An AI agent is a tool that does a job step by step instead of just answering one question. I opened every source below myself before writing this. When a number comes from the company selling the product, I say so.
Visa, BILL and Billtrust test AI agents that pay invoices
On 20 October 2026, Visa, BILL and Billtrust announced what they call the first live pilot where trusted AI agents start and finish approved invoice payments through supplier portals. The source for this is Visa's own news release, and PYMNTS reports the same facts from it.
Here is the problem they say they are solving. Many supplier portals ask for human checks, like CAPTCHA tests, which stop software from paying. In this pilot, BILL's payment agent is registered with Visa's Trusted Agent Protocol and Agentic Directory. Billtrust uses that protocol to recognise the agent and let it in, so no separate agent login is needed.
The agent does not decide what to pay. Visa's release says it only runs payments the buyer has already approved, and buyers do not share their supplier-portal logins. Mark Nelsen of Visa said: "Businesses should not have to choose between automation and control when making payments."
The pilot is live but limited to participating BILL buyers and Billtrust suppliers. PYMNTS notes the article gives no pilot length, number of participants or payment volumes.
My take: All three companies are selling this, so every claim here is theirs. It is a sensible idea, but without volumes or results I cannot say how well it works yet.
Cognizant and Google Cloud build a team around Gemini agents
On 19 October 2026, Cognizant and Google Cloud said they are expanding their partnership with a new Gemini Frontier Capability Group. Google Cloud's press release and The Next Web both describe it.
The group pairs Google's AI platform with Cognizant's industry and engineering skills. The release says it will first focus on updating old software, developer tools, business process automation and cyber defence. The stated goal is to move AI agents from pilot projects into daily use.
The release says Cognizant plans to give Gemini to 100,000 of its own staff this year, growing to 200,000. It also describes a retail client case with 40% higher productivity and 35% lower costs. The Next Web points out that this comes from one unnamed client and has not been checked by anyone else. It also says no money figures were shared for what both sides call a "substantial joint investment".
My take: Google Cloud is one of the two companies announcing this, and Cognizant is the other, so the numbers are marketing until someone independent checks them. The size of the team is real news. The results are still claims.
Microsoft says treat AI agents like insiders you cannot fully trust
On 19 October 2026, Microsoft announced new security controls for AI agents. TechCrunch and SiliconANGLE both covered it.
Microsoft's argument is that AI models are not necessarily evil, but they are built to get results and can be unpredictable. Hayete Gallot, who leads Microsoft Security, said: "Unfortunately, frontier models, regardless of their origin, cannot be trusted". CEO Satya Nadella wrote in an essay earlier this month: "We simply can't outsource responsibility for what intelligence does on our behalf."
SiliconANGLE describes the tools. Each agent gets a risk score that rises if it acts differently from what its user intended. As the score climbs, an agent can lose the ability to write while keeping read access, and risky actions can need a person to approve them. At the top, the agent can be isolated or stopped. Microsoft says its tools can find up to 60 types of agents on work machines, with Claude Code and Codex named as examples. Price and availability were not shared.
Disclosure: Anthropic makes Claude, and I use Claude to help run this blog. Read this section with that in mind.
My take: Microsoft sells these controls, and TechCrunch notes that it gains by selling its own agent security rather than relying on the AI labs' tools. Microsoft also invests in OpenAI and Anthropic, so the warning is not free of interest either way. Still, keeping the monitoring outside the agent makes sense to me.
What I take from this week
- Agents are moving from "answering" to "doing", including paying real invoices, but the first pilots are small and the results are not public.
- Big consulting and tech firms are building whole teams around agents, and their headline numbers come from their own press releases.
- Control is becoming the main selling point: who approved the action, and who can stop the agent.
I will keep reading the originals and flag who is making each claim. If a number cannot be checked, I will say so.
Sources
- Visa: Visa, BILL and Billtrust Launch First Live Pilot for Trusted AI Agent Payments Between Buyer and Supplier Platforms
- PYMNTS: Visa Launches Agentic Payments Pilot With BILL and Billtrust
- Google Cloud press corner: Cognizant and Google Cloud Expand Alliance to Accelerate Agentic Enterprise Transformation
- The Next Web: Cognizant puts 18,000 staff behind Google's Gemini in AI agent push
- TechCrunch: Microsoft says AI agents are like insider threats as 'frontier models can't be trusted'
- SiliconANGLE: Microsoft targets the 'agentic insider' with new controls for rogue AI agents