Short answer: “AI-powered” no longer separates you from anyone. In September 2026, 49.0% of B2B software homepages mentioned AI, and companies that made AI their main claim grew no faster than similar companies that didn’t. Position an AI product the way buyers evaluate it: the job and the outcome, the alternative you replace (manual work, an incumbent or an in-house build on a model API), accuracy on data like the buyer’s, what you own besides the model and where people review the output. Keep every claim provable, because the SEC and the FTC were still bringing AI-washing cases in 2026.
This guide covers what changes when the product is AI. For the general process, start with our guide on how to position a B2B startup. For the conversations that feed both, see customer interviews for positioning.
Why doesn’t “AI-powered” set an AI startup apart anymore?
Half the market says it. Kenso, a sales data company, analyzed the homepages of a random sample of 10,000 B2B software companies in six English-speaking countries. In September 2026, 49.0% mentioned AI, up from 11.7% in 2019, and for 25.8% AI was the main claim in the title, description or main headline. Those companies grew headcount no faster over the next two years than matched peers, and 9.3% of them raised money within 24 months against 8.4% of their matches, a gap Kenso describes as within the noise. It is a link rather than proof of cause, but the label didn’t come with faster growth or easier fundraising.
Gartner estimates that only about 130 of the thousands of agentic AI vendors are real. It describes the practice as “agent washing”, the rebranding of assistants, robotic process automation and chatbots that lack substantial agentic capabilities.
Buyers have adjusted. In 6sense’s 2026 buyer research, eight in ten B2B buyers evaluated solutions that contained AI, and for 58% it was strongly encouraged or mandated. When everything on the shortlist has AI, buyers ask what it does for them, how well and at what risk.
In our own review of 30 funded B2B startup websites, the highest-scoring site, Zynt, opens with “The daily list of companies that are starting to buy.” Its first screen doesn’t mention AI. Kuro, a German company building AI for general contractors, names its customer Köster and the result, 6 hours saved per week per employee.
What happens when AI claims go further than the product?
Regulators call it AI washing. In the cases so far, the company made a specific claim it couldn’t back up.
| Date | Regulator and company | The claim | What the regulator found or alleged |
|---|---|---|---|
| March 2024 | SEC: Delphia and Global Predictions | Delphia said it used AI and client data to predict winning companies. Global Predictions called itself the “first regulated AI financial advisor” | The SEC found the AI claims false and misleading. $400,000 in combined penalties |
| September 2024 | FTC: DoNotPay, part of Operation AI Comply | ”The world’s first robot lawyer” | It never tested whether its output matched a human lawyer’s work. $193,000 settlement |
| January 2025 | SEC: Presto Automation | Its own AI took drive-thru orders without human order-taking | For a period the speech recognition belonged to a third party, and the vast majority of orders needed human intervention |
| April 2025 | SEC: the founder of Nate | AI completed purchases in its app without human involvement, as told to seed and Series A investors | The app relied in large part on contract workers entering orders by hand. Over $42 million raised |
| April 2025 | FTC: Workado | An AI content detector that was “98 percent” accurate | The model was trained on academic text, and independent testing on general content found 53% accuracy |
| May 2026 | FTC: Cox Media Group and two marketing firms | An AI service, sold to small businesses, that targeted ads using conversations picked up by smart devices | The service didn’t use voice data. $930,000 in total |
| August 2026 | SEC: GenesisAI | Revenue projections, 25 “partnerships” and a customer waitlist | The partnerships weren’t enforceable and no waitlist existed |
Most of these ended in settlements in which the company neither admitted nor denied the findings. The Nate row describes the allegations in the SEC’s April 2025 complaint.
US policy on AI has loosened, but regulators still pursue deceptive claims. In December 2025 the FTC set aside its 2024 order against Rytr, an AI writing tool, as an undue burden on AI innovation. Yet by May 2026 the law firm DLA Piper counted 13 FTC AI-washing cases since 2024, seven of the latest eight over claims made to businesses, and the SEC brought three AI-claim cases in August and September 2026.
What do buyers check before they buy an AI product?
They check five things, and different people in the buying group own each one.
- Accuracy on their data. A benchmark on someone else’s data tells them little. Workado’s detector was trained on academic writing and failed on general content.
- Integration. In 6sense’s research, 25.9% of buyers whose evaluations stalled cited implementation concerns. In G2’s interviews, leaders who described what slowed adoption named compliance requirements, legacy-system integration and single sign-on and permissions.
- Security and data handling. In the 6sense study, 52.8% of buyers had AI-specific security concerns, rising to 64.0% among finance buyers. In G2’s interviews with 102 US enterprise leaders about language model purchases, 59% of those who named a dealbreaker or near-dealbreaker pointed to security, privacy or data exposure, ahead of task quality (28%) and price (13%). Of those who named someone with veto power, 87% pointed to a risk or governance team.
- Total cost. Nearly half of buyers (47.0%) wanted pricing clarified, the same share as a year earlier. They also disagreed on what they were paying for: 67.2% expected AI to be included in the standard offering and 26.4% expected an extra charge.
- Who is accountable for errors. In 2024, after an Air Canada chatbot gave a customer wrong information about bereavement fares, the airline argued that the chatbot was “a separate legal entity that is responsible for its own actions.” A British Columbia tribunal disagreed: “It should be obvious to Air Canada that it is responsible for all the information on its website.” Your buyer carries the risk of your output, and in G2’s research 45% of 58 leaders said customer-facing and sensitive workflows need a higher accuracy and safety bar.
Buyers who expected significant AI were more than twice as likely as other buyers to ask about AI capabilities (57.1% against 21.6%), and more likely to ask about security, data processing, ROI and whether the technology was proprietary or third-party. 6sense advises vendors to make those answers easy to find before the first sales conversation. Our guide to what enterprise procurement checks on your website lists the pages reviewers open.
Six positioning moves for AI products
1. Lead with the job and the outcome
Buyers pay for a finished job. Name it in their words, put a number on the outcome and say who it is for. Model names, “agents” and “LLM” belong further down the page, where a technical evaluator will look for them.
| Generic AI claim | Job and outcome (illustrative) |
|---|---|
| “AI-powered contract intelligence platform" | "Flag the risky clauses in a supplier contract before it reaches legal" |
| "Autonomous AI agents for finance" | "Match every supplier invoice to its purchase order, so your team only reviews the exceptions" |
| "Next-gen AI for customer support" | "Answer order-status tickets from your own systems, with a person approving every refund” |
In 6sense’s data, buyers whose evaluations stalled were twice as likely as buyers who purchased to say their AI expectations depended on the vendor (13.1% against 6.5%). A specific outcome gives the buying group something to compare vendors against.
2. Name the alternative you replace
Most AI products compete with one of three things, and a homepage can argue against only one at a time.
- Manual work or a service provider. Compare on cost per unit of work, turnaround and error rate, and show who still checks the result.
- An incumbent adding AI to software the buyer already owns. Under economic pressure, 69.1% of buyers expecting significant AI favored established or familiar providers, against 49.6% of other buyers. Answer with depth on one job and proof from customers like them.
- An in-house build on a model API, or a general assistant. In Menlo Ventures’ survey of 495 US enterprise AI decision-makers, 76% of AI use cases were purchased rather than built in 2025, up from 53% in 2024, but the internal team is still in the room. Compare on time to reach your accuracy and the cost of maintaining evaluations, integrations and security reviews.
3. Prove it with evaluations and pilot metrics
An accuracy number means little unless the buyer can see how you got it. Publish the test set size and source, the date, the model version, the baseline (the current process or a person doing the task) and the kinds of errors you found, and re-run it when you change models.
Avoid absolute words. Some legal research providers promoted their methods as “eliminating” hallucinations or guaranteed “hallucination-free” citations. A preregistered study from Stanford’s RegLab found that the tools from LexisNexis and Thomson Reuters hallucinated between 17% and 33% of the time. They did hallucinate less than a general-purpose chatbot, which was the claim the evidence supported.
For pilots, agree on one metric with the buyer before you start, measure the baseline first, count review time and errors alongside speed, and set a decision date. Keep the first step small. Zeit AI, a German data engineering company for finance teams, offers a free proof of concept in 24 hours.
4. Explain the moat honestly
Investors and technical buyers will ask why a model provider won’t ship your product as a feature. Three answers hold up when they are true.
- Data. Data you have the rights to use, that the model hasn’t seen and that improves the product with each customer.
- Workflow. Integrations, approvals, audit trails and domain rules, which take time to build and are costly to rip out. Bessemer Venture Partners writes that “for AI application founders, context and memory may be the new moats.”
- Distribution. Existing customer relationships, a channel, a community or a regulated status the model providers lack.
Eight Capital, an early-stage fund, says the portfolio companies that reached priced Series A rounds each owned “something a better model makes more valuable, not less.” Use that as your test. If a better model would make your product unnecessary, you have a head start rather than a moat. Say so, and show what you will build with the time.
5. Answer the wrapper question before it is asked
Darren Mowry, who leads Google’s global startup organization, said in February 2026: “If you’re really just counting on the back-end model to do all the work and you’re almost white-labeling that model, the industry doesn’t have a lot of patience for that anymore.”
Answer the question in three parts. Say which models you use, since Presto shows the risk of hiding a third party. Show what you built around the model. Then run the buyer’s task through a general assistant and through your product side by side, including the time a person needs to check each result. An illustrative answer: “We use [provider]‘s models to read documents and can switch providers. What we built is the matching logic for [industry] invoices, connectors to [ERP] and a test set of [number] real invoices that we re-run after every model update.” Put it in your FAQ, deck appendix and sales notes.
6. Say exactly where humans review the output
Presto and Nate show the risk of claiming “no human involvement” when people step in. Describe the review precisely: who reviews (their team or yours), what they review (everything, exceptions or a sample), the share of outputs that needs review today and how that is expected to change. “Your AP lead approves every exception above $5,000” tells a buyer far more than “human oversight”.
How should the investor story differ from the buyer story?
The facts stay the same and the questions change. Buyers ask whether it works for them. Investors ask whether it becomes a large business that gets harder to copy.
| Buyer version | Investor version | |
|---|---|---|
| Opening | The job and the outcome, in the buyer’s words | The same job, plus market size and why now |
| Proof | Evaluations on similar data, pilot results, references | The same evidence, plus retention, expansion and pipeline |
| The model | Which models you use, how data and errors are handled | Dependence on model providers, inference cost per customer, gross margin |
| Moat | What they would lose by switching | Why each new customer makes you harder to copy |
| Human review | Who checks what | How review time and cost change as volume grows |
At seed, the story rests on a sharp wedge (one job, one buyer), a reason why now (what changed in model capability for this task), evidence of pull such as design partners or paid pilots, a credible moat hypothesis and the team’s access to the domain.
At Series A, the moat has to show up in numbers: retention, expansion, gross margin after inference costs and a sales motion that repeats. Bessemer’s 2025 benchmarks describe “Shooting Stars” reaching about $3M ARR in their first year of revenue with about 60% gross margins, and “Supernovas” growing much faster on about 25% margins with revenue that “may appear vulnerable.” Eight Capital puts the 2026 Series A conversation closer to $3-5M ARR with net revenue retention above 120%, though that is one fund’s bar.
Emphasis can differ between the decks, but the facts can’t. The SEC’s complaint against Nate’s founder concerns claims made to seed and Series A investors. Our guide on one story across the pitch deck, website and sales deck shows how to keep them aligned.
What should you ask customers about an AI product?
Interview recent buyers, lost deals and the people who check the output, not only the champion, and ask about what already happened. Add these questions to your usual guide.
- “Walk me through the last time you did this task before using us. Who did it, and how long did it take?”
- “Did anyone try doing it with a general assistant or an internal tool? What happened?”
- “When was the output last wrong? How did you notice, and what did fixing it cost?”
- “Which mistakes would be unacceptable, even once?”
- “Who checks the output before it goes out, and how long does the check take?”
- “What did you need to see before you trusted it with real work?”
- “Who had to approve access to your data, and what did they ask?”
- “If the AI part stopped working tomorrow, what would you still use?”
Listen for the words people use about trust, such as “I can see where the answer came from” or “it tells me when it isn’t sure,” and use them in your copy. Answers to question 8 show whether your value sits in the model or in the workflow around it.
What does a message hierarchy for an AI product look like?
This example is for a fictional product that matches supplier invoices for manufacturing finance teams. Brackets mark facts you would fill in from your own interviews and pilots.
| Level | Example |
|---|---|
| One-liner | ”[Product] checks every supplier invoice against the purchase order and the delivery note, so your AP team only reviews the exceptions.” |
| Lead outcome | Close the month without an invoice backlog. Proof: “[x]% of invoices matched without manual review in a [6]-week pilot at [customer].” |
| Pillar 1: accurate on your documents | We test on [number] of your own invoices before you sign, and share the error report |
| Pillar 2: works inside your ERP | Connectors for [ERP list]. No new inbox |
| Pillar 3: your team stays in control | Every exception goes to a named reviewer, with a full audit trail |
| Proof points | Pilot metric, a named customer quote, a security page with data retention and model provider details |
| Investor line | Every customer adds supplier-specific matching rules, which makes onboarding the next manufacturer faster |
Claims checklist before anything goes live
Run every page, deck and ad through this list. It doesn’t replace legal advice, so have counsel review investor materials.
| Claim type | Risky version | Safer version | Case that shows the risk |
|---|---|---|---|
| Use of AI | ”AI-driven” on a feature built on rules | Say AI only where a model does the work | Delphia (SEC, 2024) |
| Firsts and superlatives | ”The first AI [profession]“ | A specific, checkable difference | Global Predictions (SEC), DoNotPay (FTC) |
| Accuracy | ”98% accurate" | "[x]% on [test set], [date], [conditions]“ | Workado (FTC, 2025) |
| Replacing people | ”Replaces your lawyers” or “replaces your reps” | The task it handles and who still reviews it | DoNotPay (FTC), Air AI (FTC complaint, 2025) |
| Human involvement | ”Fully automated” | The share of cases a person handles | Presto (SEC, 2025), Nate (SEC complaint, 2025) |
| Own technology | Implying the model is yours | Name the provider when buyers ask | Presto (SEC, 2025) |
| Reliability | ”Hallucination-free” | A measured error rate and how errors are caught | Stanford RegLab study |
| Traction | Waitlists and partnerships that don’t exist | Signed customers and dated pilots | GenesisAI (SEC, 2026) |
Keep an evidence file with the source, date and owner of every number you publish, and have marketing, sales and fundraising work from that one file, so the website and the deck show the same numbers.
How dp.vision helps AI companies with positioning
dp.vision is a brand, web and product studio for funded B2B tech companies. We position products from interviews with our clients’ real customers, including lost deals and the people who check the output. AI helps us with the speed and cost of production, such as visual exploration, prototypes and assets, while the strategy comes from the interviews, which people run.
Two AI products we have worked on (results as reported by our clients):
- Edutailor, an AI and XR training platform. Brand identity and a landing page in 5 days.
- Mega Panda, an AI legal tech platform. Research with 12 attorneys, and task flows cut from 8 steps to 3. The product design includes dashboards where lawyers oversee the AI’s work.
We charge $85-100 an hour, with a fixed price agreed before work starts (how much dp.vision costs). If you are comparing studios, see our ranking of AI-native branding agencies and our list of brand strategy agencies for AI and B2B tech. dp.vision is our studio and appears on both.
To test your AI positioning with your own customers, get in touch.
Frequently asked questions
How do you position an AI startup when every competitor says AI?
Position it the way buyers evaluate it: the job and the outcome in the buyer’s words, the alternative you replace, accuracy on data like the buyer’s, what you own besides the model and where people review the output. In September 2026, 49.0% of B2B software homepages mentioned AI, so the word sets no one apart.
Should an AI startup put “AI” in its homepage headline?
Only if the buyer would miss it. In our review of 30 funded B2B startup websites, the top-scoring site led with the job it does for sales teams and didn’t mention AI on the first screen. Put the model, data handling and evaluation results lower on the page and on a security or FAQ page.
How do you answer “aren’t you just a ChatGPT wrapper?”
Say which models you use, show what you built around them (data connectors, domain rules, evaluations, review flows, integrations) and run the buyer’s own task through a general assistant and through your product side by side, including the time a person needs to check each result. Publish that answer in your FAQ and deck appendix.
What do enterprise buyers check before buying an AI product?
Accuracy on their own data, integration, security and data handling, total cost and who is accountable when the output is wrong. In G2’s interviews with 102 US enterprise leaders, 59% of those who named a dealbreaker or near-dealbreaker in their latest model purchase pointed to security, privacy or data exposure.
What is AI washing, and can it get a startup in trouble?
AI washing means claiming AI capabilities a product doesn’t have or overstating what it does. The SEC and the FTC have brought cases over it since 2024, including against startups and companies selling to businesses. Accuracy percentages, “first” claims, “no human involvement” and claims that a product replaces a professional carry the most risk.
How should an AI startup’s investor pitch differ from its sales pitch?
The facts must match. Buyers ask whether it works on their data and what could go wrong. Investors ask why the business gets harder to copy, what inference does to gross margin and whether customers stay. At seed, show a sharp wedge and evidence of pull. At Series A, show retention, expansion and margins after model costs.
What should you ask customers about an AI product?
Ask about past behavior around trust and errors: when the output was last wrong and what fixing it cost, who checks it and how long that takes, and which mistakes would be unacceptable even once. The answers tell you which claims to make and which proof to show.
What is a defensible moat for an AI application startup?
Something a better model makes more valuable: data you have the rights to use that improves the product with each customer, a workflow built into the buyer’s systems, or distribution the model providers lack. If a better model would make your product unnecessary, you have a head start, so explain how you will use it.
Sources
- Kenso Research, Half of B2B software homepages now say AI. Did it pay off? and methodology (PDF)
- Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027
- 6sense, AI Inside the Solutions B2B Buyers Evaluate and Economic Uncertainty
- G2, Who Can Veto an Enterprise LLM? For 87% of Leaders, It Isn’t the Budget Holder
- Menlo Ventures, 2025: The State of Generative AI in the Enterprise
- SEC, SEC Charges Two Investment Advisers with Making False and Misleading Statements About Their Use of Artificial Intelligence (March 18, 2024)
- FTC, FTC Announces Crackdown on Deceptive AI Claims and Schemes (September 25, 2024)
- SEC, SEC Charges Restaurant-Technology Company Presto Automation for Misleading Statements About AI Product (January 14, 2025)
- SEC, Litigation release: Albert Saniger (April 9, 2025)
- FTC, FTC Order Requires Workado to Back Up Artificial Intelligence Detection Claims and final order
- FTC, FTC Sues to Stop Air AI from Using Deceptive Claims (August 25, 2025)
- FTC, FTC Reopens and Sets Aside Rytr Final Order (December 22, 2025)
- FTC, FTC to Require Cox Media Group, Two Other Firms to Pay Nearly $1 Million (May 21, 2026)
- SEC, Litigation release: GenesisAI Corp. (August 26, 2026)
- DLA Piper, FTC AI-washing action underscores enforcement in business-to-business context and Three new SEC cases show continued scrutiny of AI-washing claims
- Civil Resolution Tribunal of British Columbia, Moffatt v. Air Canada, 2024 BCCRT 149
- Magesh et al., Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, Journal of Empirical Legal Studies (2025)
- Bessemer Venture Partners, The State of AI 2025
- Eight Capital, What Reaches Series A When AI Is Eating Software
- TechCrunch, Google VP warns that two types of AI startups may not survive
- dp.vision, We reviewed 30 funded B2B startup websites
