditions:
1. Sentiment drop. If the customer's tone shifts negative, route to a human immediately. 2. Confidence score. If the AI's confidence in its response falls below your threshold (start at 80%), escalate. 3. Topic match. Billing disputes, complaints about service quality, anything involving a refund above a set dollar amount — these go straight to a human.
Write these rules in plain language. Zendesk's new platform lets you author procedures in natural language and preview the AI's planned steps before activation. Other tools are catching up.
Days 61-90: Monitor, measure, adjust.
Track five numbers weekly:
| Metric | Target (Day 90) | Red Flag | |---|---|---| | Deflection rate | 40-50% | Below 25% | | CSAT on AI-handled tickets | Within 5% of human baseline | Drop of 10%+ | | Time-to-resolution (AI) | Under 3 minutes | Over 5 minutes | | Escalation rate | 15-25% | Over 40% | | Cost per contact | 30-40% lower than baseline | No change |
Start conservative. A 40% deflection rate in month one is great. Klarna didn't hit 66% overnight. They iterated.
How to Not Tank Your CSAT
This is the fear that kills most AI customer service projects. "What if customers hate it?"
The data says they won't, if you build the triage layer right.
Klarna's CSAT stayed equal to human levels. Best Buy Canada's NPS improved. Banco Bradesco gained 22 NPS points.
The pattern holds: AI handles speed and volume. Humans handle complexity and emotion. Customers don't care who solves their problem. They care that it gets solved fast.
Three rules to protect your score:
Always offer a human path. Never trap someone in an AI loop. One click to a person. Every time.
Don't pretend the AI is a person. Label it clearly. Customers who feel deceived tank your ratings. Customers who know it's AI and get a fast answer rate you higher.
Measure AI and human CSAT separately from day one. Klarna did this. It's the only way to know if the system is working or if your overall score is masking a problem.
The Real Lesson From Home Depot
Home Depot built an intelligent routing system backed by 15 petabytes of product data, real-time inventory, and agentic AI that can actually take action — check stock, locate products, generate material lists, place orders.
CIO Angie Brown said it simply: "It's creating the experience that works for a variety of different customers, across a variety of different journeys, and meeting them in the moment for what they need."
You don't need 15 petabytes. You don't need Google Cloud. You need a clear map of your ticket types, an AI layer that handles the repetitive ones, escalation rules that catch what the AI can't, and a dashboard that tells you if it's working.
StoryPros builds AI agents that take action — not chatbots that deflect questions. The triage pattern Home Depot is running is the same architecture we use for sales and support workflows. Start with the business problem. Build the routing logic. Then pick the tools.
The AI is the delivery mechanism. The strategy is the product.
FAQ
Does Home Depot use AI for customer service?
Yes. Home Depot runs 175+ AI and machine learning projects and expanded its Magic Apron AI suite through a Google Cloud partnership in January 2026. The system uses generative AI to answer product questions, check real-time local inventory with aisle-level precision, generate material lists, and place orders. Home Depot also hired Franziska Bell as CTO in April 2026 specifically to lead enterprise-wide AI integration. About 80% of the company's software is built in-house.
What is an AI triage workflow for customer support?
AI customer service triage is a routing system that evaluates incoming tickets and decides whether AI can resolve them or whether they need a human agent. Unlike old phone-tree systems that sorted by topic, AI triage sorts by resolution capability — checking sentiment, confidence scores, and topic complexity. Klarna uses this approach to handle 1.5 million of its 2.3 million monthly conversations with AI, resolving them in 2 minutes versus 11 minutes for human agents. StoryPros builds triage workflows using n8n with retrieval-augmented generation and escalation rules that trigger on sentiment shifts, low AI confidence, or high-value disputes.
How do you automate customer support with AI without hurting CSAT?
Start by deflecting only high-frequency, low-complexity tickets — order status, password resets, basic troubleshooting. Keep escalation paths to human agents available at all times. Measure AI-handled and human-handled satisfaction scores separately from day one. Klarna's AI-handled conversations achieved customer satisfaction equal to human agents, and Best Buy Canada saw a 19% reduction in average handle time with a 40% drop in call transfers after adding AI triage through Genesys Cloud. The key is conservative deflection targets (40-50% in month one) with weekly monitoring.
What tools can mid-market teams use for AI ticket triage?
Zendesk is piloting agentic AI models with OpenAI targeting 80% automation. Intercom offers Fin for AI-first support. For custom builds, n8n provides workflow automation with retrieval-augmented generation at a fraction of the cost of larger platforms. Genesys Cloud powers AI triage for brands like Best Buy Canada and Banco Bradesco. Costs range from $29/month for small teams (Tidio) to $2,000/month for mid-market teams handling 10,000+ monthly tickets (Zendesk AI). The tool matters less than the escalation logic behind it.
What deflection rate is safe for AI customer service?
A 40-50% deflection rate in the first 90 days is a strong target for mid-market teams. Klarna eventually reached 66% AI-handled volume across 2.3 million monthly conversations. Automation Anywhere's data from 70+ deployments shows 80%+ auto-resolution in IT support contexts. The safe threshold depends on your escalation rules — if CSAT drops more than 5 points below your human baseline or escalation rates exceed 40%, pull back and retune. Conservative targets with weekly monitoring protect customer satisfaction while still cutting cost per contact 30-40%.
Related Reading
How much can a mid-market team save by adding AI triage to customer support?
Klarna saves $40 million per year using AI triage across 2.3 million monthly conversations. Their AI system costs $5-10 million per year, a 4-7x return. A mid-market team can build a basic triage layer for $500-$2,000 per month.
How fast did Klarna resolve customer tickets after switching to AI triage?
Klarna cut average resolution time from 11 minutes to 2 minutes, an 82% reduction. AI handles 1.5 million of their 2.3 million monthly conversations. Customer satisfaction scores for AI-handled tickets matched human-handled ones.
What deflection rate should a team target in the first 90 days of AI customer service?
A 40-50% deflection rate in the first 90 days is a safe target. Pull back and retune if CSAT drops more than 5 points below your human baseline. Klarna reached 66% AI-handled volume over time, not on day one.