How to sell AI services when you have no case studies
Nobody has bought from you yet and the prospect just asked who else you have done this for. Here is what to say, what counts as proof before you have clients, and the proof you are legally not allowed to invent.
Lumina13 min read
You replace the case study with something the prospect can test while you sit there. A working version of the system, running under their business name, answering the questions they were about to ask you. For a first sale that beats a reference from a company they have never heard of, because they are not taking your word for anything. They are using the thing.
Every agency owner hits this wall in week two. You can describe an AI receptionist well enough to get a meeting. Then someone asks who else you have done this for, and the honest answer is nobody yet. Most people handle that moment badly, either by inflating a friend's side project into a flagship client or by apologizing for twenty seconds and losing the room. There is a third option.
Why your prospect asks about case studies
They are not asking for a PDF. They are asking a risk question in the only language they have: has this worked before, for someone like me, and will I look stupid for buying it.
That question has almost nothing to do with your history and everything to do with how much they have to trust you. Which is why Gartner's buyer research is worth reading even though it covers much larger deals than yours. In a survey of 632 B2B buyers fielded in August and September 2024, 61% said they prefer an overall rep-free buying experience, and 69% reported inconsistencies between what a supplier's website said and what its sellers told them. People do not want to be sold to by someone whose claims they cannot check. A case study is a claim they cannot check. A demo they can poke at is one they can.
Something else is going on with local service businesses specifically. Your buyer probably has very little AI in their own operation to compare against. The Census Bureau's Business Trends and Outlook Survey put the share of US businesses using AI in any business function at 19.8% as of May 3, 2026, holding between 17% and 20% across that whole survey period. Among firms with four or fewer employees, fewer than 20% reported using it, and unlike larger firms their rate did not move between December 2025 and May 2026. So a four-truck plumbing company asking for case studies is often asking a more basic question underneath. Does this actually work at all, on a phone, in my trade, with my prices.
Answer that one and the case study question tends to dissolve.
What actually counts as proof before you have clients
Proof is anything the prospect can verify without trusting you. Rank your options by how much trust each one requires, and the order is pretty clear.
A live system they operate themselves. Strongest by a wide margin. They pick up their phone, call a number, ask about pricing, and get booked. Nothing in that sequence requires belief.
Their own data inside it. A website audit run against their real domain, in the browser, while they watch. An ROI figure seeded with their average ticket and their close rate rather than a made-up industry average. Specifics they already know are specifics they can check. If the audit finds six issues on their site and they recognize four of them, you are done arguing about credibility.
A process they can inspect. Not a promise about outcomes, a description of mechanics. What you collect from them in week one. What goes live in week two. Who answers when the agent mishandles a call. An owner who has been burned by a contractor can read a schedule and tell whether it was written by someone who has done the work.
Terms that put the risk on you. Month to month with a short notice period. A first-month price that reflects you being new. A written exit where they keep their phone number and their data. This is proof in the sense that matters, which is that you are willing to be held to it.
Published third-party material. Vendor documentation, model pricing pages, named research. You did not produce it, so it carries weight you have not earned. Use it for the general claim and your demo for the specific one.
Named references. The thing you do not have. Worth noticing that it sits at the bottom of this list early on and climbs to the top by client five, which is a good reason to stop worrying about it and start collecting it.
Build the thing they can test
An interactive demo used to be a real engineering project. That is the part that changed, and it is why selling without case studies is more viable in 2026 than it was in 2023.
What a prospect should be able to do on the page you send them, before any call:
They should be able to call a receptionist for a business in their trade and ask a real question. Not a recording, not a Loom of someone else's system. A live conversation where they can interrupt, go off script, and try to break it. Half of them will try to break it. Let them.
They should be able to chat with the same agent on the page, because some owners will not call from their truck but will type at 11pm. They should be able to run an audit against their own website and read the findings against a site they know intimately.
And they should be able to watch a missed call turn into a text and then into a booking, played out rather than described.
Inside Lumina Sales that is what the demo builder assembles. You pick the trade, drop in the prospect's business name and website, arrange the nodes to match the order you actually pitch in, and publish to your own subdomain under your logo and colors. The live demo on our homepage is the same machinery with the trade switcher left visible, so you can hear what a regenerative medicine clinic agent sounds like against an HVAC one. Our platform plan is $49 a month and includes 100 live voice minutes, 500 chat messages, and 50 website audits, which works out to roughly forty demo calls a month. First demo live usually takes about ten minutes, most of it spent picking a subdomain and uploading a logo.
The trade-specific part matters more than people expect. A generic agent that says "thank you for calling, how may I help you" proves you can wire up an API. An agent that triages an emergency call differently from a maintenance call, and quotes from the price bands that trade actually uses, proves you understand their business. We write the packs for over 100 trades for exactly this reason, and we have written before about why generic prompts lose demos.
What to say when they ask you directly
They will ask. Have an answer that takes eight seconds and does not flinch.
What I say, more or less: "You're my fourth client in this trade's category, not my fortieth. That's why the demo is live instead of a slide, and why the first month is priced the way it is. Call it right now and try to trip it up."
Three things are happening in that answer. You admit the gap before they have to drag it out of you, which stops them from wondering what else you are hiding. You immediately hand them something better than the thing they asked for. And you invite the adversarial test, which is the single most persuasive move available to someone with no track record, because nobody who is bluffing says "try to break it."
What not to do, ranked by how badly it goes:
Do not say "we have worked with many businesses in your industry" when you have not. It is a lie, it is checkable, and it poisons everything else you said.
Do not show a competitor's case study as a proxy. Owners talk to each other.
Do not pad with your day job. "I ran operations at a 200-person company" is fine context, but pretending it is a case study reads as insecurity.
Do not discount to compensate. If you feel the need to drop price because you lack proof, your demo is not doing its job. Fix the demo. Our demo call playbook covers the running order that keeps the conversation on capability rather than price.
The proof you are not allowed to invent
This section exists because the AI services space has a manufacturing-proof problem and some of the people teaching it will tell you to do things that are illegal.
The FTC's rule on consumer reviews and testimonials, finalized in August 2024, bans fake or false reviews and testimonials, including ones generated by AI and ones attributed to people who do not exist. It also covers testimonials from anyone who had no real experience with the business, reviews from company insiders that do not disclose the connection, and claims that a review site you control is independent. The rule lets the FTC seek civil penalties against knowing violators. So the AI-written testimonial from "Mike R., HVAC owner" on your landing page is not a gray area and not a growth hack.
Earnings claims are the other trap, and the one that catches agency owners rather than their clients. In September 2024 the FTC ran a sweep it called Operation AI Comply, five cases against companies making deceptive AI claims. Two of them, as summarized by Orrick, were e-commerce business-opportunity schemes built on unsubstantiated claims about income from AI-powered stores, where most clients earned little and many lost their investment. Another, against DoNotPay, concerned an "AI lawyer" marketed beyond what it could do, settled with a $193,000 fine and a requirement to tell customers about the service's limits. Then-chair Lina Khan's line was that there is "no AI exemption from the laws on the books."
The practical version for a one-person agency: never state a result you did not measure. "This recovers missed calls" is a mechanism and you can show it running. "This will add $8,000 a month to your revenue" is a result, and if you have not measured it on a real client, you are making it up. You can model it with their numbers in front of them and call it a model. That is different, and owners can tell the difference.
Running your demos in a sandbox belongs in the same honesty bucket. Every published Lumina demo carries a visible notice that the agent is a demo and bookings in it are simulated. It is the one piece of non-agency text on the page and it is deliberate. We looked at whether that notice costs you closes and the short version is that it does not, because the alternative is a prospect discovering on their own that the booking never happened.
Where this approach is weak
A live demo is not a universal substitute for a track record, and pretending otherwise would be the same mistake in a different coat.
It proves capability, not outcomes. Your demo shows that an agent can answer a pricing question in their trade's language. It says nothing about whether their particular front desk will cooperate, whether their calendar will sync cleanly, or whether their call volume justifies the monthly. A case study speaks to outcomes. A demo cannot, and some buyers correctly care more about outcomes.
Integration risk stays unproven. The sandbox is deliberately not connected to a live calendar or CRM, which is what makes it safe to demo. It also means the messiest part of delivery, hooking into whatever ten-year-old scheduling software they run, is the part the prospect has not seen. Be upfront that this is the step where timelines slip.
Some buyers will not move without references and you should stop pushing. Multi-location operators, anyone with a procurement process, franchise groups, practices with a compliance officer. They are not being difficult. Their internal process requires a reference and you do not have one. Go find smaller buyers for ninety days and come back with three.
The product itself is not universally welcome. Metrigy surveyed 503 consumers for its customer experience study and found 84.7% said they would prefer interacting with a human over an AI agent, with 80.1% still preferring a human even when assured their issue would be resolved. That is a general finding about AI agents, not about after-hours phone coverage, and the comparison that matters at 9pm is an AI versus voicemail. But a thoughtful owner may raise it, and you need a real answer rather than a reflex. We wrote up what callers actually notice if you want the longer version.
A demo also does not fix a weak offer. If what you sell is not worth the money, letting someone try it gets you to no faster. Useful, but not a win.
Turning your first three clients into real case studies
The whole point of the demo-as-proof approach is that it is temporary. Treat your first clients as instrument panels.
Capture the baseline before you launch anything. Missed calls per week, average answer time, how many after-hours enquiries go to voicemail now, their average ticket, their close rate on inbound. Ten minutes of questions on the kickoff call. Without it you have no before, and a case study without a before is an anecdote.
Instrument the obvious counters from day one. Calls answered outside business hours. Bookings created by the agent. Missed calls recovered by text. Those three are enough.
Ask for permission to publish at signature, not at month three. People say yes to "if this works, can I write it up" far more readily than to a request that arrives after they have grown attached to their competitive advantage. Get it in the agreement.
Report at thirty days whether the numbers are good or not. If they are bad, you have found out early and you have a client who trusts you. If they are good, you have a case study with a before, an after, and a name on it.
By client four the question that opened this post stops coming up, and the demo goes back to being what it is best at, which is closing the deal rather than defending your résumé.
Frequently asked questions
Can I use the demo for a trade I have never sold into?
Yes, and this is the main argument for a vertical pack library over a single generic agent. You pick the trade, the services, price bands, objections, and booking language change with it. You still need to know enough about the trade to ask good questions on the call, so read their website and three of their competitors' before you dial. The verticals list covers over 100 trades, and anything appointment-based that is not on it gets a deliberately industry-neutral pack.
Should I give away a free build to get my first case study?
Usually no. Free work attracts clients who will not implement, and a case study built on a client who never adopted the system is worthless. Discount the first month or two if you need to, keep the monthly, and keep the obligation on both sides. A client paying something will actually answer your intake questions.
How many case studies do I need before this stops being a problem?
Three in one trade does more for you than ten spread across ten trades, because the question is always "someone like me." That is also the strongest argument for niching early.
What if the prospect breaks the demo on the call?
Good. Write down exactly what they said, tell them you are writing it down, and tell them what the fix is. Half the time the fix is a line in the prompt and you can show them the corrected behavior before the call ends. Watching you fix something in two minutes is better proof of competence than a case study. The failure mode is pretending the agent handled it well when both of you heard that it did not.
Your next step this week
Pick the three prospects you have already pitched and lost to silence. Build a demo for each one, on their trade, with their business name and their website in it, and send a one-line message saying you built the thing you were describing and here is the link. No pitch in the message. You are not asking for a meeting, you are handing them something to click.
That tests this whole argument on people who already told you no, which makes it cheap. If you want the stack to do it with, start an agency workspace at $49 a month, or read the full FAQ first, including the parts that are not flattering.
