Agent Washing in AI Vendors: How to Spot It Before You Sign a Contract

Agent washing in AI vendors is the practice of relabeling an existing chatbot, rules engine, or robotic process automation (RPA) tool as "agentic AI" without adding the autonomous reasoning, planning, or decision-making that term is supposed to describe. The fastest way to spot it is to ask one question in the sales call: "What decision does this system make on its own, and what happens when it's wrong?" If the answer is a vague gesture at "smart automation" rather than a specific decision loop with defined guardrails, you're likely looking at a relabeled product, not a new one.

This isn't a pedantic distinction. Gartner — the research firm that coined the term — estimates that of the thousands of vendors now marketing "agentic AI," only about 130 have built anything that meets the bar: software that can pursue a goal, choose among actions, and adapt without a human scripting every step. Gartner also forecasts that more than 40% of agentic AI projects will be scrapped before the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls as the leading causes. For a small business weighing a five- or six-figure automation contract, that's not an abstract industry statistic — it's the exact failure mode you're trying to avoid.

What Agent Washing in AI Vendors Looks Like

Agent washing usually takes one of three forms:

1. The relabeled chatbot. A conversational interface that answers questions or routes tickets gets rebranded as an "AI agent" because it uses a large language model somewhere in the pipeline. It isn't making decisions — it's retrieving and formatting text.
2. The relabeled RPA script. A deterministic, rules-based workflow ("if field A is empty, send email B") gets an AI-sounding name. It's reliable, but it isn't adaptive, and it can't handle a case outside its scripted path without breaking.
3. The demo-ware agent. A genuinely agentic system exists, but only in a narrow, cherry-picked demo scenario. In production, with your real data and real edge cases, it either needs constant human correction or fails silently.

None of these are inherently bad tools — a well-built RPA script or a tightly scoped chatbot can be exactly the right fix for a narrow problem. The issue is paying an "agentic AI" price and setting "agentic AI" expectations for something that can't actually operate independently.

Why It's a Real Budget Risk, Not Just Marketing Spin

The financial exposure compounds in three places. First, cost: agentic platforms are typically priced on the assumption that they'll replace significant human decision-making time, so if the system can't actually make those decisions, you're paying an inflated price for narrower automation. Second, scope creep: when a "demo-ware agent" hits a case it can't handle in production, the usual fix is more consulting hours, more customization, and a slipping go-live date. Third, and most damaging, is the opportunity cost — the months spent integrating a system that was never going to deliver the autonomy you were sold, instead of fixing the underlying process through a systems audit or buying a narrower tool that actually fits the job. This isn't unique to any one vendor category: McKinsey's 2026 survey of over 1,700 executives found only 6% of organizations qualify as AI "high performers," and unrealistic claims about what a tool can decide on its own are a direct contributor to that gap.

A Vendor Vetting Checklist for Small Businesses

Before you sign, work through these in order:

1. Ask for the decision loop, not the demo. Have the vendor describe, in plain language, what input the system receives, what it decides, and what output or action results — for your actual process, not a generic one.
2. Ask what happens on an exception. Every real agentic system has a defined behavior for cases it wasn't trained or configured for: escalate to a human, flag for review, or decline to act. "It just handles it" is not an answer.
3. Ask for a reference customer with your transaction volume. A system that works at 50 transactions a day may fall apart at 5,000. Ask specifically what changed when their reference customer scaled up.
4. Ask who's liable for a bad autonomous decision. If the vendor's contract is silent on this, assume the liability is yours.
5. Price the human fallback into your ROI model. If the system will still need a person to review or correct a meaningful share of outputs, your ROI case should be built on that reality, not the vendor's best-case pitch.
6. Request an audit trail. You should be able to see, after the fact, exactly what the system did and why — not just that it "completed the task."

Red Flags vs. Green Flags

Red flags: the sales deck leads with "agentic AI" but the technical demo shows a static decision tree; pricing is per-seat rather than per-outcome; the vendor can't name a specific exception-handling behavior; references are all logos, no metrics.

Green flags: the vendor can show you a case where the system changed its approach mid-process based on new information; they're upfront about what it can't do yet; pricing ties to measurable outcomes (tickets resolved, hours saved, errors caught); they offer a scoped pilot with a defined success metric before a full contract.

Put the Vetting Checklist Into Your Contract

A vendor conversation is only as good as what ends up in writing. Once you've worked through the checklist above, carry the answers into the contract itself: name the specific decisions the system is authorized to make, define the exception-handling behavior you were promised, and attach the pilot's success metric as a condition of the full rollout rather than a side conversation. If a vendor is reluctant to put their verbal claims about autonomy into a written scope of work, treat that reluctance itself as a signal — a genuinely agentic system should be describable precisely enough to write down.

Common questions

Is every vendor using the term "agentic AI" doing agent washing? No. A growing number of vendors have built genuinely autonomous systems — the issue is that the term currently has no enforced definition, so both real and relabeled products use it. The vetting checklist above is how you tell the difference for your specific use case, rather than taking the label at face value.

Does agent washing mean the underlying tool is low quality? Not necessarily. A relabeled RPA script or chatbot can still be a good fit for a narrow, well-defined task — the problem is pricing and expectations, not the tool itself. If a vendor is transparent that their product is rules-based rather than autonomous, and prices it accordingly, that's not agent washing; that's accurate marketing.

How long should a pilot run before committing to a full contract? Long enough to see the system handle a real exception, not just the happy path — for most small businesses, that's 4–8 weeks of live use on actual transaction volume, with a predefined metric (error rate, time saved, cases escalated) agreed before the pilot starts.

What should we do if we think we already bought a relabeled tool? Audit what it's actually deciding versus executing on a script, using the questions above, and compare that against your original contract's claims. If there's a significant gap, that's worth a structured conversation with the vendor about re-pricing or scoping down — and worth building into how you evaluate the next tool.

Vetting an "agentic AI" vendor is really a process-mapping exercise in disguise: you can't tell whether a system's decision-making matches your needs until you've documented what the decision actually requires. If you want a second set of eyes on a vendor contract or a workflow before you sign, start with an audit.