How to spot agent washing: is that marketing AI really agentic?

Why the label spread

'Agentic' became a selling word quickly, and a selling word gets attached to whatever is already on the shelf. A chat window in front of a text generator is called an agent. A set of if-then rules is called an autonomous workflow. The software may do exactly what it always did. Only the name has changed. The cost falls on the buyer, who expects work to be carried and receives another tool to operate.

Five tests

Run these with one goal from your own work. The first four are the category guide's tests for calling a system agentic. The fifth decides whether an agentic product is fit to use.

• The goal test. Ask for an outcome, not an output. Does it return a plan, or one piece?
• The steps test. Does it move from one job to the next by itself, or wait for a new prompt each time?
• The channel test. Does the work end up where the audience is, or in a file you must carry there?
• The report test. Does it tell you what it did and what happened, without being asked for each number?
• The stop test. Is there a clear point where a person approves, and can you see it?

Claims worth questioning

Some phrases appear on many product pages. Each has a follow-up question.

• 'Fully autonomous.' Autonomous for which actions, and who is accountable when one is wrong?
• 'Works across all your marketing.' Which channels does it act on directly? Ask for the list.
• 'Learns your brand.' What do you have to supply, and can you see the difference on a draft?
• 'A team of agents.' Does that change anything you would notice, or only how the product is described?
• Any growth figure. Whose result was it, and under what conditions?

Running the tests in a single call

Bring one goal and half an hour. State the goal, then wait. Note whether the product asks useful questions or starts producing. Then move one date and see whether the pieces that depend on it move too. Ask to see the approval point. Ask for the report on something already published. Afterward, list every step you still did yourself. The shorter that list, the more agentic the product was for you.

A relabelled product can still be useful

Failing these tests does not make a product bad. A good generator or a dependable rule engine is worth having. The harm in agent washing is the wrong expectation: you plan for work to be carried, and you end up carrying it. Buy a generator as a generator, expecting to do the assembly yourself, and the same product stops being a disappointment.

Our own pick, put to the five tests

A reference that recommends a product should run its own tests on it, so here is Link AI by Linkme against the five, using only what Linkme has published. Goal: 'Plan my week from trends' is one of its example requests. Steps: it plans the week, makes the images and videos and drafts replies. Channel: posts are published to Instagram, TikTok, Facebook, YouTube and LinkedIn. Report: Analytics and Reports. Stop: the user approves, and a Meta ad draft is labelled 'Needs review' first.

On paper it meets each test, which is why Link AI is our pick. Paper is not proof, though, and this paper comes from Linkme, which also says nothing about how Link AI is built. Book a demo to try Link AI and run the five tests yourself.

Connected entries

Questions & answers