Marketing AI Products Playbook

Sell an AI product on the job it does and the result it delivers, not on the model behind it. Prove it on the buyer’s own work, answer the risk questions before they are asked, and price it so that heavy use raises your revenue faster than your model bill.

Every software company now says it has AI, which makes “AI-powered” close to meaningless as a message and raises the bar for proof. Three facts about today’s buyers shape the steps below.

Use is common; value is not. In McKinsey’s 2026 survey of 1,719 people in 97 countries, nearly nine in ten said their organization uses AI regularly in at least one business function, but only 37% attributed any profit (EBIT) impact to it^[1]. Many of your buyers have tried AI and seen little come of it.

Trust is low. In the University of Melbourne and KPMG’s 2025 study of 48,340 people in 47 countries, 66% used AI regularly but only 46% were willing to trust AI systems^[2].

More people join the decision. In Forrester’s survey of nearly 18,000 business buyers, purchases that included generative AI features had buying groups twice the size of those without, 14 people against seven^[3]. Expect security, legal and data teams at the table.

My answer to all three is the same: be specific, prove value before you ask for commitment, and know what each sale costs you to serve.

In my view “AI” rarely belongs in the headline. Lead with who the product is for, the job it does and the result, and explain the AI where you explain how it works.

There is some evidence for this. In experiments with more than 1,000 US adults, mentioning “artificial intelligence” in a product description lowered emotional trust and purchase intention, and the effect was stronger for higher-risk products such as medical devices and financial services^[4]. That study is about consumers, and I have found no equivalent study of business buyers. But it fits the public mood: in Pew Research Center’s June 2026 survey of 3,488 US adults, 52% were more concerned than excited about the growing use of AI in daily life, up from 37% in 2021^[5].

Write the message in five parts, in this order: who it is for, the job, the result, how the AI does it, and the limit. The first three go in the headline and the last two in the copy beneath it. The limit is not a weakness: it answers the first question a careful buyer asks.

Do not compete on the model. Capable models are cheap and getting cheaper: Stanford’s AI Index found that the price of using a model that scores at GPT-3.5’s level fell from $20 to $0.07 per million tokens between November 2022 and October 2024^[6]. And the leaders are bunched together: as of March 2026, the top models from four companies were within 25 points of each other on the Arena leaderboard^[7]. Whatever model advantage you have today, a competitor can probably buy soon.

You also have two new competitors. One is the buyer’s own team: in McKinsey’s 2026 survey, 32% of respondents said their organization had decided against buying at least one software product or feature because it could be built in-house with agentic coding tools^[1]. The other is the free chatbot the buyer already uses. Do not argue that your model is smarter. Show the work a general tool leaves to the buyer, such as supplying context, moving the output into the workflow, checking it and controlling the data, and prove that your product does that part. A timed side-by-side on one real task, with the errors counted, does this better than a feature list, and conceding what the free tool does well makes the rest of your case believable.

Differentiate on what is hard to copy:

What Duolingo’s and Chegg’s filings show

Two education companies’ filings from 2023 show both sides of this shift (Figure 2). Duolingo sold AI as a paid upgrade and placed its advantage around the model, not in it^[8]. Chegg met AI first as a competitor: its annual report for 2023 named AI tools as a risk to its business, reported 7.7 million subscribers, down 6% from 8.1 million in 2022, and described its own AI rollout from September 2023^[9]. The filings say what each company said, not how it decided, and they do not show how much of Chegg’s decline came from AI.

The lesson I take for a startup: the free tool may be the alternative your buyer compares you with, and the durable part of your position is what you build around the model.

Buyers want proof that the product works on their work. Accuracy is the risk they worry about most: in McKinsey’s 2025 survey of 1,993 participants, inaccuracy was one of the two AI risks organizations most often said they were working to mitigate, 51% of respondents said their organization had seen at least one negative consequence from AI, and nearly a third reported consequences from inaccuracy^[10]. And buyers still want a person to check the answer: in Gartner’s survey of 645 B2B buyers, 69% preferred to validate AI-generated insights with sales reps^[11].

A proof kit has five parts, built in this order:

1. A test on the buyer’s own data, with the success measure agreed before it starts (Step 5).
2. An evaluation you can show. NIST’s AI Risk Management Framework says “Accuracy measurements should always be paired with clearly defined and realistic test sets – that are representative of conditions of expected use – and details about test methodology”^[12]. So report the task, the number and source of the test cases, who judged “correct”, the date and product version, and each type of error separately: what the AI missed and what it got wrong.
3. A limits page: where the product does worse, and how errors are caught.
4. Human review points: where a person checks the output, what they can override, and what happens when the AI is unsure.
5. Security and data-use answers (Step 4).

Before launch, sort every failure you find into two groups. A launch-blocking failure stops the launch whenever it occurs; one customer’s data shown to another always belongs here. A disclose-and-fix failure goes on the limits page and into your tracking. There is no universal accuracy bar for launch, so set yours by what a failure costs the buyer, and widen access in stages, from internal use to pilots, invited groups and then the public, with a written stop rule at each.

With security, legal and data teams in the buying group, most deals bring a security questionnaire. Answer it from one maintained set of documents: a data-use page, a trust page and a dated list of subprocessors. Disclose what a buyer can check, and keep private what a competitor could copy:

• Disclose: what data the product reads and produces; where data goes, whether anything trains on it, how long you keep it and who can see it; where a person reviews the output; your evaluation method and results; and your model provider, whenever customer data reaches it.
• Keep private: system prompts, fine-tuning data, the full evaluation set (publish the method and a few sample cases), rules for routing between models, provider price terms, and architecture detail beyond what a review needs. Offer what you will not publish, such as a penetration test report, under a confidentiality agreement rather than refusing.

These answers are promises you must keep. Staff in the FTC’s Office of Technology wrote that model-as-a-service companies “must also abide by their commitments to customers regardless of how or where the commitment was made”^[13]. If you process EU personal data for a customer, the GDPR says “The processor shall not engage another processor without prior specific or general written authorisation of the controller”, and under a general authorisation you must tell the customer before you add or replace one^[14]. Hence the dated list.

If your product serves children, check children’s privacy law too. Under the US COPPA Rule a child is anyone under 13, and personal information includes “A photograph, video, or audio file where such file contains a child’s image or voice”^[15]. Since April 22, 2026, the amended rule has required a written data retention policy, published in the privacy notice, that says why children’s data is kept and when it is deleted^[16]. This is orientation, not legal advice.

Trials are common, but they are not sales. In Forrester’s survey, more than 60% of business buyers bought some form of trial, but just over a third said they planned to convert to a fully paid version with the same provider^[3]. So treat a pilot as a test with an end date. Before it starts, agree one workflow, one success measure, the data the test will use and how it is protected, and a decision date with the options written down. At the end, walk the buyer through the results in person; that is where buyers’ wish to have a person check the AI (Step 3) pays off.

During the pilot, log what users do with every output: accepted unchanged, accepted after edits, regenerated or discarded. Acceptance is a useful early signal. In GitHub’s study that matched 2,047 Copilot users’ survey answers to their usage data, the rate at which shown suggestions were accepted predicted how productive users felt better than more detailed measures did^[17]. That links acceptance to how useful people find a product, not to whether they pay, so compare the pilots that convert with those that do not before you rely on it.

Once you run self-serve trials, you can make that comparison directly. Group trial accounts by the quality of output they saw in their first week, using the signals you already log, such as the share of outputs accepted unchanged or regenerated, and compare how many in each group go on to pay. A gap shows that quality and conversion move together, not that one causes the other: accounts that got worse output may also be smaller or less committed. To show cause, test it. Give half of new trials a quality improvement and keep the other half as they are, then compare paid conversion once both groups have had time to decide. Fix the sample size before you start, because a difference between small groups is often noise.

Claim only what you can prove when you say it. Regulators treat exaggerated AI claims as ordinary deception, and under the FTC’s policy on advertising substantiation, a claim that tests prove something needs “at least the advertised level of substantiation”^[18]. This is orientation, not legal advice.

• The FTC. Announcing a crackdown on deceptive AI claims in September 2024, the FTC’s chair said “there is no AI exemption from the laws on the books”^[19]. In 2025 it acted against Workado, which marketed an AI-content detector as 98% accurate; the FTC’s complaint said its accuracy on general-purpose content was 53%^[20]. The final order bars such claims unless the company has “competent and reliable evidence” for them when it makes them^[21].
• The SEC. In 2024 it charged two investment advisers over “AI washing”: claiming AI capabilities they did not have^[22]. In 2025 it charged a startup founder who raised over $42 million while claiming his shopping app completed purchases with AI; the SEC alleged that the app “relied in large part on contract employees to manually input orders”^[23]. Investor materials count as marketing too.
• “Agent washing.” Gartner estimates that only about 130 of the thousands of vendors claiming agentic AI are real, and predicts that over 40% of agentic AI projects will be canceled by the end of 2027^[24]. Call something an agent only if it acts on its own.
• The EU AI Act. Since August 2, 2026, providers must tell people they are interacting with an AI system unless it is obvious, and providers of generative AI must mark its output in a machine-readable way; systems already on the market before then have until December 2, 2026 for the marking^[25]. If you sell into the EU, check these duties.

A simple rule for every claim: a number needs a dated, repeatable test behind it; “autonomous” or “agent” needs no person doing the work; “secure” needs a named standard or audit; and “more accurate than ChatGPT” needs a side-by-side you can show.

Pick a price metric that grows with the value the buyer gets and with your own running costs. Practice is still settling. In Bain’s analysis of the public pricing of about 200 B2B software companies in June 2026, about four in five vendors introducing AI pricing chose capacity models, such as prepaid credits or entitlements; among those adding a hybrid AI meter, about 55% metered output, about 35% effort and only about 10% outcomes; and about one in five AI-native companies still relied mostly on per-seat pricing^[26]. Buyers also want room to move: in a McKinsey survey of 150 enterprise software buyers in October 2024, 65% said being able to move usage or spending commitments from one product to another was very or extremely important^[27].

No study yet shows which model performs best, so choose by fit:

A hybrid, a base fee with included usage and a price for more, is the common compromise. Never sell unlimited use at a fixed price: every task costs you model and hosting fees, so check the price against your cost per completed task.

Work that cost out from your logs, not from estimates: input and output tokens, retries, tool calls and hosting. Then report gross margin after inference, revenue minus the costs that grow with use (model and API fees, hosting, any human review) divided by revenue, rather than burying model fees in R&D. AI margins are lower than classic software’s: in ICONIQ’s survey of about 300 executives at software companies building AI products, average gross margin on AI products was 45% in 2025, with 53% projected for 2026^[28].

To see how much the model tier matters, take one illustrative query of 3,000 tokens in and 500 out. On the standard prices each provider listed on October 22, 2026, a small model ($0.10 per million input tokens and $0.50 output) costs about $0.00055 and a mid-tier model ($2 and $10) about $0.011 at all three of Anthropic, OpenAI and Google; the top tier runs from $0.022 to $0.055^{[29],[30],[31]}. The query size is an assumption, so use your own logs. The number to track is cost per completed task, not cost per query: if users retry 4 in 10 answers once, that is 1.4 queries for every answer kept, and a $0.011 query becomes $0.0154. Caching the 2,000 tokens of instructions that repeat cuts the query to about $0.0072 on Claude Sonnet 5.5, where a cache read costs 5% of the input price, and running work that can wait as a batch job halves it again to about $0.0055 (my arithmetic; the cache-writing cost is left out)^[29].

Price lists also move. Google’s page shows gemini-4-argon at $2 input and $10 output per million tokens through November 30, 2026 and $4 and $20 from December 1, 2026, and gemini-3.8-flash doubling on January 1, 2027^[31]. Falling average prices are not a promise about your cost per customer, so re-check them every quarter.

Before a launch that could spike usage, set a spending cap with your provider and daily limits for each user in the product, and keep free and trial traffic apart so a spike there cannot stop paying customers. Anthropic’s documentation, for example, says “Spend limits set a maximum monthly cost an organization can incur for API usage”, and limits can be set for each workspace^[32]. Work that can wait is cheaper: its Batch API gives “a 50% discount on both input and output tokens”^[29]. Check your own provider’s current terms. Caps can lag: OpenAI’s spend-limits guide says enforcement is not instantaneous, and Google says billing data can trail by about 10 minutes^[33],[34]. So alert well below the cap, and give free and trial traffic its own workspace and cap.

Choose channels where buyers can see the product work. Three matter most:

• Hands-on trials and product-led growth. Let people try it on a real task quickly, with usage caps so free use cannot outrun your budget. The Product-Led Growth Playbook covers this.
• AI answers. In Forrester’s survey, 94% of business buyers said they used AI during their buying process^[3]. If AI assistants do not describe your product correctly, you are missing a research channel. The GEO and AEO Playbook covers being cited.
• Expert sales. With bigger buying groups and buyers who want a person to check the AI, you need people who can run a proof and answer security questions.

Content that shows your evaluation method, your limits and real before-and-after results builds more trust than feature lists. Use fictional examples, or real ones only with the customer’s written permission.

The real test starts after the contract. In McKinsey’s 2026 survey, 44% of respondents said AI was scaling across their enterprise, up from 38% a year earlier^[1], so most buyers have not yet made AI part of how they work. If the customer’s project stalls, your renewal goes with it.

Plan onboarding that reaches the first measurable result fast, train the users and not just the admins, and put ready-to-run templates for their everyday jobs inside the product (see Common questions). Then report value back every quarter, and choose the metric you report with care. Rank them: the outcome the buyer pays for first; tasks completed second, meaning outputs the user accepted and used; weekly active users third, counting only people who completed a task; and prompts last. More prompts can mean a user is struggling, and they raise your costs, so treat prompts per completed task as a measure of friction, not of engagement (Figure 9). Cost per query, a popular AI KPI, misleads in the same way: internally, a question asked three times costs three queries for one answer, so track cost per completed task (Step 7) and read cost per query beside it as friction.

What changes by stage?

What mistakes should I avoid?

• Leading with the model. Lead with the job and the result; the model is replaceable (Steps 1 and 2).
• Publishing one accuracy number with no method. Show the test, its date and each type of error (Step 3).
• Hiding your model provider until a questionnaire forces it out. Name it wherever customer data reaches it (Step 4).
• Open-ended free pilots. Agree one measure and a decision date up front (Step 5).
• Calling a workflow an agent. Use the word only for software that acts on its own (Step 6).
• Unlimited use at a fixed price. The heaviest users then cost more than they pay (Step 7).
• Reporting prompt volume as traction. Report the outcome the buyer pays for (Step 9).