Joe Bryan
All work

The Home Depot

A voice assistant people don’t hang up on

Callers spent 98 seconds fighting a phone tree before asking for a person. Before LLMs, I designed and hand-wrote a voice and chat assistant that asks what they need instead. Navigation time fell to 28 seconds.

Clay figure of Joe in a headset tangled in curly phone cords, who spins free in a cartoon dust cloud and holds up one straight cord
Role
Senior Conversational Designer
Teams I worked with
  • Product
  • Engineering
  • Data Science
Platforms
Phone, chat, smart speakers
The prototype’s call screen, ready to call The Home Depot’s voice assistant
Turn your sound on and place a call. A working rebuild of the assistant on the phone, by text and in web chat, with an invented customer and recreated voices. Flip to Before to hear the old menu. Open prototype

This was before large language models. The assistant ran on Google’s voice technology, but nothing wrote its own answers. Every intent, sample phrase, prompt and fallback was written by hand, then prototyped in Voiceflow and tested with customers.

98→28s time to navigate
35M calls kept off the service desk
+185% self-service containment

The problem

  1. 01

    Callers couldn’t tell whether to talk or press

    The phone system mixed “press 1” menus with spoken prompts on top of a deep decision tree. People gave up, or kept asking for a representative.

  2. 02

    Every failed call landed on a person

    The overflow swamped the call centers and pulled store associates off the floor to answer phones, so waits grew for callers and in-store customers alike.

Decisions that mattered

01Map every path to a person

Before designing anything, I traced the phone tree. Reaching someone took up to eight menus, notices and hand-offs, and time studies put the average at 98 seconds just to navigate.

The phone tree, before

Four ways to reach a person on the 1-800 line

  • 8 steps, worst case
  • 98s average to navigate

New or existing installation project1,037,467 calls in FY18

Get help with an installation6 steps
  1. 1Installation project
  2. Name the product, or key in its 3-digit code
  3. “Flooring”
  4. Verify your zip code
  5. Routed to installation services
  6. Call may be recorded
  7. Installation agent

Online orders or site questions3,563,612 calls in FY18

Ask about an order6 steps
  1. 2Online orders
  2. Call may be recorded
  3. 2New or existing order
  4. 2Already placed
  5. 2All other inquiries
  6. Self-serve order status menu
  7. Orders agent
Reach a local store6 steps
  1. 2Online orders
  2. Call may be recorded
  3. 3Local store information
  4. Store locator
  5. Enter your zip code
  6. 3Second-tier option
  7. Local store
Ask about a credit card8 steps
  1. 2Online orders
  2. Call may be recorded
  3. 4Credit card account
  4. 1Pick one of three card types
  5. Out to Citibank
  6. 2Language option
  7. Call may be recorded
  8. Last four of your card or Social Security number
  9. Card agent
  • 2Press a key
  • Speak or key in
  • Listen and wait
  • Handed off
  • Major appliances: sent back to a different 1-800 main menu
  • Self-service and FAQs: automated instructions, then “press # for the main menu”
Four paths through the old phone tree, redrawn from my map. Call volumes are from FY18.

02Study who does it well

I reviewed eight companies’ phone and chat support. The ones customers trusted gave them control: real tasks they could finish on their own, by phone or in chat. Ours stopped at FAQs in chat.

CompanyVoice and visualVoice assistantText opt-inCustomer controlSelf-service by phoneSelf-service in chat
The Home Depot NoNoYes Medium Order status · Product availability · Store hours FAQ
Lowe’s NoNoNo Low None Order status · Availability · Installation · Store hours · Credit cards
Amazon YesYesYes High Order status Order status · Payments · Gift cards · Memberships · Subscriptions · Login and security
Walmart NoNoNo Low None Order status · Availability · Store hours
Xfinity YesNoYes Medium Callback · Payments · Paperless billing by text Change plan · Equipment · View bill · Update account · Payment
Bank of America YesYesNo High Balance · Activity · Transfer · Order checks · Dispute Schedule appointment · Activity · Payment · Insights · Balance
American Express NoYesNo Medium Travel notice · Replacement card · Transactions · Add users · Credit limit · Available credit Same as phone
Delta NoNoYes High Redeem miles · Check flight · Seat changes · Plan trip Business chat by text · Check flight · Changes and cancellations · Refunds · Baggage and seats · SkyMiles

Every company offered chat on the web and in its app. The difference was what you could actually get done.

The competitive review, transcribed.

03One intent, any channel

Online, phone and chat ran on separate platforms in separate silos, so the same question got different answers. I designed around what the customer wants, the intent, and let every channel share it.

Before
  • OnlineIts own platform, its own answers
  • PhoneIts own platform, its own answers
  • Text and chatIts own platform, its own answers
After

“I need to make a return”

  • OnlineOrder history
  • PhoneVoice assistant
  • Text and chatHomer

One intent, understood the same way everywhere

“I need to make a return,” before and after.

04Give control back

52% of customers wanted to choose how we reach them. So the assistant offers to text what they need, transfers fast to the right associate when it can’t help, and sounds like one company everywhere.

52%of customers want to choose how we reach them
  1. Let customers choose where they talk to usPhone, text or chat: the customer picks, not the org chart.
  2. Give the best experience on any channelSometimes that’s a fast transfer to the right associate, or a link to finish online.
  3. One voice persona at every touchpointThe old phone system didn’t sound like one company.
The three trust principles from the strategy.

05Use the data we already had

The Home Depot collected fourteen kinds of customer data, siloed across teams. Four were enough to skip most of the menu: the caller’s phone number finds their orders before they say a word.

What we collected

  • Service interactions
  • HD.com browsing
  • Contact center interactions
  • Device ID and cookies
  • Marketing campaign exposure
  • Third-party property data
  • Pro Xtra member number
  • Tender and card number
  • Email address
  • Loyalty memberships
  • Name and postal address
  • Phone number
  • Marketing preferences
  • Sales transactions

14 kinds of customer data. The assistant needed 4.

What the assistant did with it

  • Phone numberFind the caller’s orders before they say a word
  • Sales transactionsKnow the order status without asking for an order number
  • Name and postal addressVerify the billing zip instead of a long security quiz
  • Marketing preferencesOffer to text tracking or a label, only to people who opted in
Customer data we collected, and the four pieces the assistant put to work.

06Write every turn by hand

With no LLM to improvise, nothing could be left to chance. For each intent I wrote the sample phrases that trained the intent model, the slots to fill, every prompt and confirmation, and what to say when it didn’t understand. Then I prototyped the whole conversation in Voiceflow.

One intent, by hand

return_order

Sample phrases, written to train the model
  • I need to make a return
  • I want to return something
  • How do I send this back
  • The vanity doesn’t fit
  • Return an online order
  • Can I get a refund
  • Start a return
  • I got the wrong item
Slots to fill
  • orderfrom caller ID, or ask
  • itemconfirm if the order has several
  • methodstore or mailed label
Every response, scripted
  1. Prompt“Is it the bathroom vanity from your order on March 3rd?”
  2. Confirm“You can return it at your Midtown store, or I can text you a free return label.”
  3. No match“Sorry, I missed that. Store, or a texted label?”
  4. No match, again“Let me get you to someone who can help.”
One intent written out, as an example of the method.

One assistant, three channels

The same intents and the same voice on the phone, by text and in web chat. Each one is live in the prototype.

01Before: press 2, press 2, press 2

For comparison, the old menu rebuilt from the tree. Asking about an order meant four key presses and long prompts, only to end up on hold.

The old phone menu, rebuilt: four presses to reach the hold queue. Open prototype

02After: just ask

The assistant greets the caller by name, finds the order and asks what they need. In the prototype, a return is done in about 30 seconds with no transfer.

A return over the phone: the assistant finds the order and texts the label. Open prototype

03Fix it in the text thread

When a delivery slips, the assistant texts first and solves it right there: it checks the local store, moves the order to pickup and cancels the late one.

A delayed tile order, switched to same-day pickup by text. Open prototype

04Help while they shop

On the website, Homer, the assistant’s chat persona, answers product questions in context. Here it works out how much paint a room needs and swaps the quart for a gallon.

Homer works out how much paint the room needs. Open prototype

Results

  • Navigation time fell from 98 to 28 seconds

    Callers reached an answer or the right person instead of fighting menus.

  • 35 million calls redirected from the service desk

    Less overflow, so fewer associates pulled off the store floor to answer phones.

  • Self-service containment rose 185%

    Callers finished tasks on their own, with no associate needed.