Enterprise AI Agents Are Getting Wallets and Identities. They Still Can’t Look at a Broken Machine.

The AI agent working your support queue is about to get a corporate card. Visa and Mastercard both shipped agentic payment rails this year. Mastercard’s Agent Pay gives agents tokenized identities so merchants can verify who, or what, is buying. Visa’s Trusted Agent Protocol does the same with attestation headers. Startups like InFlow now issue Visa cards where the customer is literally an AI agent, complete with identity, onboarding, and a multi-currency wallet (Forbes covered it in May).

So the agent can spend money. It can prove its identity. It can sit in your org chart with credentials and policy-governed permissions, like a new hire that never sleeps.

Ask it to look at the error light on a customer’s machine, though, and it goes silent. It has a wallet and an ID badge. It does not have eyes.

The fastest adoption curve service has ever seen

The numbers behind this are real, not vendor hype. Salesforce’s 2026 State of Service research found that AI agent adoption in customer service jumped from 39% to 66% in a single year. Even better, 70% of teams deploying agents reported measurable value within 60 days. Nothing in service operations has ever moved this fast.

I believe those numbers. AI agents genuinely crush a huge slice of support work: password resets, order status, refund policies, billing questions. Anything that lives in a database, an agent can now handle faster and more politely than most humans.

Meanwhile, the industry is racing to give these agents more agency. Payments. Identities. Permissions. Google even has a protocol, AP2, for signed purchase mandates, so an agent can commit money on your behalf with a cryptographic paper trail. The direction is obvious: agents are becoming economic actors, not chat widgets.

Everyone is sold. The plumbing is not.

Here’s where it gets uncomfortable. ZDNET’s look at Salesforce’s field service data found that 95% of field service organizations already use AI, and 85% plan to spend more. The industry is fully sold.

Yet the same research shows the average enterprise runs over 1,000 software applications, and only 28% of firms share employee and customer data across the business. Read that again. Nearly everyone bought the agent; barely a quarter built the nervous system it needs.

Even more telling: field service leaders ranked transparency into how AI makes recommendations as their top criterion when picking partners. In other words, the people closest to physical work don’t just want AI. They want to see how it reached its conclusion, because they act on it while standing next to a humming, sparking, very real machine.

That instinct is correct. It also points at a deeper problem nobody in the agentic-commerce gold rush wants to discuss.

Agency without perception is a liability

Think about what we’re actually building. An agent with a wallet can order a replacement part. An agent with an identity can open a work order, dispatch a technician, and approve the invoice. That’s real organizational power.

Now think about what triggers those actions in field service and hardware support: a physical fault. A miswired cable. A blinking amber light. A pump that sounds wrong. A gasket seated at a slight angle that no customer on earth can describe accurately in text.

The agent making those economically consequential decisions works entirely from words. It reads the customer’s typed guess about what’s wrong, matches it against a knowledge base, and acts. If the customer says “the light is red” when it’s actually amber, the agent dispatches the wrong part with perfect confidence. Then it pays for the mistake with its shiny new wallet.

We wrote before about why agentic support still can’t see the customer’s problem. Giving that same agent payment credentials doesn’t fix the blindness. It raises the cost of every blind decision.

Gartner already predicts 40% of enterprises will demote or decommission autonomous agents by 2027 because governance gaps only surface after production incidents. I’d bet a meaningful share of those incidents will trace back to one root cause: the agent acted on a description of reality instead of reality itself.

The missing sense

Humans figured this out a long time ago. Before a good technician orders a part, they ask to see the machine. Before a good support rep escalates to a truck roll, they ask the customer to point their camera at the thing. Vision grounds decisions in the physical world, and it’s cheap insurance against expensive mistakes.

AI agents deserve the same grounding, and so do the humans supervising them. When a support interaction involves anything physical, the fix is not a longer prompt or a bigger knowledge base. The fix is looking at it. A live camera feed from the customer’s phone turns “the light is red, I think” into ground truth in about fifteen seconds. No app install, no guesswork, no confident wrong answer.

That’s the entire premise behind Viewabo: give support teams, and increasingly their AI agents, a live view of the customer’s actual problem before anyone commits money, parts, or a technician’s afternoon to a guess.

The industry spent 2026 giving agents wallets and identities. That was the easy part, because payments and credentials are solved problems being adapted to a new actor. The hard part, the part that determines whether those 66% adoption numbers survive contact with physical reality, is perception.

An agent that can pay for anything but see nothing isn’t an autonomous worker. It’s a purchasing department with the lights off. Before you hand yours a wallet, hand it eyes.