An AI agent cuts the data work of switching providers, and Google starts selling Gemini 4 Carbon

The same work used to take two and a half weeks. Northern Trust, a Chicago bank, reports on the investments of large investors, and each new client arrives with years of transactions recorded in its own way. The agent converts that history into the bank's standard format and sends the cases it is unsure about to analysts, who correct them.

That kind of conversion is where most of the cost of changing providers hides. A payroll bureau or a freight forwarder keeps your history in its own codes, so moving means weeks of someone matching old product and customer codes to new ones. Many companies stay with a supplier they dislike because nobody has those weeks to spare.

The provider that wins a new customer gains as much. For example, a contract warehouse that signs a retailer must load the retailer's product list and two years of orders before the first pallet arrives. If an agent matches the codes and a planner checks only the lines it flags, a team that set up one new customer each month could set up four. Three days is about a quarter of two and a half working weeks.

Within a year, buyers will ask how many days it takes to go live as routinely as they ask the price. A provider whose customers stayed because leaving was hard will have to keep them with service. That includes your company, if its customers stay mainly because moving their data would be hard.

A typical company that pays for AI spends $16.32 per employee a month. A bill of $10,200, more than many people earn, only makes sense when AI does a whole job.

Customer support is the second whole job to move, after writing software: its share of spending went from 0.8% in June to 13.3% in October. Requests arrive in writing, and each one ends in a result anyone can see, a ticket closed or reopened.

The companies at the top budget AI the way they budget staff, against the cost of the work it takes over.

In Antigravity, Google's app for coding with agents, typing /swarm turns one request into a team: a coordinator splits the work, and verifier agents run tests on every change.

A team is only as good as its tests. When Google's researchers used agents built on Carbon to check the code behind Signal's encrypted messages, they put back two bugs that had been fixed. Their checks caught both.

Before trusting a team to build a tool that converts supplier price files, put a negative price and a missing product code in the sample, and see whether the verifiers reject them.

Databricks, a data software company, has 12,000 employees. Past about 150 people, bad news gets softer at each layer of management before it reaches the top. So Ali Ghodsi, its chief executive, has the company's agent, Genie, read the records instead. It went through the weak scores in a staff survey and helped him reject fixes that his executives had proposed.

Those records are honest because they were written for someone else: survey answers for HR, call notes for the next salesperson. Once staff learn that the boss's agent reads them, they will start writing for the boss, and the softening comes back. How long the agent keeps hearing the truth depends on what staff believe it is for, and the manager gets one chance to tell them first.

Since September, DeepL Voice has spoken its translations in the speaker's own voice, but it copied that voice from a sample at the start of the call, so the tone never changed. Its new model, Aure, follows tone, emphasis and pace as the call goes on.

So the listener now hears how something was said, and the speaker cannot hear what arrived. A dry remark about a late delivery can land as an insult, and nobody on the speaker's side would notice. After a translated call with a supplier, send the agreed prices and dates in writing.