Stop asking which AI agent to buy. The question that matters is simpler and harder: what work are you willing to let one do?
That question just got more urgent. On September 28, Meta launched Meta Enterprise Platform, a new business line that packages its Muse agent, Meta Business Agent and developer tools for companies. Meta joins Microsoft, Salesforce, ServiceNow, Google and a long list of startups already selling agents. Your inbox is about to fill up with demos.
Every one of those demos will show an agent doing something impressive. None of them will tell you whether that task belongs in your business. That part is on you.
Chatbots Answer. Agents Act.
The difference matters more than the marketing admits. A chatbot that gives a bad answer wastes a minute. An agent that takes a bad action like issues a refund, overwrites a customer record, changes a vendor’s payment details, creates a problem somebody has to find and fix.
So the real decision isn’t about technology. It’s about delegation. You already make these calls with people. You don’t give a new hire signing authority on day one. You watch their work, widen their scope as they earn it, and keep some decisions on your own desk.
AI agents deserve the same discipline. Most companies skip it because the demo made everything look easy.
What I Learned Watching Incidents at Scale
At Kyndryl, I ran operations on a Fortune 500 account where the pressure to automate never let up. Everyone wanted systems that fixed problems on their own.
We did the opposite first. We instrumented the environment so we could see what was actually happening. Incident detection improved 38%, and the account documented more than $2M in savings. We didn’t fix more incidents. We saw them sooner. Only then did it make sense to decide which fixes a machine could be trusted to make.
That sequence is the whole lesson. You can’t safely delegate work you can’t see, can’t measure and can’t undo. The five questions below are how I apply that lesson to AI agents today.
The Five-Question Test
Pick one specific workflow. Not “customer service” something like “respond to order-status emails” or “code incoming vendor invoices.” Score each question 0, 1 or 2. Be honest. The total tells you what to do next.
1. If the agent gets it wrong, can you undo it?
Drafting a reply is fully reversible. Delete it. Updating a CRM field is mostly reversible if you keep history. Sending money, deleting records or emailing a customer can’t be taken back.
Score 2 if the action is easy to reverse, 1 if reversing it is painful, 0 if it’s permanent.
2. How much damage can one bad action do?
This is blast radius. A wrong tag on a support ticket annoys one person. A wrong price pushed to your online catalog hits every customer who checks out before someone notices. Count regulatory and reputational damage too, not just dollars.
Score 2 if one mistake stays small and contained, 1 if it touches customers or money in a limited way, 0 if it can spread, compound or put you in front of a regulator.
3. How often does the work break the pattern?
Agents are strong on routine and weak on the unusual. If 95% of your order-status emails are “where’s my package,” an agent can handle the bulk and route the rest. If half your invoices need a phone call to sort out, the agent will spend its life guessing.
Score 2 if exceptions are rare and easy to spot, 1 if they’re common but recognizable, 0 if nearly every case is a judgment call.
4. Is the data it reads trustworthy?
An agent acts on what it can see. If your customer records live in three systems that disagree, the agent will confidently pick one, and you won’t know which. This is where most agent projects actually fail, long before anyone argues about models.
Score 2 if there’s one clean system of record, 1 if the data is decent but scattered, 0 if nobody on your team would trust it without checking.
5. Would you know it went wrong and who owns the fix?
This is the question vendors skip. Does the agent log what it did and why? Does anyone review those logs? When it makes a mistake at 2 a.m., whose phone rings? If the honest answer is “nobody’s,” the agent isn’t ready, no matter how well the demo went.
Score 2 if actions are logged, reviewed and owned by a named person, 1 if logging exists but nobody watches it, 0 if you’d find out from an angry customer.
Reading Your Score
8 to 10: Let the agent act. Keep a named owner and spot checks. Review a sample of its work every week for the first few months.
5 to 7: Let the agent draft, and have a person approve. This is where most useful work lands today. The agent does the reading, sorting and drafting. A human clicks send. You recover most of the time savings, and you build the track record that tells you when to widen its authority.
0 to 4: Don’t hand this to an agent yet. Fix what the test exposed first. A low score almost always points at data, visibility or ownership. Those are infrastructure problems, and no agent solves them for you.
One rule overrides the math. A 0 on question 1 or question 2 means a human approves every action, whatever the total says. A task can score well everywhere else and still be the one that costs you a customer or a lawsuit.
A Worked Example
Picture a 40-person distribution company weighing two agent use cases.
The first is answering “where’s my order?” emails. Drafts are reversible (2). A wrong answer irritates one customer (2). Most questions follow the same pattern (2). Shipping data comes from one carrier integration (2). But nobody has been assigned to review what the agent sends (1). Total: 9. Let it act, as soon as someone owns the weekly review.
The second is updating vendor payment details when a supplier emails a change. Payments are hard to recall (0). One fraudulent change can send a month of payables to a criminal (0). Real requests are routine, but fraudulent ones are built to look routine (1). The input is an email, the least trustworthy data you have (0). Nobody reviews changes before the next payment run (1). Total: 2, with zeros on both override questions. That task stays with a person who picks up the phone and verifies. Probably forever.
Same company. Same agent platform. Opposite answers. The technology didn’t decide. The work did.
Where to Start This Week
Pick the three workflows your team complains about most. Run each one through the test with the person who actually does the work, not the person who manages it. They know the real exception rate. They know which data they quietly double-check.
You’ll likely find one workflow that’s ready now, one that belongs in agent-drafts-human-approves, and one that exposes a data or ownership gap you needed to fix anyway. All three are useful outcomes. Only one of them requires buying anything.
The companies that get real value from agents won’t be the ones who adopted first. They’ll be the ones who knew exactly what they were handing over.
Find out which of your workflows are agent-ready. Our AI Infrastructure Assessment scores your data, integrations, security and governance, then gives you a prioritized roadmap before you spend a dollar on agents. Start the conversation →
See how we sequence AI work. From assessment through strategy, implementation and adoption, infrastructure comes first. Explore our services →
Keep reading. More practical frameworks for AI decisions that hold up in production. Visit the Summit AI blog →
Russell Love is the Founder & CEO of Summit AI Business Solutions, based in Browns Summit, NC. With 20+ years of enterprise transformation experience at IBM and Kyndryl, Russell helps businesses build the foundations that make AI actually work.