Agent washing sales tools: five tests before you buy

Agent washing sales tools is the practice of relabeling a drafting feature, a fixed workflow or a chatbot as an "agent" without changing how the product works. Five tests separate the two: what changed in the product, who chooses the steps, what a run looks like, what happens when the agent is wrong, and what one finished output costs.

An analyst firm said in mid-2025 that only about 130 of the thousands of vendors claiming agentic AI were real. The competitor scan I read this week suggests the ratio hasn't improved. So here is how I'd test it, working through it the way I would on a call with a RevOps lead.

Why agent washing sales tools is hard to spot from the homepage

In the last two weeks I've watched a signal tool list a dialer and forecasting under its own product menu, a point tool for email coaching lead with a new "sales agent" while its page meta tags still say email coach, and an outbound vendor swap a rip-and-replace hero for "runs alongside your team". Some of those are real product moves. Some are copy edits. From the homepage, you can't tell which.

The reason is that "agent" has no spec, and that is what lets agent washing sales tools spread so easily. One widely cited engineering write-up draws the useful line: a workflow runs predefined steps, while an agent decides its own steps and tools as it goes. A workflow is fine. Plenty of RevOps work should be a workflow. The problem is paying agent prices for one.

Test one: the changelog diff

Ask what shipped, not what was announced. Pull the changelog, the docs and the pricing page for the 90 days before the rebrand, and compare them to the 90 days after. A real change leaves traces: new tool permissions, new objects it can write to, new limits, new pricing units.

If the changelog is quiet and the hero copy is loud, you've found the edit. A second tell is residue: old category names left in page titles, billing labels or help articles. Agent washing sales tools almost always leaves this kind of half-finished repositioning behind, and your reps will notice the gap too.

Test two: who picks the steps

Give the tool a goal, like "find accounts that match our ICP, check whether we already talk to them, and prepare outreach for review". Then see whether it chose the sequence or you configured it.

One correction to a common checklist: needing a human approval does not make a tool a fake agent. I build an approval-first product, so I'm biased, but the distinction matters. Approving the action (send this email, update this field) is a control. Being forced to approve every step of a plan the tool can't form on its own is a limitation. Ask which one you're looking at. Our page on autonomous versus supervised AI covers where we draw that line, and the approval gates explainer covers why.

Test three: the trace

Ask for one run, live, on your own data, with the tool calls visible. I call this the replay test. You want to see what it looked up, what it decided, what it wrote and where it stopped.

Then run the same input twice. A real agent may take a different path and should still land somewhere defensible. A relabeled feature takes the identical path both times, because it was a template. Neither is bad, but only one is an agent.

Test four: what happens when it's wrong

Every agent is wrong sometimes. The question is who finds out and how fast. Ask five things:

• Is there an audit log of every read and write, by object and field?
• Can you scope permissions per object, or does it hold one shared login?
• Can you undo a batch of writes?
• Who is alerted when it contacts the wrong person?
• Where does the log live, and can you export it?

Vague answers about "human oversight" are the answer. This is also where the data underneath shows up: an agent working on stale records fails confidently, and a vendor practicing agent washing sales tools will rarely volunteer that.

Test five: the bill per finished output

Seat price tells you nothing about an agent. Credit-metered tools make it worse: in the pricing I reviewed, one vendor's monthly credits do not roll over and work pauses when the pool runs dry, while another vendor's per-seat tiers sit next to a separate credit meter.

So ask for cost per completed output, such as one researched account or one follow-up ready to send, at your volume. I wrote up the arithmetic in forecasting your real monthly bill. For better or worse, the number from a vendor demo is usually the best case.

How to score agent washing sales tools in a pilot

The question most guides skip is how to turn these tests into pass or fail. Here is the scorecard I'd use. The thresholds are my assumptions, not benchmarks, so adjust them to your volume.

• Scope: 50 to 100 of your own accounts, in your real CRM, for two weeks.
• Completion: at least 80% of tasks reach a usable output without you rebuilding the setup.
• Accuracy: a person reviews a sample of 25 outputs, and 90% or more should be correct against your ICP and your records.
• Recovery: seed three deliberate errors and confirm each is traceable in the log within a day.
• Cost: price per accepted output stays within your target at full volume, not trial volume.

Fail any one and you've learned something cheap. A pilot has to run alongside your existing sales stack to be honest, since that's where the tool will live. Whether it still works on day 30 also depends on who owns it once reps adopt it.

I don't have a great answer for how to score a tool whose output quality drifts over months. A two-week pilot won't catch that. I'd put a recheck on the calendar at day 60.

What I'd do this quarter

If you're buying now, run test one on every shortlisted vendor before the first demo. It takes an hour and removes half the list. Run tests two through four in the demo, and test five in the pilot.

If you'd rather have someone run it with you, implementation help is worth picking on time to value, not on a deck. To see how we answer these five tests for our own product, how Coco works shows the approval flow end to end. Agent washing sales tools only works when buyers accept the homepage as evidence, so don't.