Every platform says it's 'AI-powered' now: here's how to actually evaluate one

The phrase now covers a text box that drafts subject lines and a system that can allocate your budget overnight. Same two words, two very different purchases. You probably got here after searching for the best AI marketing tools and finding ranked lists of AI marketing tools, none of which explained how to tell those two apart.

This isn't a ranked list. It's a way to tell the three kinds of AI apart, a decoder for the claims you'll read on every website, and the questions that reveal which kind any given AI marketing platform actually has. It works on full suites and on narrow AI tools alike, because it tests the claim rather than the category.

Why 'AI-powered' stopped meaning anything

Every platform in this category added AI language at more or less the same time. Some shipped real capability. Others took features that had been running on rules and regression for years, changed the button colour, and 'artificially' called it intelligence. Both groups now describe themselves in identical words, and that's the problem.

No shock, the money's moving faster than the judgement. Gartner's 2026 CMO Spend Survey of 401 marketing leaders found CMOs putting 15.3% of budgets into AI while only 30% rated their AI readiness as mature, and around 70% admitted their own processes weren't mature enough to scale it. Budget reaches AI marketing tools well before the ability to judge them does.

There's a structural reason too. The genre that's meant to help buyers compare AI tools is the ranked list, and ranked lists reward breadth over interrogation. A roundup of the best AI marketing tools has too many products and too little room for each one. Enough to describe features, nowhere near enough to test them. So the claims go in unexamined, over and over, and everyone starts to agree with them.

None of this means the category's hollow. Plenty of AI marketing software does exactly what it says, and plenty of AI marketing genuinely works. It means the claim has stopped being evidence, and now whoever signs the contract has to be more informed. The AI tools haven't got worse. The language around them has.

The three kinds of AI in a marketing platform

Most articles about the best AI marketing tools sort by function: email tools, content tools, analytics tools. That’s rather useless for evaluation, because it tells you what a product touches rather than what it's allowed to do without you. Sort by autonomy instead.

Assistive AI helps you work faster

Assistive AI shortens a job you were going to do anyway. You ask, it produces, you decide whether to keep it.

The eCommerce version goes something like, you need a batch of subject line variants for a sale, and what used to take an afternoon takes minutes. Content creation is the obvious win. The real one is the tedious middle of the job, the repetitive marketing tasks that produce nothing anyone remembers. Product descriptions. Category copy. Headlines for landing pages that have gone stale. The content marketing calendar. Metadata!

Content creation of this kind is now genuinely good, and the generative AI tools behind it have come a long way. Video generation has landed in enough platforms to be worth asking about. Where generative AI still struggles is anything needing taste, which is why the marketing content wants an editor rather than a rubber stamp.

What assistive AI doesn't do is decide. It has no opinion on whether the sale should run, who gets it, or when. Nothing reaches a customer unless a person puts it there. If a vendor's entire AI story lives here, that's not a scandal, but know it's the cheapest tier to build and the most heavily marketed.

Predictive AI forecasts what customers will do

Predictive AI reads history and scores the future: churn risk, lifetime value, likely next order date. Underneath is machine learning trained on your own transaction and behavioural data, which is why it can't work on day one.

The eCommerce version will feel like every contact carries a churn score, and the ones crossing a threshold drop into a win-back journey on their own. That's predictive AI doing the thing humans are worst at, spotting a pattern across your whole database at once. Customer segmentation stops being rules you maintain and becomes scores that update themselves, and the customer segmentation you end up with is usually one no human would've drawn.

This is where data analytics and prediction get blurred in demos. Data analytics tells you what happened last quarter. Prediction tells you what happens next, and the machine learning behind it should come with a statement of confidence. Good predictive analytics admits when it's uncertain, which separates serious AI marketing tools from dashboards with a new label. If a vendor shows you data analytics and calls it predictive, the tell is that everything on screen is in the past tense.

Again though, what it doesn't do is act. A score's an input. An automation you built decides what happens to a customer the model rates high risk rather than low.

Autonomous or agentic AI decides and acts

Autonomous AI, sold as agentic AI or as AI agents, takes an outcome and goes after it. It chooses, executes and adjusts without being asked each time.

The eCommerce version is shaping up like you set a revenue target and the system picks the segments, the creative and the send time, then reallocates towards whatever's working mid-campaign. Nobody approves each step. This is marketing automation in a rather different sense from the marketing automation you already know, where you drew the flowchart and the software did as it was told.

What it doesn't do, in almost every product on the market, is operate without boundaries you set. Remember this when you’re comparing; most platforms have the first type, many the second, and the third is where vendors differ enormously while using identical language. Two products both calling themselves agentic can sit at opposite ends of a spectrum, one drafting a campaign for approval and the other spending money on its own judgement. No list of the best AI marketing tools will tell you which, because the feature name is the same in both.

What six common AI claims actually mean

Every phrase below turns up in AI marketing tools sold to eCommerce teams, and in every roundup of the best AI marketing tools you've read this year.

AI-powered personalisation

Could mean: a model choosing, per person, in the moment, from your full catalogue, which product to show and which message to attach.

Usually means: rule-based customer segmentation with a recommendation widget added on. A handful of segments, a handful of variants, conditions somebody wrote once and never revisited them.

Ask: does it decide per person, or pick from variants I built? Then ask how many genuinely different things your customers would end up seeing. If it's a handful, you've got rules with a new name.

Predictive analytics

Could mean: machine learning working from your own history, producing scores you can use as triggers and filters, with stated accuracy and confidence.

Usually means: RFM analysis and lead scoring. Useful, well understood, and decades old.

Ask: what's it predicting, on what data, and how far back does it need? Vendors who know their models answer in specifics. Vendors who don't keep it vague and hope you'll move on. Any predictive analytics that works before you've given it history is a rules engine with better marketing.

AI agents

Could mean: a system that executes multi-step work towards a goal, choosing its own steps.

Usually means: a chat interface running commands you type, or a workflow that drafts several things and stops. Both useful. Neither autonomous.

Ask: what can it do without asking me, and what can it never do? The second half matters more. A vendor who can't name what their AI agents are forbidden from doing hasn't thought about the controls, which is a rather larger problem than the answer they're avoiding.

Generative AI built in

Could mean: generation grounded in your assets, writing from your brand voice, your live product feed and your campaign history.

Usually means: a general-purpose model in a sidebar. The thing you've already got open in a browser tab, with their logo on it.

Ask: does it know my catalogue and my brand voice, or is it a chatbot in a panel? Test it there and then. Ask for a product email for a specific SKU, then check the price and the specification. Generative AI that can't see your data will invent both confidently.

Self-optimising campaigns

Could mean: continuous reallocation towards a defined objective, with that objective visible and editable.

Usually means: automated A/B testing that picks a winner on open rate.

Ask: what's it optimising towards, and can I change it? Open rate and revenue produce very different behaviour. If nobody in the room can name the objective function, nobody in the room knows what the software does.

Real-time

Could mean: an event reaches the profile and changes what a customer sees within seconds.

Usually means: a batch job that runs every fifteen minutes. Sometimes overnight.

Ask: real-time from what event, to what action, measured how? A cart abandonment that fires while the customer's still on site and a segment that refreshes overnight are both sold as real-time. Only one of them catches anybody before they close the tab.

The questions that get you a straight answer

Does it decide, or does it recommend and wait for you?

Ask this first, about every AI feature they've named. Almost every vague claim in AI marketing collapses into clarity the moment somebody answers it.

A system that recommends produces something and stops. A person reviews, edits, approves, and only then does a customer see it. A system that decides acts on its own judgement and tells you afterwards, if it tells you at all.

A good answer comes back immediately, because a vendor who's built either one knows which. "It drafts the campaign and drops it into the editor for your team to approve" is a real answer. So is "it reallocates spend inside your guardrails, and you get a log."

A dodge sounds like flexibility. "It can do both, depending how you configure it." Usually that's a way of skipping the follow-up: what's the default, and who can change it? The default is what'll happen once the novelty's worn off and the person who set it up has moved on.

Your customers have the same instinct, incidentally. In a Gartner survey, willingness to let AI make the actual purchase topped out at 11%, and that ceiling was in the lowest-stakes categories. Letting AI narrow the choices ran to 31%. Recommend and wait, in other words.

What it decides

Get specific about scope, because most AI marketing tools are vague here by default. Which decisions are in play: audience, creative, timing, channel, budget? Can it change a live campaign, or only build new ones?

Then oversight. Who signs off before an autonomous action goes live, is that person named in the system, and is the approval gate configurable per action type? You want a different threshold for a test send than for your entire database. Ask what the audit trail records, whether it survives a staff change, and whether there's an off switch a marketer can reach without raising a ticket.

What data it needs

Most disappointment with AI marketing tools traces back here rather than to the model. Which customer data does each feature need, and what happens when it's patchy? How long before a predictive model's worth trusting? What was it trained on: your account only, or pooled across the vendor's client base?

Then integration. How does this sit in your tech stack, and how much of the plugging in turns out to be your engineering team's problem? AI workflows that depend on customer data you don't collect aren't features, they're projects you're paying for. Ask which parts of your tech stack the AI workflows read from and which they write back to, because a system that can't write back is a reporting tool with ambitions.

What happens when it gets it wrong

Ask for a failure story, not a hypothetical. Every vendor with real deployments has one, and the willingness to tell it is itself the answer.

How do you find out, and how fast? Can you roll a decision back, or only stop the next one? If a model drifts, does anything alert you, or does campaign performance quietly sag for a quarter until somebody goes looking? Who's accountable when an autonomous action does damage, and is any of that in writing, or is it all in the relationship?

What it costs to run

Pricing for AI tools here has moved fast, so be blunt about it. Are AI features included, tiered, or metered, and if metered, by what unit? What does a heavy month look like against a normal one? Software that charges by AI action can be excellent value and can also produce a bill nobody forecast.

Then the honest total. What internal time does this need every month, and who owns it after they leave? Marketing teams underprice this consistently, and a platform that needs a specialist you haven't got costs considerably more than its price tag.

The same logic condensed, for the claims you'll hear most often.

"Web events hit the profile in under two seconds. Predictive scores recalculate hourly. Here is the difference on a live account."

Red flags worth noticing in a demo

None of these prove a vendor's overselling, and several turn up in demos from products that belong on any list of the best AI marketing tools. Three in one demo, though, and you've learned something.

1. The demo runs on immaculate sample data. Perfect names, complete purchase histories, nobody who bought once and vanished. Ask what happens when a good share of your profiles have no purchase history at all. The hesitation before the answer tells you more than the answer does.

2. Nobody can say what the model was trained on. Not the architecture, which you don't need, but the data, which you do.

3. 'The AI figures that out' arrives in response to a direct question. The most common tell in the category, and the most effective, because it reframes an unanswered question as a feature and most people let it go.

4. Results get quoted without a baseline. An uplift against what, measured over what period? Anyone comparing the best AI marketing tools should ask every time.

5. Everything is described as AI. When the reporting module, the form builder and the send scheduler have all been rebranded, the word has stopped meaning anything at all. Some of the best AI marketing tools are noticeably conservative about what they call AI, and that restraint is usually a huge green flag.

6. 'Actionable insights' does a great deal of work and no question gets answered. Ask what action, by whom, on what screen. If nobody can walk it end to end, there are no actionable insights.

7. Integration gets waved through. AI tools that assume a clean tech stack tend to discover yours during implementation rather than during the sale, which is a discovery you pay for.

8. Nobody can name a limitation. Every serious vendor of AI tools knows where their product's weak, and the best AI marketing tools tend to come from the ones willing to say so, because it buys credibility for everything else.

What to test before you sign any contracts

Run a pilot on your own data. The gap between a demo and a deployment shows up precisely where a model trained on clean examples meets a catalogue that's been edited by everyone who's ever worked there.

Agree success criteria before you start, in writing, with numbers. Not "see how it goes". Pick two metrics, name the baseline, say what result makes this a yes.

Give predictive models runway. A churn model needs history behind it and time in front of it. A few weeks tests the onboarding, not the capability. Give it a full quarter and judge campaign performance at the end.

Test the failure path on purpose. Feed it a segment with sparse data. Approve something you'd normally reject and see whether anything flags it. Try to roll an action back. Point the content generation side at your worst landing pages.

Put one real person on it who isn't the champion. Ask whoever has to use the thing whether it saved them time or just moved the work elsewhere. 

A pilot also does something no article can, including this one. It tells you which of the marketing strategies you already run would genuinely change if the AI marketing tools were better, and which wouldn't shift at all. Plenty of marketing strategies underperform for reasons no model can reach, and it's worth knowing which of yours those are before you sign.

Where Manago AI fits

Manago AI (previously SALESmanago) runs AI across five types: predictive, agentic, generative, recommendation and conversational. The structural point is that AI here is the execution layer on top of a customer data platform and marketing automation core, not the product itself. The scores and the drafts are only ever as good as the unified customer data underneath them, which is why the CDP comes first in the architecture.

Having spent an article handing you questions, it'd be poor form not to answer them ourselves.

Does it decide, or recommend and wait? It recommends and waits. The documented behaviour of our agentic AI is that it proposes, drafts and assembles, and your team reviews and approves before anything goes live. Nothing auto-launches. Every output, whether that's a segment, a campaign or a recommendation, lands in the editor fully editable, and you can reject it, edit it or replace it. There's an approval gate before every send.

That's a deliberate position rather than a missing feature. By this guide, that puts our AI agents closer to a capable assistive and predictive system than to autonomous spend allocation.

What data does it need? Predictive scoring, recommendations and generative AI output all run on live account data: your contacts, campaigns, segments and performance history, not generic patterns. Predictive AI and the marketing automation around it improve as that history builds, which is also why customer engagement tends to lift through sharper targeting rather than through more volume.

Questions buyers keep asking

Closing

There'll be another demo next week, and another rep promising more or less what the last one promised.

What's changed is that you can sort the claims into three buckets while the person is still talking, and that you've got one question that cuts through most of it: does this decide, or does it recommend and wait for me? Ask it early, and enjoy the show.

Good AI marketing tools survive that question comfortably. So do honest vendors selling modest capability, which is a perfectly reasonable thing to buy. The ones that don't survive it were never going to fit your marketing strategies anyway, and you've found that out before signing rather than after.

'AI-powered' will tell you nothing. What the software is allowed to do without asking you will tell you everything. If you want to see what that looks like in a platform where the approval gate is the default rather than a setting, that's what we've built.