AI in the sale. What owner-led companies are doing with it, and what it's worth.October 8, 2026 · 7 stories
This week
What they changed the second time
An AI agent that recommended a competitor, an owner whose first AI build broke in days, a bank that apologized, and what each changed.
Editorial illustration / AI-generated
Hey,
Most AI stories you'll read are either a launch announcement or a cautionary tale. This week is about the part in between: companies whose first attempt went badly, sometimes publicly, and what they changed before it worked.
There are fewer stories than usual, because each one needs room for both halves. A support agent at Salesforce that recommended a competitor. A five-person restoration company whose first AI platform broke within days. An agency that pulled its AI SDR. A bank that cut jobs for a voice bot and then apologized. In almost every case, the fix was something about the business around the AI: the content, the handoff to a person, who owns what, or what got measured. Three of the stories are told by the companies selling the tools, so watch the labels.
01
What broke: too many rules, too much old content
Salesforce's own AI support agent recommended a competitor, quoted 2018 release notes and made it hard to reach a person. It fixed the content, the rules and the handoff.
Editorial illustration / AI-generated
~26%
abandonment rate when the agent launched
85%
resolution rate ten months in, about 45,000 conversations a week
5–8%
handoffs to a person now, up from about 1% at launch, after making a human easier to reach
~$100M
taken out of support costs, with satisfaction “as good or better”
Salesforce newsroom (July 2025), diginomica interview with Bernard Slowey (July 2025), CIO.com interview with Joe Inzerillo (October 2025)
Salesforce put its own Agentforce agent on help.salesforce.com in fall 2024 and has been unusually open about what went wrong. Early on, a customer asked about a competitor and the agent answered. “Our marketing team wasn't thrilled about that, so we added a grounding rule to never talk about our competitors,” chief digital officer Joe Inzerillo told CIO.com. The rollout had started small, with 10% of logged-in users in week one, “which gave us fewer than 150 conversations to sift through manually.” The rule then blocked legitimate questions, like how to connect Microsoft Teams. Another time, asked about a recent update, the agent quoted release notes from 2018, which support executive Bernard Slowey put down to “content collisions”: “When there's too many articles about the same thing, you have poor content hygiene.” And people left: “When we first launched, our abandonment rate was at about 26%,” Slowey told diginomica. Handoffs to a person ran at about one percent, which looked good at first. “What we realized was we made it too hard to get to the Support Engineer.”
The fixes were mostly not technical. They dropped the list of banned topics and told the agent to act in the best interests of Salesforce and its customers. “Less is more to start,” Inzerillo says. They used agents to “cull through content and clean it up,” and “if a piece of content hadn't been used in a year, we deleted it.” Slowey says they rewrote the prompts using “that soft skills training” their support staff go through. Customers can now “start the conversation with ‘talk with human,’” and Slowey says handoffs of five to eight percent are fine.
Ten months in, Salesforce says the agent handles about 45,000 conversations a week with resolution “in the 85% range.” Inzerillo says it has taken “about $100 million in cost out of our support function.”
The workflow
Agentforce, Salesforce's own AI agent, as its own first customer
Answers support questions from curated help content, and hands the conversation to a support engineer whenever a customer asks for a person.
What it changed Resolution in the 85% range and about $100 million out of support costs, as Salesforce reports it. Handoffs to people rose from about 1% to 5–8%, on purpose.
Vendor-reported Salesforce sells Agentforce, so this is the vendor grading its own product, and the newsroom piece is its own. The interviews are with independent outlets. Slowey says abandonment is “down about eight or nine percent” without saying whether that's points or percent, so it isn't in the numbers above. The outside interviews still carry Salesforce executives' own figures, and the $100 million isn't stated as annual or cumulative. The competitor rule, “less is more” and the $100 million are from CIO.com's interview with Inzerillo (October 2025). Salesforce is far larger than any reader, but the fixes are ones a small team can copy.
Three things here would apply to a business of any size. Delete the old content before you blame the bot. Tell it what you're trying to do for the customer instead of giving it a list of banned topics. And make the way to a person obvious. I'd read a rising handoff rate as the agent knowing its limits, which is what you want from a new hire too.
02
What broke: the platform, the memory, the laptop
A restoration company owner's first AI build broke within days. The rebuild took new-job intake from 30–45 minutes to about 90 seconds.
Editorial illustration / AI-generated
Joyce Gabriel co-founded Expert Restoration, a water, mold and fire restoration company in Washington state. It did “just over $800,000 in revenue” in 2025 with a team of five. Every new job meant the same routine: verify the claim, log it in the CRM, build a Drive folder, text the company, call the operations manager, dispatch the crew, all while adjusters called about other jobs. “Every. Single. Loss.” She “started looking seriously at AI agents in early 2026”: Atlas for finance, Freddy for operations. “The first platform we tried broke almost immediately,” she wrote in C&R Magazine. “QuickBooks authentication wouldn't cooperate. Tokens expired. Connections failed.” They scrapped it and rebuilt on another platform. Then a procedure she had written “was sitting in the workspace but not in his long-term memory file.” Then the laptop it ran on failed, so they moved Freddy to a cloud server “at midnight, while also handling two brand new jobs.”
30–45 min~90sec
To take in a new loss at Expert Restoration
The owner's own account
By early May, QuickBooks and Google Drive were connected and field updates fed one database. She now gets a finance brief every Monday, the management team gets a morning brief in Slack, and “new loss intakes that once took 30 to 45 minutes dropped to around 90 seconds.”
The workflow
Two AI agents she built: “Atlas” on QuickBooks and “Freddy” for operations, on a cloud server
Freddy takes a new loss, creates the folders and starts the dispatch steps. Atlas watches open invoices and sends a weekly finance brief.
What it changed New-loss intake down from 30–45 minutes to about 90 seconds, in the owner's words.
Anecdotal One owner's first-person column, part one of a series, with one self-reported time figure and no revenue figures yet. She doesn't name the platforms she used. Her bio says she is building AI systems for restoration, so she may sell this in future.
This is the closest thing I've found to what an owner actually goes through. What broke were a login, a file in the wrong place and a laptop. The AI's answers aren't the problem she describes. She calls it “the ugly middle,” and I'd budget for it: a few months where the setup is the work, before the time savings show up.
03
What broke: more emails, wrong buyers. Results: not published yet
A marketing agency pulled its AI SDR. Its CEO says it “worked exactly as told, and that was the issue.”
Editorial illustration / AI-generated
Eric Siu runs Single Grain, a digital marketing agency, and in September he wrote about uninstalling the AI SDR his team had been running for outbound. “It worked exactly as told, and that was the issue. It emailed everyone on the list with the pitch we gave it, at a pace no rep could match, and every weak assumption baked into the list and the offer got amplified the same way.” His description of what breaks: the wrong buyers on a loose list, domain reputation (“Within weeks even your human reps land in spam”), and a tool that keeps emailing someone who “already said ‘not now.’” His rebuilt approach, which he also gives as advice to other teams, keeps AI on research, list prioritization and first drafts, and puts a person on every send. “I would rather send fifty good emails a human approves than five thousand an agent guesses at.”
He changed the number the team reports: “cost per meeting that actually shows up,” not emails sent, replies or meetings booked.
The workflow
AI for prospect research, list ranking and first-draft personalization, with a person approving every send
AI narrows the list and drafts the opening lines; a rep edits and sends each email.
What it changed No figures published. The scoreboard moved from emails sent to meetings that actually show up.
Anecdotal The CEO's own post, with no numbers on either side. Single Grain sells marketing consulting and links to it at the end of the piece.
I'd take his scoreboard even if you never touch an AI SDR. Emails sent and meetings booked can both go up while outbound gets worse. Meetings held with someone who could buy is the number that tells you whether it's working, and I'd track it before switching any tool on.
04
What broke: one team doing two jobs
A software company's AI started out with 30% satisfaction. A year of rebuilding later, it was at 71% and support was bringing in revenue.
Editorial illustration / AI-generated
30% → 71%
customer satisfaction with the AI agent
40%
more new revenue, as Intercom reports it
30% → 55%
share of questions the AI resolved, Dec 2023 to Mar 2025
~1 year
before the team was “satisfied with the results”
Intercom customer story (2025) and The Growth Code interview with Emil Korpar (June 2025)
Databox, a business analytics software company, had grown its support and sales development teams to 40 people for a spike in demand that didn't arrive, then had to cut back. It combined the two teams and put Intercom's AI agent, Fin, in front of customers. The agent started out resolving about 30% of questions, and customer satisfaction with it was 30%. Emil Korpar, who runs support, told The Growth Code that the fix was mostly content. They split the knowledge into three sources: how-to guides, a library explaining each metric's quirks and limits, and a community of real use cases written “in customer language, not product language.” The other fix was the team. The interview says Korpar “tried having reps do both reactive support and proactive revenue hunting,” and it didn't work, so they split the roles, and reps now carry two goals: satisfaction and revenue.
Intercom reports resolution up to 55%, satisfaction with the agent up to 71%, team productivity up 50% and new revenue up 40%. Korpar told The Growth Code: “It took us all together about a year to be satisfied with the results.”
The workflow
Intercom's Fin AI agent, plus three rebuilt knowledge sources
Fin answers routine questions from the help center, metrics library and use-case library. Reps split into reactive support and proactive expansion, with a revenue goal.
What it changed AI satisfaction from 30% to 71% and new revenue up 40%, as Intercom reports it.
Vendor-reported The headline numbers come from Intercom, which sells Fin. The Growth Code interview is independent but reports slightly different figures (satisfaction from 29% to 75%+) and a vaguer revenue range. Both are from 2025, outside our usual window. Databox is a software company, not an owner-led one.
The part I'd steal is the metrics library: a page for every place where the product doesn't do what customers expect. Most businesses have a list like that, and it usually lives in one person's head. The second lesson is the split. When the same people were asked to answer tickets and find revenue, they didn't do both well.
05
What broke: cost became the only goal
Klarna's CEO says cost drove its AI customer service push and quality suffered. It now pitches a human as the VIP experience.
Editorial illustration / AI-generated
853
full-time agents' worth of work done by its AI assistant, the company says
In 2024 Klarna said its AI assistant was doing the work of 700 customer service agents. By May 2025, CEO Sebastian Siemiatkowski told Bloomberg that “as cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality,” and Klarna started hiring human agents again. Forrester's Kate Leggett later told CX Dive: “They overpivoted to cost containment, without thinking about the longer-term impact of customer experience.” On the 20VC podcast in February, Siemiatkowski described the reaction: “People were angry to some degree as well,” and “it was like an eye opener for us.” The rebuilt model splits the work. The AI keeps the simple, repetitive questions, and, as he told Bloomberg, “there will be always a human if you want.” “We said the future of VIP experience will be the human connection, the relationship,” he said, and Klarna is recruiting some of its own customers to work in its customer service, “just like somebody can go and drive an Uber for a while.”
Klarna now says the assistant does the work of more than 853 full-time agents, has saved $60 million, handled two-thirds of inquiries and cut repeat issues by 25%, with satisfaction “on-par” with a human. Siemiatkowski says satisfaction with the in-house human agents is “through the roof.”
Directional Klarna's own figures, reported by CX Dive and Maginative (citing Bloomberg); none are independently checked. Its customer service and operations costs rose from $42 million to $50 million year over year in Q3 2025, which complicates the savings story. The CEO says nobody lost their job in the original change; Maginative reports Klarna shrank mainly through natural attrition.
The line I'd remember is that cost was “a too predominant evaluation factor.” If the only number you track is what the AI saves, my guess is you'll keep pushing it into conversations it handles badly. Klarna's answer was to decide which questions should stay cheap and fast and which ones a person should take, and to say so to customers.
06
What broke: cutting staff before the volumes were in
A bank cut 45 jobs for a voice bot, then apologized and reversed it. Its messaging AI, built with handoffs from the start, resolves about 85% of self-service chats.
Editorial illustration / AI-generated
At the start of August 2025, Commonwealth Bank of Australia confirmed it had cut 90 jobs, including 45 roles in direct banking, after introducing a voice bot on its inbound customer enquiries line. Three weeks later it reversed the decision. The bank said its assessment “did not adequately consider all relevant business considerations and this error meant the roles were not redundant,” and added: “We have apologised to the employees concerned and acknowledge we should have been more thorough.” The union called it an attempt to “dress up job cuts as innovation.” On messaging, the bank had taken a different route. Its system, built with Microsoft, sends simple requests to a bot, regulated ones like fraud disputes down a fixed path, and anyone who sounds vulnerable to a person, with the whole conversation passed along. “Where a customer's language indicates vulnerability, we expect a human involved,” says Rachel Round, who runs self-service. Its head of engineering, Shashank Verma: “Customers have very low tolerances for failure.”
84.6%
Self-service messaging interactions resolved end to end, May 2026
Commonwealth Bank, via Microsoft
“In May 2026, approximately 84.6% of self-service messaging interactions were resolved end-to-end in the messaging channel,” according to Microsoft's account.
The workflow
An AI routing layer built with Microsoft, migrating nearly 700 old chatbot topics
Reads what the customer wants and sends it to a bot, a fixed process or a human specialist, with the conversation summary attached.
What it changed About 84.6% of self-service messaging interactions resolved end to end, May 2026.
Vendor-reported The reversal is from the bank's statements to ACM newspapers. The 84.6% comes from Microsoft, which built the system with the bank. The two aren't one story told in order: the messaging bot launched in November 2024, before the voice-bot cuts, and Microsoft's piece doesn't mention the reversal. Putting them side by side is my framing. The voice side is still being built.
The bank cut staff on its voice line before the voice bot had proved itself, while its chat system was designed from the start to know when a person should take over. The chat system is the one with the strong numbers. If you're adding AI to your phones or chat, I'd wait for real volumes before changing anyone's job.
07
What the research says
Companies getting real value from AI are nearly three times as likely to have redesigned the work around it
Editorial illustration / AI-generated
McKinsey's latest State of AI survey, of 1,719 people in 97 countries, defines AI high performers as companies that attribute 5% or more of their earnings (EBIT) to AI. “Nearly three-quarters of high performers report fundamentally redesigning workflows because of their AI use, up from 55 percent last year. By comparison, just one-quarter of other respondents report doing so.” Only 37% of all respondents say AI has added anything to EBIT. Gartner, surveying 321 customer service leaders, found only 20% had actually cut agent staff because of AI, and predicts that “by 2027, 50% of companies that attributed headcount reduction to AI will rehire staff to perform similar functions, but under different job titles.” Its researcher Emily Potosky: “Relying solely on AI right now is premature and could lead to unintended consequences.”
Share reporting they fundamentally redesigned workflows because of AI
Share reporting they fundamentally redesigned workflows because of AI
AI high performers, 2025
55%
AI high performers, 2026 (nearly three-quarters)
75%
Everyone else, 2026 (one-quarter)
25%
McKinsey State of AI, August 2026, 1,719 respondents. High performers attribute 5% or more of EBIT to AI. The 75% and 25% are approximate: McKinsey says “nearly three-quarters” and “one-quarter.”
Directional McKinsey's respondents skew to large companies, and the survey shows correlation, not cause. Gartner's rehire figure is a prediction, not a measurement. Both sell advisory services on AI.
This fits most of the stories above: Salesforce rewrote its content and its handoffs, Databox split its team, and Expert Restoration rebuilt its plumbing. I'd expect the redesign to be most of the work.
The workflows behind the results
Every tool in this issue, what it does, and what it changed.
Put the stories next to each other and the same fixes come up. Old or conflicting content got cleaned out (Salesforce, Databox). Getting to a person got easier, and handing someone off stopped counting as a failure (Salesforce and Klarna; the bank built it into its chat system from the start). The scoreboard changed from activity or savings to an outcome (Single Grain's meetings held). And someone owned the plumbing until it held (Expert Restoration).
If you've tried AI somewhere in your sale and it went badly, here's a 20-minute post-mortem before you write it off. Write down what it was supposed to do and the one number that would show it worked. Then check four things: was it working from current information, could a customer reach a person easily, was it measured on an outcome or on activity, and did anyone own it week to week? My guess is at least one of those was missing, and that's usually what to fix before trying again.
Cheers, TeeJay Johnson Owner, Foundry Revenue Partners
Editorial illustration / AI-generated
Free pre-assessment
Is your sales process ready for AI?
The free pre-assessment looks at how your sales engine is running today against your goals, before anything gets recommended.