Every vendor now sells agents. Fewer are selling agents.
Gartner has a phrase for what's happening in the AI market: agent washing. It's the rebranding of existing products, such as AI assistants, robotic process automation (RPA) and chatbots, as "agents" without substantial agentic capability. In the same research that predicted over 40% of agentic AI projects will be cancelled by the end of 2027, Gartner estimated that only about 130 of the thousands of vendors claiming to offer agentic AI are the real thing.
That's not a reason to avoid AI. It's a reason to know how to test what you're being sold. Here's a seven-question checklist you can run on any "AI agent" or "AI OS" before you sign anything.
1. Does it work across your real systems, or just one app?
A chatbot inside a single product can't see anything outside it. A genuine agent connects to the places your business actually lives: your mailbox, your accounting software, your documents. Ask for the list of integrations and, more importantly, ask what it can do in each, read, write or both.
Red flag: "It integrates with everything via Zapier."
2. Can it handle messy, unstructured input?
Real business work arrives as emails, PDFs, photos and forwarded chains, not clean forms. RPA and scripted automations break when the input varies. Ask the vendor to show it handling an email it has never seen.
Red flag: the demo only uses perfectly formatted inputs.
3. Does it act without being asked every time?
An assistant waits for your prompt. An agent should notice that an invoice is overdue, a request has stalled or a renewal is coming, and prepare the next step on its own. Ask what it does overnight, when nobody is typing.
Red flag: everything starts with you opening a chat window.
4. Does it show its working?
If it tells you "4 of 31 open orders are stuck", you need to be able to check that. Look for exact counts, named sources and timestamps, and for plain admissions when it couldn't reach some data. An assistant that guesses confidently is worse than none, because you can't tell when to stop trusting it.
Red flag: fluent answers with no sources.
5. Does a named person approve what goes out?
This is the most important one. Anything that leaves the business, whether an email, a quote or an update to a client, should wait for a named, accountable human. Ask: can it send without approval? Whose email address does it send from?
Red flag: "It can run fully autonomously", presented as a selling point.
6. Is there an audit trail?
When something goes wrong, and eventually something will, you need to know who or what did what, and when. Ask to see the log, and ask whether it covers AI actions and human approvals alike.
Red flag: no log, or a log you can't export.
7. Does it start with a problem you actually have?
Gartner describes most agentic projects as early-stage experiments driven by hype and often misapplied. Ask the vendor for the three problems it solves best, then check whether any of them is on your list. A platform that can do everything often proves itself at nothing.
Red flag: the pitch is about the technology, not your work.
How an AI OS should score
We built Clara against this same list, because they're the questions we'd ask as buyers.
| Question | What it looks like in Clara |
|---|---|
| Real systems | Connects to Microsoft 365, Gmail and IMAP, Xero and Sage, OneDrive, Google Drive and Dropbox |
| Unstructured input | Turns incoming email and attachments into tracked work |
| Acts unprompted | Inspects the whole business every night and reports each morning |
| Shows its working | Exact counts, named sources, timestamps, and an admission when data is missing |
| Human approval | Mandatory, with a named person; sent from their real address |
| Audit trail | Every record, document and approval logged with who, what and when |
| Real problem | Starts with the admin that crosses your tools: requests, orders, invoices, follow-ups |
You shouldn't take our word for it. That's the point of the list. Bring it to the demo and test us.
The short version
Agent washing thrives when buyers can't tell the difference. The seven questions above make the difference visible in about fifteen minutes. If a vendor passes, you've found something real. If not, you've saved yourself from becoming part of the 40%.
For more on what separates a real AI operating system from the rest, read What Is an AI OS Platform? and our take on why so many agent projects get abandoned.