Is Your Vendor Selling You a Real AI Agent? A Buyer’s Checklist for Indian SMBs

Short answer: “Agent washing” is the practice of rebranding an existing chatbot, script, or RPA tool as an “AI agent” without adding the capability the word implies. It is widespread enough that Gartner put a number on it, and it is the reason a lot of automation projects quietly die a year in. You do not need to understand the technology to avoid it. You need about six questions, and the confidence to keep asking until you get plain answers.

What is agent washing?

It is marketing language running ahead of the product. Gartner coined the term for vendors rebranding existing AI assistants, robotic process automation, and chatbots as agentic without substantial agentic capability, and estimates that only around 130 of the thousands of vendors making the claim are the real thing (Gartner). The word “agent” now sells, so it gets applied to things that merely follow a script.

How big is the problem, really?

Big enough that the analyst house expects most of the current wave to be abandoned. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls (Gartner, June 2025). It has not eased since: in May 2026 the same firm issued a fresh warning about agent washing in the supply-chain planning software market specifically (Gartner, May 2026).

Why does this matter more for a small business than a large one?

Because you feel the whole cost and you rarely get a second budget. A large enterprise running twelve pilots can write off four of them as the price of learning. A twenty-person company that spends six months and a meaningful share of its cash on one project does not get to call that a portfolio. The asymmetry is not technical, it is financial, and it means the vetting has to happen before the contract rather than during the retrospective.

What is the actual difference between a chatbot and an agent?

An agent decides and acts; a chatbot responds. The practical test is whether the thing can take a goal, choose its own steps toward it, use tools or data on the way, and recover when something does not go to plan. If every path it can take was drawn by a human in advance, you have a workflow with a language model on the front, which may be exactly what you need, but it should be priced and described as one.

Capability	Scripted bot	Genuine agent
Chooses its own steps	No, follows a fixed tree	Yes, plans toward a goal
Uses tools and live data	Usually a fixed lookup	Calls systems as needed
Handles the unexpected	Falls back to a default reply	Retries or escalates deliberately
Leaves an audit trail	Message logs only	Actions and decisions logged
Honest label	Automation, and often the right answer	Agent

What questions should I ask a vendor before signing?

Ask questions whose answers cannot be dressed up, which means asking about mechanics and money rather than capability. A vendor with a real product answers these quickly and without reaching for adjectives; a vendor who is washing will convert every one of them into a story about transformation. The six below are the ones we would ask on your behalf.

• What decisions does it make on its own, and which ones always stop for a human?
• Show me it failing. What happens when a customer asks something out of scope?
• What does it cost per conversation at my volume, not per seat per month?
• Who owns the data and the configuration if I leave in six months?
• What did the last three customers of my size actually switch off after go-live?
• Which parts are your product, and which are a wrapper around someone else’s model?

What does a straight answer look like?

A straight answer contains a limit. “It handles order status, returns, and delivery timelines, it does not handle refunds without approval, and roughly one conversation in six still reaches a human” is a real answer from someone who has run the thing in production. “It handles everything, it learns your business automatically” is a brochure. The presence of a boundary is the strongest single signal you get in a sales call.

How should a real agent be priced?

In a way that moves when your usage moves, because a genuine agent has a genuine variable cost underneath it. Language models bill by token, so anyone running a real agent for you knows roughly what a conversation costs them, and a vendor who cannot answer that question either has no agent or has not looked. Flat per-seat pricing is not automatically dishonest, but it is a strong hint that what you are buying is software with fixed behaviour rather than a system that thinks per request.

Pricing model	What it usually signals
Per seat, per month	Traditional software. Fine, if it is described as software.
Per conversation or per resolution	The vendor knows their unit economics and is confident in them.
Setup fee plus usage	Common for genuine builds, ask what happens to the build if you leave.
Percentage of savings	Attractive, but check who defines and measures the saving.

Is an agent always the right answer?

No, and a vendor who says otherwise has told you something useful about themselves. Plenty of the highest-return automation we build is not agentic at all, it is a deterministic workflow that moves data between two systems on a schedule and never surprises anyone. Deterministic is cheaper, easier to audit, and fails in predictable ways. Reach for an agent when the work genuinely requires judgement over unpredictable inputs, not because the word is in the budget line.

What should I do next?

Take your shortlist and put the six questions above to each vendor in writing, then compare the shapes of the answers rather than the promises. The exercise costs you an afternoon and it is the cheapest risk control available to you, because the failure mode Gartner is describing is not a technology failure, it is a purchasing one.

We are an agency, so treat this as interested advice and apply the same questions to us. If you want a second opinion on a proposal already on your desk, book a free 15-minute call and we will tell you what we would push back on, the read is yours whether or not you work with us. For the wider decision see our honest comparison of building in-house versus hiring an agency, and why most SMB AI pilots never reach production for what goes wrong after the contract is signed.