What is AI marketing? A two-question test for UK firms

Short answer: AI marketing is software that does marketing work by learning from data and acting without step-by-step instructions. The UK government’s own description of AI gives you two questions to put to any tool: does it learn from data, and does it act without a person directing each step? Some tools sold as AI fail one or both. That can be fine, but do not pay an AI premium for it.

The label “AI” is now stuck to a lot of a marketing stack, from email schedulers to chat widgets. Some of it deserves the label. A lot of it is the same rule-following software you have used for years, with a new name on the pricing page.

You do not need a technical background to tell the difference. You need two questions, and the willingness to ask them of the tools you already pay for before you buy anything new.

This page sits under our guide to AI marketing for UK small businesses. For how widely UK firms have adopted these tools, see our AI marketing statistics for the UK.

Where the UK government draws the line on what counts as AI

In March 2023 the UK government presented a white paper to Parliament, A pro-innovation approach to AI regulation. In section 3.2.1 it says there is no widely agreed definition of AI. Rather than invent a rigid one, it describes AI by two characteristics.

The first is adaptivity. These systems are trained, once or continually, and they work by inferring patterns and connections in data that humans often cannot easily discern. The paper notes this can make their intent or logic hard to explain.

The second is autonomy. Some AI systems can make decisions without the express intent or ongoing control of a human. The paper notes this can make responsibility for outcomes hard to assign.

Two clarifications keep this honest. The white paper is a policy paper, not a law. And it described an approach under which existing regulators, rather than a new AI regulator, apply five cross-sectoral principles: safety, security and robustness; appropriate transparency and explainability; fairness; accountability and governance; and contestability and redress. This describes the 2023 approach, so check for later changes before relying on it.

For a small business the useful part is the two characteristics. They turn a vague word into something you can test.

Question one: does it learn from your data?

Adaptivity, in plain terms, asks whether the tool changes its behaviour because of what it has seen. Does it find patterns in data and adjust, or does it do exactly what it was told yesterday and will do the same tomorrow?

A few things to look for:

• Training. Was it trained on data, and does it keep learning from yours, or is it a fixed set of rules?
• Inference. Does it work out something you did not tell it, such as which enquiries look like past customers?
• Improvement. Does it get better with more data, or does it behave identically after a year?

A scheduled email send follows a time you set. It does not learn anything. A decision-tree chatbot follows scripted branches that someone wrote in advance. A large language model writing assistant is different: it generates new text from patterns in its training data. Whether that counts as learning from your data depends on whether your own material shapes what it produces.

Be precise here. Learning from data in general is not the same as learning from yours. A writing assistant trained elsewhere may pass the adaptivity test in the white paper’s sense without ever touching your customer records. Both can be true, and it is worth knowing which one you are buying.

Question two: does it act without you pressing go?

Autonomy asks who makes the move. Does the tool decide and act, or does it suggest and wait for you?

A tool that lists ten keyword ideas and leaves you to choose is advising. A tool that raises and lowers your bids across thousands of auctions without anyone approving each change is acting. The white paper’s wording is that some systems can make decisions without the express intent or ongoing control of a human.

Autonomy comes in degrees. A drafting tool that writes a reply and waits for you to send it sits near the advising end. A tool that sends the reply itself sits near the acting end. Most real tools have a setting somewhere between the two, and it is worth finding it.

It is also possible to be autonomous without being adaptive. An autoresponder fires the moment an enquiry arrives, with nobody pressing anything. It acts on its own, but it does so by a fixed rule and learns nothing. That is why we ask both questions and do not stop at one.

Running the test on tools a UK business already pays for

Here is the test applied to generic tool types. The verdicts are our judgment, not a ruling, and a specific product may differ from its category, so check the one you pay for.

Tool type	Learns from data?	Acts without you?	Our verdict
Rule-based email autoresponder	No, same reply to every enquiry	Yes, once the rule is set	Automation, not AI. Fair at a low price
Scheduled email or social send	No, follows a time you set	Yes, at the set time	Automation, not AI
Decision-tree chatbot	No, scripted branches	Yes, within the script	Automation, not AI
Large language model chatbot	Partly, generates text from training patterns	Yes, replies without approval	Passes both, depending on the setup
Large language model writing assistant	Partly, generates text from training patterns	Usually no, you review and use the draft	Passes one clearly. A drafting aid, not an agent
Google Ads Smart Bidding	Yes, learns from conversion data	Yes, sets bids in each auction	Passes both

Google’s own description of Smart Bidding, from Google Ads Help, is “bidding strategies that use Google AI”, aimed at conversions or conversion value in every auction. It is known as auction-time bidding, and the strategies are Maximise conversions, Target CPA, Maximise conversion value and Target ROAS.

The table is the point of this page. In our view, much of what a small business already pays for sits in the top three rows. That is not an insult to those tools. A cheap autoresponder that works is good value. The only problem is paying AI prices for it.

Passing both questions raises the stakes as well as the value

A tool that learns and acts is more useful, because it can do work you could not do by hand, such as adjusting bids auction by auction. But the two characteristics the white paper picked out are the same two it flagged as sources of difficulty. Adaptivity can make logic hard to explain. Autonomy can make responsibility hard to assign.

For you, that has a practical meaning. When a rule-based tool does something wrong, you can open the rule and see why. When a tool that learns does something wrong, the reason may not be visible, and you are still the one answerable to the customer.

A scoring tool is a good example. A lead score that learns from your past enquiries and ranks the next ones passes both questions, and in our view it is also profiling under UK GDPR. We cover what that involves in our guide to lead scoring and UK GDPR profiling.

So passing both questions is not a reason to avoid a tool. It is a reason to keep someone watching it. Check what it did last week, set limits on what it may spend or send, and know how to switch the autonomous part off.

What the ASA says suppliers can and cannot call AI

The Advertising Standards Authority treats “AI” in an ad as a claim like any other. The advice below is as summarised by law firm Bird & Bird on 4 October 2024. According to that summary:

• Do not falsely claim a product uses AI.
• Do not exaggerate what an AI product can do.
• Do not claim an AI product does something better than a non-AI product unless you have good evidence.
• Using AI to develop a product is distinct from the finished product having AI in it.

The rules behind this are in the CAP Code. Bird & Bird refers to rule 3.1 on materially misleading claims, 3.7 on holding documentary evidence, 3.11 and 3.12 on exaggeration, 3.9 and 3.10 on making significant limitations clear, and 3.33 and 3.35 on comparisons.

In an ASA article, Regulating the illusion of intelligence in ads (27 August 2026), the ASA puts the principle simply: what the ad promises, the product should deliver. It also says advertisers should hold evidence for objective claims about what a product can do.

One limit is worth knowing. The CAP Code applies to ads, not to the products themselves. So the ASA looks at the claims a supplier makes in its advertising. It does not test whether the software works. The same rules apply to your own ads, so if you tell customers a service is AI-powered, you need to be able to back that up. Our guide to common AI marketing mistakes in the UK covers regulator fines for things like fake reviews. This section is general information, not legal advice.

Five things to ask a supplier before the demo ends

Put these to a salesperson, and write down the answers. A good supplier will have clear ones.

1. What does it learn from? This is adaptivity. Ask what data it was trained on, and whether it keeps learning as it runs.
2. Whose data is that? Ask whether it learns only from your account, from pooled data across customers, or from data you cannot see. The answer affects both results and privacy.
3. What does it do without me? This is autonomy. Ask which actions it takes on its own, which need your approval, and where the limits are.
4. What evidence backs the “better than” claim? If the pitch says it beats the manual or non-AI way, ask for the evidence. The ASA’s advice is that this claim needs good evidence behind it.
5. Can I switch the AI part off, and what changes if I do? If the tool still works as a plain rule-based one, you have a fallback. If everything stops, you are more dependent than the demo suggested.

If an answer is vague, that tells you something. A tool that cannot say what it learns from may not learn from anything.

A hypothetical: a two-van plumbing firm in Cardiff sorts its subscriptions

This is a hypothetical, not a client. A plumbing firm in Cardiff has two vans and a handful of monthly subscriptions. The owner decides to run the two questions on each before adding anything new.

The email autoresponder sends the same acknowledgement to every enquiry. It learns nothing, and it acts on a rule. Automation, not AI, and perfectly good for what it does.

The scheduling tool books jobs into slots according to availability the owner entered. It follows rules and learns nothing from data. Automation again.

The Google Ads campaign uses a Smart Bidding strategy. It learns from the conversions the firm has set up, and it sets bids in each auction without approval. It passes both. That means the owner should check which conversions are being counted, because that is what the tool learns from.

The website chat widget turns out to be a decision tree with scripted branches. It acts without the owner, but it learns nothing. Automation, so the owner stops treating it as an AI feature.

The writing assistant drafts replies and social posts from patterns in its training data. The owner reviews every draft before it goes out, so it acts only when pressed. A Welsh business may also want bilingual Welsh and English replies, and in our view any Welsh output from an AI writing tool should be checked by a Welsh speaker before it is sent.

The result is a list with one tool that passes both questions, one that passes one, and three that are plain automation. Nothing was cancelled on the strength of a label. The owner now knows where to keep an eye on things, and what to ask before the next renewal. We report no outcomes or savings, because this is an illustration.

Do you need AI marketing yet?

Not necessarily, and not all at once. Our view is that a small business gets more from running the test than from buying a tool.

1. List what you already pay for. Run the two questions on each tool, using the table above as a guide.
2. Stop paying AI prices for automation. If a tool fails both questions, it may still be worth keeping. Just do not accept that it is AI.
3. Start with one job. Pick one task that is repetitive and has data behind it, such as bidding or sorting enquiries, and try a tool that passes both questions on that job alone.
4. Keep someone responsible. The more a tool does without you, the more it matters that a named person checks it.

If you want a second opinion on your stack, our AI automation service can help you sort which tools earn their place. To see how your marketing is set up today, request a free audit.

Frequently asked questions

What is AI marketing, in plain terms?

AI marketing is software that does marketing work by learning from data and acting without step-by-step instructions. A tool that sets ad bids, drafts copy or ranks enquiries from patterns it has found counts. A tool that follows rules you typed in, or a schedule you set, is ordinary automation. Both can be useful, but only one deserves an AI price.

How does the UK government define artificial intelligence?

It does not give one legal definition. The 2023 white paper says there is no widely agreed definition and describes AI by two characteristics: adaptivity, meaning systems trained to infer patterns in data that humans often cannot easily discern, and autonomy, meaning some systems can make decisions without the express intent or ongoing control of a human.

Is an automatic email reply AI marketing?

Usually not. A rule-based autoresponder sends the same reply to every enquiry, so it learns nothing from data, and it acts only because you set the rule. That is automation, and it can be very useful. It becomes AI only if the system writes or chooses replies itself from patterns it has learned, which is a different product.

Is Google Ads Smart Bidding a form of AI?

Google describes Smart Bidding as bidding strategies that use Google AI, aimed at conversions or conversion value in every auction. In our view it passes both questions: it learns from conversion data and sets bids without you approving each auction. You still choose the goal and the conversion data it learns from.

Can a supplier call its product AI-powered if it is not?

Not if the claim is false. The ASA’s advice, as reported by Bird & Bird, is that advertisers should not falsely claim a product uses AI or exaggerate what it can do. They should also hold good evidence before saying an AI product beats a non-AI one. Using AI to develop a product is different from the finished product containing AI.

Who checks whether a marketing tool’s AI claims are true?

The ASA checks advertising claims under the CAP Code, so it looks at what a supplier says in its ads, not at the product itself. It says what the ad promises, the product should deliver, and that advertisers should hold evidence for objective claims about what a product can do. The same rules apply to your own ads.