The one word I’d argue with in Contentful’s definition of agentic AI

Contentful published a definition of agentic AI on 24 August, and I agree with nearly all of it. An agent is given a goal, access to tools and memory, works on its own, and then, in their words, defers “to human operators for final decisions or approval”. Their governance section is better than most I’ve read this year: responsibility assigned to a named stakeholder, a checkpoint that pauses the final action until someone with context signs it off, an audit log underneath. If every vendor wrote that paragraph the category would be in better shape.

Ours is shorter. Agentic = describe the change in plain language; the platform ships it across every property, with approval and audit built in.

The one word I’d argue with is approval, and it’s sitting in both definitions.

What their approver is approving

Their worked example runs like this. An agent writes a product description, translates it into several languages, creates the image variants, then alerts a human team member for final approval. One product, one workflow, one person at the end of it. The thing being approved is the thing you can see: you read the description, you glance at the translations, you click.

Now take the buyer I spend my days with. Picture a franchise operator running 60 digital properties on a lean team of four. The instruction is “move the winter promotion to the top of every regional homepage, except the three regions where it’s still summer, and swap the hero image wherever the local store has its own”. One sentence. 60 places it lands, and it doesn’t land the same way in each.

Who approves that, and what are they looking at when they do?

A pause works when the plan fits on one screen

When there’s one item, the item is the plan. Pausing before publish and showing a human the result is exactly right, and I’d buy that product for that job.

When there are 60 items derived from one instruction, the pause has to show the human something, and this is where I think the word starts to hide things. If what they see is the instruction, they’re approving a sentence and trusting that the agent read it the way they meant it, 60 times over. If what they see is a summary, “promotion updated across 57 properties, 3 skipped”, they’re approving arithmetic. A review is when the regional lead for Queensland opens the plan and sees that her Cairns homepage keeps the summer hero and her Brisbane one doesn’t, before either page has changed.

I’m describing the word and the shape of the job here, rather than Contentful’s product, which I haven’t run at that scale and which their article doesn’t claim to cover at it.

The unit of approval is the property

So the approval we built runs per property. Before anything executes, the workflow produces a diff for every property it will touch: this page, this block, this image, from this to this. Approvals are configured by workflow type, by role and by property group, so the Queensland lead sees Queensland and whoever owns pricing sees anything that touches a price. If step three fails at property 31, the run stops and the partial state rolls back, rather than leaving 30 properties changed and 29 waiting for someone to notice. And every prompt, every plan and every change is logged with the user, the property and the time, because “who asked for this” is the first question anyone asks afterwards.

None of that is exotic engineering. It’s the middle part, between the instruction and the outcome, and it’s the part that gets skipped when the demo is one product description. We call the wider idea agentic operations, and the same boundary shows up in other stacks: Shopify Sidekick runs a multi-step task beautifully inside one store, and stops at the store line.

Five things to ask before you believe the word

Whatever platform you’re looking at, including ours, ask for these in the demo.

1. Show me the plan before it runs, one line per property.
   A good answer is a screen listing every property with the change it will receive. A weak one is “it applies the change everywhere you select”.

2. Who approves, and can that differ by property group and by what’s being changed?
   A good answer shows approval rules by property group and change type, in settings. A weak one is “an admin approves”.

3. Run it against 60 properties and make step three fail at property 31. What state am I in?
   A good answer: the run stops, completed steps roll back, and you see which property failed and why. A weak one is “you get a report of what succeeded”.

4. Show me the log for one change: the prompt, the plan, the user, the property, the time.
   A good answer is one record carrying all five. A weak one is “there’s an activity log”.

5. Run the same instruction twice. Show me the second plan is empty.
   A good answer: nothing would change, so there is nothing to approve. A weak one re-applies everything.

If a vendor can do those five, argue about price. If they can’t, the word agentic on the slide means the agent writes things and a person publishes them, which is a perfectly good product with a different name.

The honest objection

The honest objection is that most changes are one item. A blog post, a product page, a translation. For those, Contentful’s model is exactly right and a per-property diff is overkill you’d never open. I agree. The changes that hurt are the ones that fan out, and nobody gets to choose which kind they’ll need at four o’clock on a Friday before a long weekend.

So I’d keep their definition and add a clause. Agents act on a goal, with tools and memory, and defer to a human for approval, where approval means seeing what will change, and where, before it does. We’ve built that far. I expect the rest of the market to get there too, and that’ll be good for everyone who runs more than one property.

If you’re comparing platforms this quarter, take the five questions above into the next demo and see how far the word stretches.

Frequently asked questions

What does human in the loop mean in agentic AI?

It means the agent stops before the final action and waits for a person to approve it. The useful question is what that person is shown at the stop: the instruction they typed, a count of what happened, or the actual change queued for each place it will land.

Is agentic AI the same as generative AI?

No. Generative AI produces a draft when asked. An agentic system is given a goal, chooses its own steps, uses tools and memory, and acts. Contentful’s article draws the same line, and it is the reason approval matters more in the second case than the first.

What is a per-property diff?

A view produced before anything executes, showing every property the run will touch and the exact change queued for it: this page, this block, this image, from this to this. It is what makes the pause a review rather than a formality.

Who should approve an agentic change?

Whoever owns the thing being changed, which is rarely one person across an estate. Approvals that can be configured by property group and by change type let the regional lead see their region and the pricing owner see anything touching a price. See what separates an agentic CMS from AI features.