Scalewell / Blog

LinkedIn video ads for B2B with AI: scripts by role, versions from one shoot, and tests small audiences can finish

How to script B2B video ads by buying committee role, cut versions from one recording with AI tools, and test them in audiences too small for LinkedIn's A/B tool.

Illustration of one filmed expert whose footage is cut into a few video frames in different shapes, with two chosen for testing

AI makes LinkedIn video ads cheaper to produce, but a B2B audience can only judge a few versions at a time. So write one script per buying committee role, record once, let AI tools cut that footage into a short list of lengths and ratios, and test one change per round against view-through and pipeline. It works for YouTube too.

You've probably sat through the demo. An AI video ad generator turns a product page into 30 polished versions before lunch, and then someone remembers that your ICP on LinkedIn is, say, 70,000 people. Split thirty ways, each version reaches about 2,300 of them at most, nothing reaches a verdict, and the "winner" is whichever ad happened to land in front of a bored CFO on a quiet Thursday.

This guide is for the paid media or demand gen lead running LinkedIn and YouTube for a B2B company: scripting by role, the AI tool types that cut versions as of November 2026, where the files go in each platform, small-audience tests, and disclosure rules for synthetic people and voices. Static images and ad copy have their own guide on AI generated ads for B2B, so this one stays on video.

AI made video versions cheap, so the scarce resource is audience

The platforms now generate video for you, so the hard decision is which few versions deserve the audience. Google's March 2026 announcement says Veo, its video model, is available in Asset Studio inside Google Ads, where you "upload up to three static images" and Veo generates a video "up to 10 seconds long" from each one. For AI YouTube ads, Google Ads also creates video enhancements on its own, extra versions in other orientations, shorter cuts chosen by Gemini, and English voice-overs, and the help page says "They're turned on by default."

LinkedIn generates less and is stricter on testing. Its AI drafting in Campaign Manager, which can generate "headlines, intro text, and suggested visuals," is documented only for single image ads, not video, and its Brand Kit says "The single image ad format is currently the only available ad format for brand kit." Meanwhile LinkedIn's formal A/B test won't start unless "each variant in the test must have an audience size of at least 1,000,000," according to its A/B testing API documentation.

Once AI handles the scripts, cuts and test readouts, the hours that used to go on video editing come back to your team, and Scalewell helps paid teams set that workflow up while they keep deciding which versions earn a place in front of buyers.

Script one video per buying committee role, from a single recording

B2B video advertising works better when each cut speaks to one person on the buying committee, because the CFO, the person who'll use the product and the security reviewer each need a different first five seconds.

The efficient way to get role-specific video is to record one subject matter expert, such as a product lead or a customer success manager, answering four questions in one session, then have AI write the scripts around those answers and cut the footage. Choose a real person who knows the product over any synthetic presenter (the failure list below shows why), and four answered questions with a few retakes each hold more cuts than a small audience can test in a quarter.

RoleWhat the first 5 seconds must sayProof that worksCall to action
Economic buyer (CFO, VP)The cost of the problem in their termsA named customer result or a calculatorPricing page or ROI calculator
Day-to-day userThe task that gets easier, shown on screenA real product screen recordingProduct tour or free trial
Technical or security reviewerThe risk you remove and howCertifications, architecture, integrationsDocumentation or security page
Champion (the person selling it internally)The argument they can repeat to their bossA short quote from a peer in the same roleCase study or demo

Google's YouTube creative research gives each script its structure. Its ABCD framework says to "Draw attention from the beginning," "Integrate your brand naturally," "Connect with emotion and storytelling" and "Clearly state a call to action," and Google reports that ads following it showed a "30% lift in short-term sales likelihood," based on a Google and Kantar analysis of 11,000 ads from April 2021. That sample is mostly consumer advertising, so treat the number as a reason to follow the structure rather than a forecast for your pipeline. LinkedIn's own video ad tips point the same way: "Show what you want your audience to see in the first 10 seconds of the video," because "Viewer attention drops after that point."

Give an AI assistant the recording's transcript and your sales notes, and ask for scripts in this shape:

You are writing B2B video ad scripts for LinkedIn and YouTube.
Attached: transcript of a recorded interview with [name, role], plus
[win/loss notes or call transcripts] and our ICP definition.

For each role (economic buyer, day-to-day user, technical reviewer, champion):
1. Pick the strongest 2 to 4 sentences the speaker actually said that answer
   that role's main question. Quote them exactly with timestamps.
   Do not write new lines for the speaker.
2. Write an on-screen text line for the first 3 seconds (max 8 words) that
   works with the sound off.
3. Lay out three cuts: 6 to 15 seconds, about 30 seconds, and 60 to 90 seconds,
   listing the timestamps to use in order.
4. Name the proof shown on screen (customer name, product screen, certification)
   and flag any claim that needs approval before it runs.
5. End each cut with one call to action from this list: [CTAs].
Return a table per role. Flag any role the transcript cannot support.

The instruction to quote the speaker exactly keeps the AI as an editor of what a real person said, which is the line between a fast workflow and a synthetic testimonial.

Cut the versions with AI tools, sorted by the job each does

AI video ad tools fall into six types, and most B2B teams need two or three. The table describes each type as of November 2026, naming one vendor only as an example from its own website; it isn't a ranking.

Tool typeWhat it doesExample of the typeWhere it fits in B2BMain risk
Long-video clippersFind highlight moments in a long recording, cut them into short clips, caption and reframe themOpusClipTurning a webinar or the expert recording into 6 to 30 second cutsPicks moments that are lively, not the ones that answer a buyer's question
Text-based editorsTranscribe footage so you edit video by editing the text, and remove filler wordsDescriptFast, precise cuts from the script table aboveLittle, as long as a person checks the final cut
Avatar presentersTurn a script or document into a video read by an AI avatar, in many languagesSynthesiaProduct walkthroughs and localized explainersReads as cheap to technical buyers; disclosure rules apply
Voice and dubbing toolsTranslate a real speaker's audio into other languages in a clone of their voiceElevenLabsTaking one English recording into other marketsNeeds the speaker's written consent
Generative footageCreate new moving footage from images or promptsVeo in Google Ads Asset StudioBackground motion and product scenes for YouTubeGeneric look, rendering errors on screens and hands
Platform-side automationThe ad platform makes extra versions from your assetsGoogle video enhancements and auto-generated videosCovering ratios you didn't produceVersions you never reviewed run under your brand

The output of this step is a cut sheet: every role, at three lengths, in the ratios each platform wants. LinkedIn's video ad specs recommend 4:5 vertical and say "We recommend uploading videos that are 15-30 seconds long" so they qualify for every feed and Audience Network placement. Google's Demand Gen video specs list 9:16 as "Recommended for YouTube Shorts." In this example the cut sheet holds 36 files (four roles, three lengths, three ratios), a library to draw from two at a time.

Branching diagram: one source splits into four branches, then twelve, then thirty-six tips, with two highlighted paths leading to a doorway.
One expert recording, four role scripts (economic buyer, user, technical reviewer, champion), three lengths (6 to 15 seconds, about 30 seconds, 60 to 90 seconds) and three ratios (4:5, 16:9, 9:16) make a hypothetical library of 36 cuts (4 x 3 x 3). Only two cuts enter each test round.

Upload LinkedIn video ads in Campaign Manager and YouTube ads in Google Ads

LinkedIn video ads need the file uploaded directly to LinkedIn, and YouTube ads need the video on YouTube first, so the two platforms start from different places. LinkedIn's specs page says "Video ads must contain a video file uploaded directly to LinkedIn," which rules out pasting a YouTube link.

In Campaign Manager, per LinkedIn's video ad setup page:

  1. Create or open an ad set with the brand awareness, website visits, engagement, video views, website conversions or lead generation objective.
  2. On the Ads in this ad set page, click Create new ad(s), then Upload or select to add the video.
  3. Under captions, click Upload file and add an .SRT file. LinkedIn says "Only English video captions are currently supported," so burn subtitles into the video for other languages.
  4. Add the headline, destination URL and call to action, or pick a Lead Gen Form if the objective is lead generation, then click Save ad.

The specs page sets the limits: three seconds to 30 minutes, 75 KB to 500 MB, MP4, and up to 25 uploads in 24 hours from a desktop device, so the 24 LinkedIn files (4:5 and 16:9) fit in one day.

In Google Ads, a Demand Gen campaign is the usual home for B2B YouTube video because it serves on YouTube, including in-stream and Shorts, and offers a Leads goal at setup. Per Google's Demand Gen setup page, click Create, then Campaign, pick your goal, and at the ad group choose New ad. On the Videos card select + Videos and pick from your YouTube channel. The specs page allows one to five videos per ad and a 5-second minimum, warns that clips under 10 seconds can't run as YouTube in-stream ads, and accepts unlisted videos if you don't want them showing on your channel.

Decide before launch whether Google's AI versions are on. Demand Gen video enhancements are switched off per ad, in the ad builder under Review campaign, on the Creative optimizations card. Google's auto-generated video ads have an ad-level setting in Demand Gen, and in Performance Max they stop once you upload your own videos. During a test, turn enhancements off for both versions, because a Gemini-shortened cut of version A competing against your edit of version B is no longer the test you planned.

Test B2B video ads one change at a time in audiences far below a million

For a B2B audience too small for LinkedIn's formal A/B test, run two versions inside one ad set with even rotation, change one element, and read the result on a metric your volume can support. LinkedIn's ad rotation page describes "Rotate ads evenly" as entering "each creative into the auction evenly, giving the ad a fair opportunity to compete for impressions," and you set it when you create or edit an ad set. Brand awareness ad sets use even rotation automatically.

Even rotation is fairer than optimized delivery but is still no true audience split, so treat a result as strong evidence rather than proof. Our guide to AI creative testing has the sample size arithmetic and the readout prompt; run its pre-test check before every round so you know the test can finish on your budget.

Run the rounds in this order:

  1. The opening. Test two first-five-second openings on the same short cut for one role. The opening decides whether anyone watches the rest, and two openings can differ a lot, so this is the round most likely to show a difference large enough to detect.
  2. The length. Put the winning opening on a 6 to 15 second cut against the 60 to 90 second cut. LinkedIn's tips page reports, from its 2025 data, "up to a 300% lift in completion rates" for 7 to 15 second videos, but also that "longer form content drove as many clicks and views as shorter videos," so your audience has to answer it.
  3. The role match. Run the economic buyer cut against the user cut to a single job function segment, to see whether role scripting pays for its production time.

To read the numbers in Campaign Manager, click Columns, then Customize Columns, and tick the metrics in the Video section. LinkedIn's video metrics page counts a view as "Two or more continuous seconds of playback while the video is at least 50% on screen," a completion at "97-100% of its length," and completion rate as completions divided by video impressions, a definition that changed from plays to impressions in March 2025. For an opening test, use views at 50% divided by impressions as the decision metric; for a length test, add average watch time, since a 90-second ad will usually complete less often.

The test has worked when three things are true: the winning version clears the confidence level you set before launch (90% is a sensible default), the result came on the end date you planned rather than when someone peeked, and the CRM check doesn't overturn it. That last step uses LinkedIn's video retargeting audiences, which the engagement retargeting documentation offers at 25%, 50%, 75% and 97% viewed, with lookbacks of 30, 60, 90, 180 or 365 days. Build a 75% viewers audience for each version, follow them with a demo or case study offer, and check after one sales cycle which version's viewers became opportunities. Small audiences also tire quickly, so watch frequency with the check in our ad fatigue guide.

On YouTube, the same two-version approach inside Demand Gen is the realistic option for most B2B budgets; the testing guide's table sets out the limits of Google's own video experiment tools.

Disclose synthetic people and voices before a platform or a regulator does it for you

AI video carries more disclosure risk than AI images, because it's where synthetic faces and cloned voices show up. Google Ads now has an AI content label setting: go to Asset studio in the Tools menu, open the Asset library, select an asset and choose whether to label it. Google says labels "will also show on the ad itself" in the European Union, India, New York, California, South Korea and Vietnam, that it may label assets made with its fully automated features on your behalf, and that the setting "doesn't guarantee your compliance with specific regulations."

LinkedIn shows a C2PA icon on single video ads when the uploaded file carries Content Credentials, and its help page admits "it's not yet possible to identify and label all AI-generated and modified content." LinkedIn's advertising policies have no AI-specific clause, but they say "Do not imply you or your product are affiliated with or endorsed by others without their permission," which a generated "customer" would breach.

New York's General Business Law section 396-b, in force since June 9, 2026, requires anyone who creates an ad with actual knowledge that it contains a synthetic performer to "conspicuously disclose" it, with civil penalties of $1,000 for a first violation and $5,000 for each one after; audio ads and AI used solely to translate a human performer are excluded. The FTC's 2024 rule targets testimonials that "misrepresent that they are by someone who does not exist, such as AI-generated fake reviews," per the FTC's announcement.

Write down three rules. Customers, employees and experts are always real people with signed releases. An AI avatar may present a product walkthrough, labeled, and never plays a customer. A real speaker's voice may be dubbed into another language with their written consent. The wider policy sits in our guide to generative AI in advertising, and the pre-launch checks belong in your creative approval workflow.

Where AI video ads go wrong in B2B

Most AI video marketing failures in B2B come from making more video than the audience can judge, or from letting automation publish what nobody reviewed.

  • Testing the whole cut sheet at once. Twelve versions in a 70,000-person audience produce twelve noisy numbers and one confident slide.
  • Avatars in front of engineers. A synthetic presenter explaining an API to a technical buyer tells them, within seconds, how much the vendor invested in reaching them.
  • Vertical crops that cut off the point. An automatic reframe can crop on-screen text or a product screen out of a 9:16 version.
  • Captions that misspell the product. Auto-captioning gets brand names and acronyms wrong, and LinkedIn's tips page says "a large portion of LinkedIn members will watch your ad with the sound off."
  • Cheap views from the wrong people. A version that wins on view rate by pulling in students and job seekers loses in the CRM.

On a few thousand dollars a month for one audience, tests will rarely conclude, so use the role scripts and cut sheet anyway, rotate on judgment and the CRM check, and save tests for the opening.

Where to start

Book one recording session with your best product expert this month, run the scripting prompt on the transcript, and launch two openings for one role with even rotation. If you want that loop to run every month with AI doing the cutting and the reading, Scalewell sets it up while the team still chooses what gets tested and what gets cut.

Frequently asked questions

Are LinkedIn video ads worth it for B2B?

LinkedIn video ads are worth it when your buyers are hard to reach elsewhere and you judge them over a full sales cycle. Treat video as the way accounts get to know you before they're in market, and measure it by opportunities among viewing accounts, not first-month cost per lead.

What is a good completion rate for LinkedIn video ads?

There's no reliable public benchmark, and many published figures divide by plays, while Campaign Manager now divides by impressions, so they don't compare with your numbers. Use your own account's history by length and audience as the baseline, and compare versions of the same length only.

Can I run AI avatar videos as LinkedIn ads?

Nothing in LinkedIn's advertising policies bans AI avatars, as of November 2026, but they can't imply an endorsement that doesn't exist. Label avatar videos, never present one as a customer or employee, and check local law, since New York now requires advertisers to disclose a synthetic performer they know an ad contains.

Do I need separate videos for YouTube Shorts?

You need a vertical 9:16 version, which Google recommends for Shorts. Google's video enhancements can create vertical versions from horizontal video automatically, but an automated crop can cut off text and product screens, so cut the vertical version yourself and review it like any other.

How many video ads should I put in one LinkedIn ad set?

LinkedIn's help pages suggest two to five variations per ad set and allow up to 100 creatives. For a small B2B audience, run two at a time when you want a readable result, adding more once the opening and length are settled.