AI Adoption GuideSalesPropose
Personalized demo video
AI generates tailored video walkthroughs per prospect using an avatar or product clips, with tools like Tavus or HeyGen.
Sales processProspectQualifyDiscoverProposeNegotiateCloseHandoffRenew
By Don, DoneThat’s AI coach · updated
A short walkthrough is the job, not a generated demo
Personalized demo video is a speed play after discovery: assemble a short walkthrough from approved product clips and a few CRM fields, then let the AE watch it and send it.
It does not replace a live demo. It covers the buyer who asked about one workflow, and the committee members who never joined the call. Name their hook, show the matching SKU, and stop.
Generation tools in this class include Tavus and HeyGen. They put an avatar or a voice over product footage. This page does not rank them. An enablement library such as Highspot is where those clips should already live, tagged and approved. Salesforce is where the opportunity product, the contact, and the send record belong. If a stack also auto-sends, speaks as the AE, and flips stage to Demo Completed on a play, you have left this use case.
The person on screen is a production choice. A stock avatar is not the AE. A cloned AE face is still not a live take recorded for this account. Label a synthetic presenter in the opening seconds, or do not use a face the buyer will treat as the person who will join the next call.
Pull one hook from discovery and only the matching clips
Build from three inputs. Do not start from a prompt that says "make a personalized demo for this account."
The hook. One thing the buyer said, or a priority written on the opportunity. "The technician still emails a spreadsheet quote from the driveway" is a hook. "You are growing fast" is not. If discovery notes are empty, do not invent a pain. Wait, or send nothing.
The SKU. The product, edition, or module on the Salesforce opportunity, not the flagship suite in last year's tour. If the product field is blank, block generation and flag the AE.
The clip pack. Screen recordings already approved by product marketing, stored in Highspot or the same kind of library, tagged by SKU, edition, and language. The generator may stitch and narrate. It should not invent UI, or claims that are not on those clips.
Sequence the cut: hook, matching workflow, next step the AE will honor this week. A working session, a sandbox, or a written proposal are real closes. "Let me know if you have questions" is not. Do not recap the account. They know who they are. A pre-call account briefing is for the AE before the next live meeting, not for the first minute of the video.
Keep it short. A twelve-minute tour with a first-name slate is the generic recording with a mail-merge on top. If you cannot show the named workflow without walking the rest of the platform, you are missing an approved clip. Record that clip once, get it approved, then generate. Do not ask a model to hallucinate screens.
Customer proof on screen only if case-study retrieval RAG returns an in-policy asset for this vertical. No match means no logo wall. Inventing a customer in a video is worse than a missing beat.
Deal numbers belong in the ROI business case generator for this opportunity. Do not narrate typical savings unless the AE already used those figures with this buyer. The video does not carry scope, price, or terms. That package is still an auto-drafted proposal. Send the cut as a companion, not as a substitute SOW.
The same grounding rule as hyper-personalized outbound copy applies here: name a real, current fact from this buyer, then stop. Video does not get a longer leash because it looks more produced.
Worked example: Jordan Hale at FieldKit
This is an illustrative scenario, not a case study and not reported results.
An AE at a field-service vendor is proposing to Jordan Hale, Director of RevOps at FieldKit, a mid-market HVAC company. Discovery notes and the Salesforce opportunity say quoting still happens in a spreadsheet after the technician leaves the site. Jordan asked to see on-site quoting, not the dispatch board. Dispatch is a different SKU FieldKit already declined.
The usable cut
The AE reviews a short assembly: Jordan's name and FieldKit on the slate, the driveway-spreadsheet sentence first, the approved on-site quoting clip from Highspot, and a close that offers a working session on a quote from last week. If an avatar is used, it is labeled as a generated presenter. The AE still sends from their own thread.
The hook is Jordan's. The SKU matches the opportunity. The clip is the one product marketing approved for that module.
The unusable cut
A generator returns a twelve-minute "FieldKit personalized demo." A face that looks like the AE (it is not) says "as I mentioned on our call." The recording is last year's platform tour. Dispatch and routing appear anyway. A customer logo from a national telco is on screen; that story is not cleared for HVAC. When Jordan's operations analyst plays the opening, Salesforce flips the opportunity to Demo Completed.
Four failures, one send: wrong SKU, fake AE face, a tour nobody asked for, and a stage change that pretends a view is a demo. The AE should kill that cut and rebuild from the quoting clip and the named hook.
If Jordan's VP never joined discovery, this async cut is for them. It is not a substitute for the live working session Jordan asked for. Do not skip that session because a video went out.
Wrong SKU, a fake AE, a twelve-minute tour, or a view marked complete
Reject the draft if the SKU on screen is not on the opportunity, if the face pretends to be the AE, if the cut is a full product tour, or if you are about to mark a view as Demo Completed.
A mismatched clip is a product claim. The buyer will assume the proposal includes what they saw, or that nobody listened. Tag every clip with SKU, edition, and language so a generator cannot pull footage whose product is not on the opportunity. If the SKU field is empty, do not generate. Red-flag and wait.
A face that is not the AE, presented as the AE, is a disclosure problem. Buyers notice. Some will not say so. They will treat the next live call as a bait-and-switch. Use a clearly branded presenter, a voiceover over product UI, or an AE likeness only with that AE's consent and an on-screen label. Never imply the AE recorded this take for this account unless they did.
Length is a respect problem. A twelve-minute tour signals that personalization was a name plate on a generic asset. Cut until the video is only the workflow they named. If the library cannot support that, stop generating and record the missing clip.
A view is not a completed demo. Playback in an email tracker, a video platform, or Highspot is a content event. The demo stage in Salesforce should move when a qualified buyer finished a live or agreed working session, under the definition your RevOps team already uses. Treating a few seconds of play as Demo Completed poisons forecast and tells the AE to stop selling.
The AE reviews, then the AE sends
Volume is what AEs can watch, not what the generator can render overnight.
Put the draft next to the opportunity, not in a side portal nobody opens. The AE confirms hook, SKU, names, disclosure, and next step, then sends from the thread the buyer already uses. Enablement owns the clip library, the tags, and the template. Enablement does not own the send.
Do not auto-advance stage on send or on view. Log the asset against the contact if you want a trail. Keep Demo Completed as a human field.
If the first dozen cuts come back with the same edit (wrong edition, leftover "as I mentioned," a suite tour), fix the template and the clip tags before you raise volume.
This use case pays back when the bottleneck is assembly time after discovery, on deals where a short async artifact helps a committee catch up. It does not pay back when the buyer asked for a live demo, when the SKU is still unclear, or when you have no approved clips for the module on the opportunity. In those cases, record one honest walkthrough, or wait.
Is this worth automating for you?
Whether this pays back depends on how much time it takes your team today. Most teams estimate that from memory, and the estimate is usually wrong in one direction or the other.
DoneThat reconstructs where the time actually went, with no timers to forget, so you can measure the baseline before committing to a project and check the gain afterward.
Measure the baseline first