Guides
The agentic demo for work that can't be screen-shared
An agentic demo answers the buyer's questions and shows them the product, with nobody on the call. If what you sell is a room, a lot or a building, your version places the design in their space and hands you the lead.
Your portfolio is a brochure that happens to load fast
It cost real money. The photography is good, the projects are the right projects, and it does exactly what you built it to do, which is prove you are competent.
Then a stranger lands on it at ten at night with one question in their head, and it is not the question the page answers. They are not asking what you make. They are asking whether the studio fits down the side return, whether that scheme would look absurd under their ceiling height, whether the pergola swallows the only patch of sun they get. A gallery cannot answer any of it, because every photograph on it is of somebody else's house.
So they close the tab and decide to think about it. Thinking about it is where projects go to die, and you never learn it happened, because the people who leave do not leave a name.
The part that should worry you is how little of the decision you were ever present for. Gartner surveyed 750 buyers on complex purchases and found they spend roughly 17% of their total buying time meeting suppliers, and that is all suppliers combined. With three firms in contention, any one of you gets about 5%.
Out of all the time a buyer spends on a purchase, 17% is spent with suppliers
Those figures come from enterprise B2B, not from a couple choosing a landscaper, and the two are not the same purchase. What survives the translation is the shape. Anything expensive and slow gets decided in the gaps: weeks of quiet comparison, a scroll through Instagram at midnight, an argument at the kitchen table, then a short conversation with you at the end. The 83% is where the job is won. For most practices there is nothing standing in it.
What an agentic demo is, and whether yours counts
Software sales solved this first, because it had to. The old motion was a Request a demo button, a calendar link, a four-day wait, and a rep sharing their screen through a script they had recited two hundred times. It worked and it did not scale, and it put a four-day hole exactly where interest peaks.
The replacement arrived in stages: recorded walkthroughs, then click-through interactive tours, then agentic demos. The category is now well enough established to have its own vendor shelf. Walnut's 2026 review of B2B demo platforms names agentic demos as one of the two highest-velocity trends in the category, and the working definition across those vendors is consistent: an adaptive, AI-guided experience that understands what the visitor is trying to do, selects the most relevant proof, and determines the next best action, whether that is routing to a person or triggering follow-up.
Read that definition against what happens on your own site and the fit is closer than you would expect. Three things have to be true.
It works out what the visitor wants. They ask whether you do anything narrow enough for a terraced garden, and the agent surfaces the two models in your catalog that qualify rather than the full range. Nobody navigates a menu. The catalog reorganises itself around the question.
It acts rather than answers. This is the load-bearing distinction, and worth stating bluntly because the word agentic is doing a lot of unpaid work in marketing right now. A chatbot answers: ask about the footprint, get a paragraph with dimensions in it. An agent takes the next step for you. It offers to place that model, takes an address or a photo, and returns the visitor's own garden with the studio standing in it. The conversation produced something.
It decides what happens next. The images are the moment to ask for an email, so it asks then, once, and the lead lands in your dashboard with the design and the location attached rather than as a name in an inbox.
The obvious objection
If your product is a building, there is no live interface to drive. A SaaS agent demos the real software. Yours cannot, because the thing you sell does not exist yet and will not exist for nine months. So is it a demo at all, or just a picture?
Take the objection seriously, because a prospect will. The honest answer is that the demo has moved from the product to the placement. What a SaaS buyer needs to resolve is can I work this thing, so you show them the interface. What your buyer needs to resolve is does it work here, so you show them here, with it in place. Both replace a general description with a specific instance, which is the only thing that ever moves a decision.
It is fair to say the artefact is a depiction rather than the article. Say so on the page. A scaled, four-angle visualization across day and dusk is a picture of an idea, not a survey and not a quote, and every client you have ever had already understands the difference between a render and a building. Trust survives the caveat. It does not survive the discovery that you skipped it.
Two more limits worth owning. It does not close: it takes the questions you answer identically every week and hands you the ones needing judgement. And it knows nothing you have not told it, which is a feature, because an assistant improvising about your lead times will eventually improvise a promise you have to keep.
The clock you have never once beaten
Suppose the gallery works and someone fills in the form. The clock that decides whether they become a client starts the moment they hit send, and you are up a ladder with your phone in a bag in the van.
Relative odds when an enquiry is answered in five minutes rather than thirty
The Lead Response Management study run by James Oldroyd with InsideSales found the odds of making contact with a web enquiry were around 100 times better when the first attempt came within five minutes rather than thirty, and the odds of qualifying it around 21 times better. Oldroyd's later Harvard Business Review write-up found the average company took roughly 42 hours to respond at all.
That research is old, it was built on phone follow-up, and the multipliers are relative odds rather than a forecast for your pipeline. Treat it as the shape of a curve. The direction has never been seriously contested, and buyer patience has not improved since 2007.
You cannot fix this with discipline. Nobody ever has. You fix it by having something competent answer the routine half of the conversation while it is still happening, and by making sure that something knows what you would have said.
Nine things that decide whether yours works
Decide what the visitor should leave having seen
This is the only genuinely hard step, and skipping it is why most setups feel vague. Write the sentence out before you touch a single setting: "a visitor should leave having seen ______ in their own ______."
For a prefab or ADU maker that is a specific model on their lot. For an interior studio, a scheme in their room. For a remodeler, the same room rebuilt. For a lighting designer, one space at dusk. Everything downstream follows from that sentence: which products you load, whether a photo is required, what the greeting says. If you cannot write it, no software will rescue you.
Give the agent the answers you give on the phone
The knowledge base is not your About page and it is not a brochure. It is the twenty answers you have given out loud this month. Roughly what a project this size costs. How far you travel. What the lead time actually is right now, not last spring. What happens in week one. What you will not take on.
Write those in plain sentences. Scanned pages and uploaded PDFs are worth having, but they carry the published version of your practice, which is the polished one. The unpublished version is what turns a browse into an enquiry, and no scraper can reach it because it only exists in your head. Budget half an hour. It is the highest-yield half hour in the whole setup.
Load a real catalog, not a sampler
The agent can only offer what is in front of it, so three products with four good photographs each beat twenty half-finished ones. Write each short description for a client rather than a spec sheet, because it does two jobs at once: the visitor reads it, and the visualizer uses it.
Then fill in the internal description as well. That field never appears in public and exists to tell the visualizer what a client would never read: materials, proportions, how the thing should sit on a slope. It is the cheapest quality improvement available and almost nobody uses it.
Let the visitor bring the space
The argument collapses if the demo runs on a stock photograph. Two inputs make it theirs: an address, which pulls street and satellite imagery of the actual plot, or a photo taken on the spot.
Interiors always need the photo, because a room nobody has seen cannot be rendered. Exteriors can go either way, and in practice the address route converts better for one unglamorous reason: someone standing in their kitchen at ten at night will type a postcode and will not go outside to photograph the garden in the dark. Offer both and let them choose.
Ask for the email at the moment of imagination
A contact form at the foot of a page asks for commitment before the visitor has any. The form is not badly designed. The moment is simply wrong.
The right moment arrives about four seconds after the images appear. They want to see them properly, keep them, send them to someone. Asking there is a trade, and it reads as one. Ask once, never twice, and then get out of the way.
Write the guardrails, then read what the agent actually receives
Guardrails are where you stop it sounding like everyone else. Keep them short and behavioural: be concise, say a team member will confirm anything uncertain, never discuss competitors, never quote a fixed price. A thousand characters of clear instruction beats three thousand of hedging.
Then open the full system prompt preview underneath and read it. Most people never do. It is the only way to see the assistant as it actually exists rather than as you imagine it, and anything wrong in there is wrong in every conversation you will ever have.
Test it as a stranger before you test it as the owner
Your instinct is to ask the questions you know it can answer, which teaches you nothing. Use the Playground properly and ask the awkward ones. How much. Do you work in my area. How long will it take. Can you do it cheaper. Why are you more expensive than the other quote I have.
Read the answers as a nervous homeowner spending real money, not as the person who wrote the knowledge base. Every answer that reads as evasive is one you can fix in five minutes now or lose leads to for a year. The honest tradeoff: Playground activity draws credits exactly like a real visitor, so testing has a small real cost. Pay it anyway.
Read the conversations that stopped
The Conversations tab holds every transcript with the contact details and the visualizations attached. Read ten in your first fortnight, and not the ones that converted. The ones that stalled.
You are looking for a single thing: the question it fumbled. That question is a hole in your knowledge base, and it is almost always something you would have answered instantly in person. Paste the answer in. Do that four or five times and the assistant stops sounding generic, because it now knows things only you knew. This loop is what separates a widget from an asset.
Measure four numbers and ignore everything else
Sessions and impressions tell you nothing about whether it pays for itself. Watch four: how many people opened it, how many generated a visualization, how many left contact details, and how many became a conversation you would have taken anyway.
The funnel view tells you where the drop happens, which tells you what to fix. Heavy loss at product selection usually means weak photography. Loss between choosing a placement and generating means the wait feels long. Loss at the email means you asked before you had given them anything worth having.
What a visitor can actually do at ten on a Sunday night
| What you offer them | Effort to set up | Works while you are on site | What the visitor leaves with |
|---|---|---|---|
| Gallery and contact form | Already built, already failing | Technically yes | Admiration, and no reason to come back |
| Recorded walkthrough video | A day of filming and editing | Yes, identically for everyone | Your project, not theirs |
| Booked consultation call | Low, but it recurs forever | No, only in your calendar | A date four days out, if they wait |
| Live chat with a person | Cheap until it is not | Only during office hours | An answer, if someone is at a desk |
| Agentic demo with placement | An afternoon | Yes, and specific to their space | Their own property with your design in it |
Frequently asked questions
What is an agentic demo?
An agentic demo is a product demonstration run by an AI agent rather than a salesperson. It works out what the visitor is trying to do, picks the most relevant proof from what you have loaded, and decides the next step, all without a call being booked. Two visitors asking different questions get two different demos.
Can you run an agentic demo if your product is physical?
Yes, though the mechanics differ. A software agent drives the interface the buyer would actually use. A design or build agent generates the buyer's own room, lot or address with your design placed in it. Both replace a description with a specific instance, which is the part that changes the decision.
How is that different from a chatbot on my website?
A chatbot ends in a paragraph. An agent ends in something the visitor asked to see: your design, on their property, from four angles. The commercial difference is that nobody trades an email for a paragraph.
Does it replace the consultation call?
No. It handles the half of the conversation you repeat every week, covering what fits, roughly what it costs, how far you travel and what the lead time is. It hands you the conversations that need judgement, with the design and the location already attached.
How accurate are the visualizations?
They are scaled and photographic, generated from street and satellite imagery of the address or from the visitor's own photo, and returned as four angles across day and dusk. They are a picture of an idea, not a survey, a planning drawing or a quote. Say that to visitors and the trust holds.
What does it cost to run one?
AIBCO's free plan includes 50 credits a month and up to 7 products, which covers roughly 8 visualizations or 50 chat messages. Paid plans start at $29 a month for 300 credits. One visualization returns four angles of the same placement for a single charge.
Where to start
Pick one product. Load it properly, four photographs and a description written for a client rather than a spec sheet. Write your twenty phone answers into the knowledge base. Publish the hosted page, put the link in the one place your traffic already lands, and read the first ten transcripts.
That is an afternoon, and it is the whole thing. The people spending 83% of the decision without you are not waiting for better photography. They are waiting to see it in their own space, and whoever shows them first sets the terms everyone after gets compared against.
If the site itself is the thing holding you back, the companion piece on marketing a design practice without a good website covers the rest of the routing problem.
Gartner survey of 750 B2B buyers on complex purchases (2019) for the 17% figure and the 5% to 6% per-supplier estimate; Gartner press release, March 2026, covering a survey of 646 B2B buyers conducted August to September 2025, for the 67% rep-free preference and the 45% AI usage figure, alongside the 61% recorded in the equivalent 2025 survey; Shopify's product changelog (2022) for the average 94% conversion lift among merchants adding 3D content to their stores; the Lead Response Management study by James Oldroyd with InsideSales (2007) and Oldroyd's Harvard Business Review follow-up (2011) for the response-time multipliers and the 42-hour average. The description of the agentic demo category draws on vendor category writing from Walnut and Supademo (2026), which is useful for definition and worthless as evidence of results. The Gartner figures describe complex B2B buying rather than homeowners commissioning design work, and the Shopify lift comes from ecommerce catalogs rather than design services. Both are offered as the direction of an effect, not a guaranteed outcome. Individual results vary.