Forty days
Bec runs customer service for an outdoor furniture business in Perth. Seventy people, a showroom, a warehouse in Canning Vale and a busy website. In September they switched on an AI agent for web chat after hours.
On the second Saturday a customer wrote in. Her new dining table had a scratch across the top. It had been delivered 40 days earlier, but she'd only just taken the plastic off.
The agent checked the returns policy. Thirty days. It politely declined, offered her a care guide, and closed the chat.
By Monday there was a one-star review. Bec read the transcript at seven in the morning with a coffee she didn't enjoy.
Here's the thing. Her team would have replaced that tabletop without blinking. Scratches on a wrapped table almost always come from freight, and the business claims the cost back from the carrier. The team has handled it that way for years, out to about 90 days. It just isn't written anywhere. It's in Bec's head, and in the heads of the three people she trained.
The agent did exactly what the policy said. The policy just wasn't what the business does.
The warning signs
Customer service is usually where culture on paper meets a real person first. In most businesses we see, these signs show up.
1. Everyone knows the real policy except the policy. The document says 30 days. The team says "depends". New starters learn the difference by getting it wrong.
2. The answer depends on who picks up. One person waives the restocking fee for regulars. Another never does. Both are acting in good faith.
3. Escalation is a name, not a rule. "Check with Bec" is the process. When Bec's on leave, the process is on leave too.
4. The values promise more than the script allows. The website says "We make it right". The script says "Unfortunately, that falls outside our policy."
5. Your best people answer fastest because they know the workarounds. That's a skill worth having. It's also a sign the workarounds are where the real policy lives.
6. Customers ask for someone by name. Lovely for that person. It also means the customer has worked out the answer depends on who they get.
The gap
What the company says: "We make it right."
What's written down: A 30 day returns policy from 2022 and a three page refund procedure.
What really happens: Freight damage replaced out to 90 days. Trade customers get credit, not refunds. Restocking fee waived for repeat customers, if you get the right person.
Before you go and rewrite the policy, look at who's right. Usually it's the team. They've learned what keeps customers and what doesn't, one complaint at a time. The friction isn't that people bend the rules. It's that the bent version, the one that works, never made it back onto paper.
The AI test
This is the department where AI tools tend to turn up first, so it's where the gap gets found first.
A new person on the phones would have asked someone about the table. An AI agent can't ask. It has your written policy, your FAQ and whatever else you've connected, and it acts on those. It has no Bec.
Bec's team could have caught this before switching on. Here's the version we'd suggest. Take 50 recent tickets. Give the agent only your written policies and ask what it would do. Then compare with what your team actually did. When one business ran this, the agent and the team agreed on 31 and disagreed on 19. In 14 of those 19, the team was right and the written policy was wrong. Demo
That's not a failed pilot. That's 14 unwritten rules found in an afternoon, before a customer had to find them for you.
What an agent needs is the same thing a new starter needs, just written down:
- which exceptions are allowed, and up to what amount
- who can approve what
- when to stop and hand the customer to a person
If you give it those three things, it will usually do better than a new starter on day one. If you don't, it'll do exactly what Bec's agent did.
What good looks like
Picture Bec's Saturday again, with one extra page in place. The agent reads that freight damage on a wrapped item gets replaced up to 90 days after delivery. It reads that a replacement under $800 needs no sign-off Demo. And it reads that anything over that, or any customer who is clearly upset, goes to a person first thing Monday with a note attached.
The customer gets a new tabletop, or at worst a message saying Bec will call her on Monday morning. No one-star review. Same team, same products, same agent. The only difference is that the rule Bec has been carrying for years is now written where everyone, and everything, can find it.
The connection
Customer service mostly catches problems that started somewhere else. That makes it a good place to trace back from.
- Sales sets the promise. If a rep said "anything goes wrong, we'll sort it", customer service inherits that line with no policy behind it. That's week 5.
- Marketing writes "hassle-free returns" on the website. The agent reads the policy, not the website. That's week 6.
- Operations owns the freight damage process and the carrier claims. If that process isn't written down, customer service can't promise anything with confidence. That's week 7.
- Finance decides what refunds and replacements cost and who signs them off. If the dollar limits only exist in someone's head, so does the authority.
- The leadership team wrote "We make it right" without saying what "right" costs. Back to last week.
The ident check
Yes or no. Run it with your customer service lead.
- Could everyone on your service team name the three things the business most wants from a customer conversation, without notes?
- Could each of them say what they're allowed to approve on their own, and up to what amount?
- Could each of them name one old policy or exception that has stopped?
- Is every exception your team regularly makes written down somewhere an AI tool could find it?
- Would two people on your team give the same answer to the same complaint?
- Is there a written rule for when a customer gets handed to a person?
- Does what your website promises match what your written policy says?
Questions 1 to 3 are the recall test, adapted for the front line. If question 2 is a no, that's usually the most useful place to start.
Cheap first
- Pull ten recent tickets. Costs nothing, takes an afternoon. For each one, ask: what did we actually do, and does the policy say that? Mark the ones where it doesn't.
- Write down what your most experienced person knows. One hour with your Bec and a notepad. Just a list of the exceptions they make and why. Don't judge them yet.
- Run the same-decision test. Give five real cases to a team member and to an AI tool that only has your written policy. Compare the answers side by side.
- Decide out loud. For each exception: it becomes policy, or it stops. Write the keepers in. Tell the team which ones are gone.
- Then the harder fix. Agree who can approve what, up to which dollar amount, and write the handoff rule. That one page does more for your customers, and for any AI agent you add later, than a new values poster ever will.
The Monday close
This week, pull ten tickets and check them against your written policy. Count how many match.
Next week we start Part 2: how culture travels. First stop is the team leaders, the translation layer. They turn leadership's intentions into what happens on the floor, and they're usually doing it on their own, in their own words.
Questions people ask
- Why do AI customer service agents get answers wrong?
Usually because they follow the written policy exactly, and the written policy isn't what the business actually does. The exceptions experienced staff make every day were never written down, so the agent can't know them.
- What does an AI agent need before it answers customers?
The same thing a new starter needs, written down: which exceptions are allowed and up to what amount, who can approve what, and when to hand the customer to a person.
- How do you test an AI agent before switching it on?
Take 50 recent tickets, give the agent only your written policies, and compare what it would do with what your team actually did. Every disagreement is either a policy to fix or an unwritten rule to write down.
Read the episode transcript
The Ident Series, Episode 3: Customer service: where AI meets your culture first
Show notes
Episode title: Customer service: where AI meets your culture first
Summary: Bec's new after-hours chat agent turns down a scratched table delivered 40 days ago, and Monday brings a one-star review. Her team would have replaced it without blinking. Trav follows the plan down to the front line, where customers ask the questions nobody wrote down.
- Five things to listen for, from "everyone knows the real policy except the policy" to customers asking for someone by name
- Before you rewrite the policy, check who's right. Usually it's the team
- The 50 ticket test: what the agent would do against what your team did
- What an agent needs: allowed exceptions, who approves what, and when to hand over to a person
- This week's job: pull ten tickets and check them against your written policy
Full article: getsorted.valuebull.com.au/ident/customer-service/
Script
G'day. It's Trav, and this is The Ident Series. Episode three. Customer service, where AI meets your culture first.
Quick recap. Week one, the five signs your culture's only on paper. Week two, the leadership room, where plans go missing on the way out the door. This week we follow it all the way down to the front line. Because customer service is where customers ask the questions nobody wrote down.
Here's this week's story. Made up, but you'll know it.
(pause)
Bec runs customer service for an outdoor furniture business in Perth. Seventy people, a showroom, a warehouse out at Canning Vale, busy website. Back in September they switched on an AI agent for web chat, after hours.
Second Saturday, a customer writes in. Her new dining table's got a scratch right across the top. It was delivered forty days ago, but she's only just taken the plastic off.
The agent checks the returns policy. Thirty days. So it politely declines, offers her a care guide, and closes the chat.
By Monday, there's a one-star review. Bec's reading the transcript at seven in the morning, with a coffee she's not enjoying.
(pause)
Here's the thing. Bec's team would have replaced that tabletop without blinking. Scratches on a wrapped table nearly always come from freight, and the business claims the cost back off the carrier. They've done it that way for years, out to about ninety days.
It's just not written down anywhere. It's in Bec's head. And in the heads of the three people she trained.
The agent did exactly what the policy said. The policy just isn't what the business does.
(pause)
So, what does this look like in your business? Here's what I'd listen for.
One. Everyone knows the real policy, except the policy. The document says thirty days. The team says, depends.
Two. The answer depends on who picks up. One person waives the restocking fee for regulars. Another never does. Both doing it in good faith.
Three. Escalation is a name, not a rule. Check with Bec is the process. When Bec's on leave, so's the process.
Four. The values promise more than the script allows. Website says, we make it right. Script says, unfortunately that's outside our policy.
Five. Customers ask for someone by name. Lovely for that person. But it means the customer has worked out the answer depends on who they get.
(pause)
Now, before you go and rewrite the policy, have a look at who's right. Usually, it's the team. They've learned what keeps customers and what doesn't, one complaint at a time. The friction isn't that people bend the rules. It's that the bent version, the one that works, never made it back onto paper.
(pause)
And this is the department where AI tools usually turn up first. So it's where the gap gets found first.
A new person on the phones would've asked someone about that table. An AI agent can't ask. It's got your written policy, your FAQ, whatever else you've connected, and it acts on that. It doesn't have a Bec.
So here's what I'd do before switching one on. Take fifty recent tickets. Give the agent only your written policies, and ask it what it would do. Then compare with what your team actually did.
I'll give you an example, and I'll be upfront, the numbers are illustrative. Say the agent and the team agree on thirty-one, and disagree on nineteen. And in fourteen of those nineteen, the team was right, and the written policy was wrong.
That's not a failed pilot. That's fourteen unwritten rules, found in an afternoon, before a customer had to find them for you.
What an agent needs is the same thing a new starter needs. Just written down. Which exceptions are allowed, and up to what amount. Who can approve what. And when to stop and hand the customer to a person. Give it those three things, and it'll usually beat a new starter on day one.
(pause)
And remember, customer service mostly catches problems that start somewhere else. Sales sets the promise, we'll sort anything out. Marketing writes hassle-free returns on the website. Operations owns the freight claims. Finance decides what a replacement costs and who signs it off. And leadership wrote, we make it right, without saying what right costs.
It all lands on whoever's answering the chat on a Saturday.
(pause)
Right. Ident check. Seven questions, yes or no. Run these with your customer service lead. I'll go slow.
One. Could everyone on your service team name the three things the business most wants from a customer conversation, without notes? (pause)
Two. Could each of them say what they're allowed to approve on their own, and up to what amount? (pause)
Three. Could each of them name one old policy or exception that's stopped? (pause)
Four. Is every exception your team regularly makes written down somewhere an AI tool could find it? (pause)
Five. Would two people on your team give the same answer to the same complaint? (pause)
Six. Is there a written rule for when a customer gets handed to a person? (pause)
Seven. Does what your website promises match what your written policy says? (pause)
First three are the recall test, adapted for the front line. If number two's a no, that's usually the best place to start.
So, this week. Costs nothing, takes an afternoon. Pull ten recent tickets. For each one, ask yourself, what did we actually do, and does the policy say that? Mark the ones where it doesn't. Then count them.
And if you've got another hour, sit down with your Bec and a notepad. Just list the exceptions they make, and why. Don't judge them yet. Then later, decide out loud. Each exception either becomes policy, or it stops. Write in the keepers, and tell the team which ones are gone.
(pause)
Next week, we start part two. How culture travels. First stop, team leaders. The translation layer. They turn what leadership means into what actually happens on the floor. And most of the time, they're doing it on their own, in their own words.
I'm Trav. This has been The Ident Series. See you next week.