Introduction
Whenever I talk about autonomous commerce, one reaction comes up: This is great, but it feels like a long way off. It isn’t. It only feels that way if you try to make everything autonomous at once.
The hardest part is rarely the technology, but choosing the first thing to hand to an agent. Pick well, and you learn quickly on real data and build the trust everything else depends on. Pick badly, and you end up with a stalled pilot or, worse, an expensive mistake.
Here’s what I’ve learned from building our commerce platform and agent layer with my teams, and from seeing how enterprises across the industry are approaching autonomous commerce.
Note: This post draws on a conversation I had with Roman Tejada, AI Innovation Lead at Google Cloud. Watch the full conversation on demand.
Why does the first use case matter so much?
A prototype won’t teach you much. The real lessons only show up when an agent is actually running against your real data, where being wrong costs you something. So, one narrow agent in production teaches you way more than a broad pilot that lives in a sandbox.
Your first agent also matters because that’s where your people decide whether they trust this. And it needs to work without breaking anything; it has to earn that trust the hard way, through what actually happens.
What makes a good first use case for autonomous commerce?
I look for five things:
- It’s repetitive. Think of a decision people make over and over, with plenty of data on both sides, where an experienced person usually knows the right call but can only make it so many times a day.
- It’s reversible, with a small blast radius. If the agent gets it wrong, fixing it should be cheap and easy.
- There’s a clear lever to pull. The agent needs something it can act on. If a capability isn’t exposed as an API, then as far as an agent is concerned, it doesn’t exist; it will reason through the problem perfectly, then stop and queue the work for a person.
- Success is measurable. You should be able to say, with evidence, whether it worked.
- Nobody will fight you over it. The best first use cases are boring ones that people are happy to hand over.
What does that look like in commerce?
These examples come from the kinds of work we show our operational agents doing:
Many teams begin with assistive experiences like these and move toward autonomy as they get comfortable and confident. The journey runs from guided, to scheduled, to autonomous.
What shouldn’t you start with?
Pricing your top revenue SKU during peak season. Your first use case should be one where, if it goes wrong, the cost is small and the fix is easy. Save the high-stakes work for later, once you have evidence from the first few.
What does a good first use case look like in practice?
Roman shared an example I like, from outside software. A store kept missing its morning numbers. There was plenty of demand, but prep started at seven and customers started arriving earlier, so half the display was empty at the peak and people walked across the street to a competitor.
The fix was a schedule change. If it didn’t work, they could change it back the next day. The answer was already sitting in the data, but nobody had time to look at it.
That’s exactly the profile to look for: Measurable, reversible, low drama and real money. As Roman put it, the first one only has to get people to care and to build trust.
Put the guardrails in on day one
Even a small, boring use case needs a human in the loop, observability and an audit trail from the start. Don’t retrofit them later. If you’re not sure what to put in place, I’ve covered the questions to answer before any agent acts in production in Before You Let an AI Agent Act in Production: 7 Questions to Answer First.
How do you know when to let the agent do more?
Measure the outcome and let confidence build one step at a time. At first, your organization does the work and the agents help. Over time, you move toward agents doing the work, with your people stepping in at the moments that matter. It’s a journey of building trust and confidence by letting go one step at a time.
That shift, from AI-assisted humans to human-assisted AI, should be a decision you make on evidence. It isn’t something that happens while nobody is watching, which is why observability and outcome measurement matter so much.
Build it on your own data, and do it in weeks
Roman’s advice to technical leaders is to build on your own data first and decide afterward. You will learn more from a working pilot in a few weeks than from another strategy deck.
Nobody ever learned how to swim from slides. You have to jump in, get it done, choose something that’s valuable, sort of a low-hanging fruit in terms of effort, but it'll give you value and it’s reversible.
Roman Tejada
Innovation Lead, Cloud AI Transformation, Google Cloud
It also matters because your business is not the same as another enterprise’s, or a competitor’s. Each company has its own profile, and the best outcome is an agent that adapts to yours.
Missed the live session? Watch our webinar, Autonomous Commerce in Practice: From Tasks to Outcomes, on demand.
FAQs
A repetitive, measurable, reversible task with a small blast radius, where the answer is already in your data and the agent has a clear action it can take through an API.
Weeks rather than quarters. A working version running on your own data teaches you more than extended planning.
Yes. Put approval gates, observability and audit in from day one, then loosen them as you gather evidence.
An agent can only do what your systems allow, so the capabilities it needs must be available via APIs. Skipping the foundation makes everything harder.
That’s why you start with something reversible and low in blast radius, with approvals, monitoring and a way to stop the agent in an instant.
