Strategy & Insights
November 16, 2026

Personal AI agents are already calling your business. Here's how to get ready

One agent and one weekend can rewrite your service policy. Getting ready doesn't depend on which protocol wins.

Yoav Kolodner, CEO, Tribal
Personal AI agents are already calling your business. Here's how to get ready

Daniel, a good friend of mine, has an espresso machine that stopped heating 3 weeks after its warranty ran out. One Friday afternoon, he handed the problem to Instinct, his personal agent, and went off to enjoy his weekend.

Instinct kept working. Over 2 days, it opened 5 support threads across 3 channels: chat, email, and the web form. Somewhere along the way it found an old help article promising a free repair for that exact fault, and it quoted the article in every thread.

On Monday morning, the manufacturer called Daniel. Good news: a technician would come out, free of charge. Also, could he please stop bombarding their support team?

The part I keep coming back to is who made the call. An outdated help article and an agent who wouldn't stop asking for approval of that repair. Nobody at the manufacturer chose it.

That was 1 agent over 1 weekend. Now picture every one of your customers sending one.

The swarm is already here

Meta launched Muse in September. By October 21, it had more than 10 million users in the US and Canada, and one of its launch examples was lowering a bill. Instinct raised $1 billion at a $10 billion valuation for an agent that books flights, calls restaurants, and sorts out problems with your cable bill.

These agents don't sleep, and they don't stick to one channel. One Instinct user asked it to find a restaurant table; his agent's log showed roughly 200 requests an hour, and Resy deactivated his account. Amazon blocked Muse from shopping on its site altogether.

Deloitte Digital made a point recently that every service leader should sit with. Friction quietly holds your contact volume down. In 2025, 77% of US households had a problem with a product or service, and 68% said complaining took real effort. So people let a disputed fee go, because an hour on hold isn't worth $12.

An agent will gladly spend that hour. It queues the task, retries the chat when it crashes, quotes your own refund policy back to you, and asks again until it gets a yes.

The friction that kept your volume manageable is about to stop working.

The front door is getting standardized

The industry is moving fast on the handshake. Sierra recently released version 1.0 of the Personal Agent Protocol, known as Poppy. Muse now looks for Poppy first on every site it visits. Shopify switched it on for every merchant, and more than 100 companies have signed up as design partners, Salesforce among them.

Poppy answers 2 questions: who is this agent, and what did the customer allow it to do? That's real progress. An agent that used to look like an account attack now arrives with an ID and a permission slip.

Getting through the door still takes work. An agent can be stopped at the edge, at the API or on the page, and those controls often come from 3 different security vendors. Letting an agent past the CDN does little if a CAPTCHA catches it on the checkout page.

And then look at what happens once it's inside. When a bank customer asks their agent to send a payment, the bank runs the request through the account's approval policy before any money moves. The protocol carries the request. The policy decides.

In most B2C companies, that policy is the hard part.

Built for reps

I spent years working inside Salesforce, and I know where that policy usually lives.

Service Cloud and systems like it were built for reps. A rep logs a case. A service agent follows a screen flow and clicks Next. Every step assumes a human in front of a screen.

Over the years, the real business logic sank into that design. The discount rules sit in a flow. The eligibility check sits in Apex code. The refund limit sits in an integration nobody has touched since 2019. And the voice bot and the chat bot, often from 2 different vendors, each carry their own copy of the rules, slightly out of sync.

An agent that arrives through Poppy needs the process itself: what to check and when to bring in a human. Most systems today can only offer it a screen.

So it does what Instinct did with Daniel's espresso machine maker. It finds whatever works, and it keeps going until something gives.

Your own teams are adding to the pile

The same thing is happening inside the company. In Microsoft and LinkedIn's 2024 Work Trend Index, 78% of AI users said they bring their own AI tools to work.

At one company I know well, operations and sales each had their own criteria for provisioning a customer environment. Each team built a very capable Claude agent in an afternoon. Now provisioning runs through 2 different processes, and both of them look right to the people who built them.

That's 2 teams. Picture a company of 10,000 people, with every team shipping its own agent and its own version of the rules. You'd never put any of it in front of a customer, and you can't easily govern it inside either.

Don't wait for the framework

Poppy is one standard among several. MCP lets assistants like Claude call a company's tools. Google's A2A hands off long-running tasks from one agent to another. Shopify's Universal Commerce Protocol runs the cart and checkout. Meta has opened Muse connectors to developers. More will arrive this year, and a few will quietly fade.

It's tempting to wait and see which one wins. I think that's the expensive choice. Deloitte Digital lands in the same place: the standards are still evolving, and the capability layer underneath them carries over whichever way they settle.

That layer is your processes.

We call the goal headless. Each process is defined once, governed once, and callable by any agent or channel: a customer's Muse arriving through Poppy, an employee's Copilot, your own web chat, your own voice line. The same rules apply everywhere, and each request is checked and logged the same way.

Once a process works like that, plugging it into the next protocol takes days. The slow part, pulling the rules out of screens and getting everyone to agree on them, is already behind you.

What getting ready looks like

You can start with the process that hurts most and add one at a time. For each one, the work comes down to 4 steps:

  1. Find it. Map how the process actually runs across Salesforce and the systems around it, including the logic buried in flows, code and vendor bots.
  2. Define it once. Pull the rules into a single governed definition: what to check, what's allowed, who approves and what to tell the customer.
  3. Make it headless. Expose it to customers' agents, employees' tools and your own channels, with the same rules behind all of them.
  4. Count agent traffic separately. Deloitte warns that volume, abandonment and satisfaction scores will mislead you once agents call on customers' behalf: the same request can arrive 3 times through 3 channels, and an agent never fills in a survey.

None of this needs a new CRM, and none of it needs you to bet on a protocol.

The layer that matters most

The interfaces will keep changing. So will the agents, and the protocols they use to find you. The layer that runs your business processes is the one that lasts, and it's the one that decides whether Daniel gets his discount.

At Tribal, we've spent years working inside exactly that layer, in some of the most heavily customized Salesforce orgs around. ADAMA runs Salesforce across 19 countries, and its CIO says the team now moves significantly faster and deploys to production with confidence. A large telecom provider took a grid connection process that lived inside a screen flow, and its service reps now run it through Copilot, with customer channels next.

And a B2C manufacturer that started getting discount requests from Instinct now applies its own policy every time, whichever agent asks. That's how Daniel's weekend should have ended: the agent got its answer, and the company's own policy decided it.

Each of them kept the systems they trust. What changed is that each process now lives in one place, where every agent and every person runs it by the same rules.

Start before Monday

Daniel's espresso machine maker found out about personal agents from an awkward phone call on a Monday morning. You can find out about yours on purpose, before thousands of them show up at once.

Start with a free scan. In under 24 hours, Tribal scans your systems' metadata, maps every process and shows you which ones to make headless first, with your own agents, like Copilot, working on top.

‍

Book your Agent Readiness Scan here

‍

‍

‍

Read more

Onwards. With Tribal