THE NUMBER
700
Klarna replaced 700 customer service workers with AI, announced $40 million in annual savings, and two years later began quietly rebuilding its human CS team after customer satisfaction collapsed. The tools work when the job is scoped right. That's what this newsletter is about.
3 THINGS HAPPENING RIGHT NOW
The most explicit case yet of a company naming AI agents as the reason for layoffs
Cloudflare cut 1,100 jobs - roughly 20% of its entire workforce - after internal AI agent usage grew 600% in three months. CEO Matthew Prince said "A lot of the support roles are not going to be the roles that drive companies going forward." Cloudflare beat earnings. The stock dropped 24%. It's the most explicit case yet of a major public company naming agents specifically - not "automation" in the abstract - in an official restructuring announcement.
The agent skills small business owners can actually use today
Anthropic launched a plug-and-play package that puts AI agent skills inside tools small businesses already use: QuickBooks, HubSpot, Canva, DocuSign, Google Workspace, and Microsoft 365. Fifteen pre-built skills cover payroll, financial operations, marketing campaigns, and HR tasks - no setup, no developer, no new software to learn. A payroll review that currently takes a small business owner three hours becomes a 10-minute approval workflow. Anthropic is also running a free workshop tour across 10 cities, with 100 local business owner seats per stop.
One scoped agent, more than double the qualified pipeline
One B2B software company reported dramatically cutting average response time after deploying an agent for lead qualification. Qualified lead volume more than doubled. Administrative time per sales call dropped significantly. One clearly scoped job for one agent.
THE DEEP DIVE
When the math stopped adding up
Klarna launched its AI customer service system in 2024 with a clear headline: 700 workers replaced, $40 million saved annually. The coverage was immediate and loud.
Then came the edge cases. Customers disputing charges with unusual circumstances. Calls from people in financial difficulty who needed patience, not efficiency. Multi-step problems requiring coordination across systems the AI couldn't access. Satisfaction scores dropped.
By 2026, Klarna shifted to a hybrid model and started rebuilding its human CS capacity. The Klarna case has become a cautionary reference point in boardrooms, putting the burden of proof on any AI restructuring proposal.
The real math: $40 million in claimed savings doesn't account for the cost of eroding customer trust at scale, or the restructuring charges that come with rebuilding what you just dismantled.
The lesson isn't "don't use agents." The agents that work are narrow: answer this category of question, qualify this type of lead, schedule this kind of appointment. The ones that fail try to do everything a person does in every situation. If your business takes inbound calls or emails, the agent is the screener - not the relationship. Keep the relationship human.
ONE THING TO TRY THIS WEEK
Before you build an agent, map the job. Most failures happen because the scope was wrong from the start - too broad, trying to replace a whole role instead of a specific task.
Open Claude Code. Type this, filled in for your own business:
I run [describe your business - e.g. "a 4-person bookkeeping firm"].
Here is a recurring task that takes too much of my time each week:
[describe it in plain language - e.g. "Every Friday I pull data from
three different software tools, combine it into a spreadsheet, and
write a summary email to each client. It takes about 2 hours."]
Give me a map:
- Which parts of this could an AI agent handle without my judgment?
- Which parts require me?
- What is the smallest version I could test first?
Don't build anything yet. Just give me the map.
Read what comes back. The "agent can handle this" column is your starting point for your first build. The "requires you" column is where you stay in the loop - like Klarna should have.
Pick one item from the first column. That's your next move.
Stuck? Reply to this email. I'll help.
WHAT'S COMING
Next issue: the back office, automated. A bookkeeper closes her month-end books in 90 minutes - the work that used to eat eight hours of reconciling and chasing receipts now happens while she sleeps.
Manu