Field NotesStory5 min read
The new AI support desks still give 40 percent of the work to people
A six-person startup now runs whole support departments for other companies. Here is how it splits work between agents and people, and what to copy.
Adam Blazer
Fractional Chief Intelligence Officer
On March 2, TechCrunch profiled 14.ai, a Y Combinator company that takes over a startup’s entire customer support operation. It has six people. They rotate to cover clients around the clock, and AI agents do most of the work in between.
The story that made the rounds was about speed. A supplement brand with a large ticket backlog handed its support to 14.ai on a Thursday morning. By that afternoon, the backlog was cleared across social media, SMS, email, chat and voice. The brand’s previous support agents had been based in the Philippines.
That is a good headline. The more useful detail for anyone running a company is buried lower in the piece.
The split that matters is 60 and 40
Tom Blomfield, the Y Combinator partner quoted in the article, said that with the right integration, AI can solve about 60 percent of support tasks on its own. People handle the other 40 percent. He expects that balance to shift over time, and he described 14.ai as a support department that is both AI and human.
- Solved by AI on its own60%
- Handled by a person40%
That number is more honest than most vendor pitches, and it should shape how you plan. An AI support rollout doesn’t end with zero people. It ends with fewer people doing harder work: refunds that need judgment, angry customers, edge cases the agent has never seen. If you plan for 100 percent automation, you will be disappointed and your customers will be annoyed. If you plan for 60, you can staff for the 40.
Note the condition, too. The 60 percent depends on the right integration. An agent that can’t see your order system can only apologize. An agent that can look up the order, read the tracking number and check the return policy can close the ticket.
The bigger players are betting on the same idea
Two days after the 14.ai story, TechCrunch reported that Decagon, which builds AI support agents for large companies, ran its first employee tender offer at a $4.5 billion valuation. The company says it has more than 100 large customers, including Avis Budget Group, 1-800-Flowers, Quince, Oura Health and Away Travel. Its agents resolve customer questions across chat, email and voice.
The same article cited a Gartner estimate of 17 million contact center agents worldwide. That is the size of the work these companies are going after.
The two stories describe the same shift at different sizes. Decagon sells software to big brands that keep their own teams. 14.ai sells the whole department as a service to small companies that don’t want to build one. In both cases, the agent handles the repeatable questions and a person handles the rest.
What 14.ai does that you can copy
You don’t need to hire an agency to use the same approach. Several things in the TechCrunch piece translate directly to a 30 or 300 person company.
They connect to the systems first. 14.ai says it can plug into a client’s support system within a day. Every useful answer an agent gives depends on data it can reach. Before you pick a tool, list the systems a support rep opens during a typical ticket: the help desk, the order system, the shipping tracker, the billing platform. The agent needs read access to the ones that matter.
They put every channel in one place. The company monitors email, calls, chat, TikTok, Facebook, Telegram and WhatsApp. Most small companies have support scattered across an inbox, a social account someone checks on Fridays and a phone line. An agent can’t answer a message it never sees.
They start with the backlog. Clearing a pile of old tickets is a low-risk first job. The customers have already waited, the questions are known, and you can check every answer before it goes out. It is also the fastest way to see what share of your tickets an agent can actually close.
They test on their own business. 14.ai runs its own consumer brand, a glucose gummy for people with Type 1 diabetes, partly as a testbed to see how far AI can run on its own. The lesson for an operator is to try the agent on your own lowest-stakes queue before you point it at your biggest customers.
They read the conversations. Co-founder Michael Fester described the company as a revenue growth engine as well as a support agency, because it captures what customers are saying and turns it into insight. Your support queue is the most honest feedback channel you have. If an agent is reading every ticket, have it tag the reasons people write in and send you a weekly count.
- 01Message arrives on any channel
- 02Agent looks up the order
- 03Agent answers routine cases
- 04A person handles exceptions
Where this goes wrong
The failure mode isn’t hard to predict. A company turns on a support bot with no access to order data, tells it to deflect as many tickets as possible and measures success by how few tickets reach a human. Customers get polite non-answers and give up, which looks like a win on the dashboard and shows up later as churn.
The 60/40 framing protects you from that. If you plan for people to keep 40 percent of the work, you will design a clean handoff, and the person picking up the ticket will see what the agent already tried. You will measure resolution, meaning the customer’s problem was solved. Deflection only means the customer stopped writing.
There is a staffing question to settle early as well. The people who stay on the queue will spend their day on the hardest 40 percent. That work needs more judgment and more authority than the old mix of password resets and tracking questions. Give those reps the power to issue a refund or a replacement without escalating, or the agent’s speed on easy tickets will be wasted waiting on approvals for the hard ones.
What to do this week
- Export the last 30 days of support tickets and tag each one by type: order status, returns, billing, product questions, complaints.
- For each type, write down which system a rep opens to answer it. Types that need a single lookup are your first agent candidates.
- Pick one channel, usually email, and one or two ticket types to start with.
- Write the handoff rule in one sentence. For example: anything involving a refund over a set amount, or any message with an angry tone, goes to a person with the agent’s notes attached.
- Run the agent on your existing backlog first, with a person approving every reply, and count how many it got right.
By the end of next week you will know your own split. It might be 60 and 40. It might be 30 and 70. Either way, you will be planning from your numbers instead of a vendor’s.
Sources
- A married founder duo’s company, 14.ai, is replacing customer support teams at startups (TechCrunch, March 2, 2026)
- Decagon completes first tender offer at $4.5B valuation (TechCrunch, March 4, 2026)
