News

The AI/Tech Horizon

The AI/Tech Horizon

Monday, November 16, 2026

Where Are We and Where We Are Headed.

Triangulated by Sterling Vernon
The attention paid to technology tools in general, and artificial intelligence in particular, has been relentless. BBS wanted to examine this topic specifically as it relates to advertising production — what the state of the art is today, and where it is headed. That meant looking back over fifty years of how this work has actually been done, and then looking ahead one to two years. Anything past that felt like an overreach.

We have identified four areas where technology changes can aid the process and help the budget.

  • 1. Previs is becoming more streamlined: the cost of showing an idea before anyone commits to making it is falling. 
  • 2. Improved meeting capture and note taking software can help make sure that client direction conveyed in a round of review actually is addressed. 
  • 3. More concepts can be put in front of real audiences, and earlier in the process. 
  • 4. The range of shots generative tools handle well keeps widening, while a specific set of them still struggles.

We looked at each topic area through two sets of eyes, a procurement professional's and a CMO's, and asked the same three questions: can the work be done faster, can it be done for less, and is something now possible that was not possible before — which can be worth paying more for. Procurement often has additional concerns, especially during contract renewals: what capabilities should be secured during sourcing so an agency or production partner has the option to develop more concepts earlier and to a higher standard if needed, for instance.

On where the technology goes next, our crystal ball is no better than yours. On what these changes do to a production budget, BBS has forty-one years of reading them, and we will say which calls we are confident in and which are speculation. Tell us where you think we have it right, and where you think we have it wrong.

Previs Costs Are Falling

Animatics have been part of advertising since the 1970s and previs has been an important development step for commercials since the mid-2000s. Budget often determines how much of these intermediate steps can be included in the process for each spot. An advertiser with five directions on the table may only choose to fund further development of one or two of them into the next step in fidelity. And on smaller jobs, budget constraints may dictate skipping these steps entirely. That can force decisions before anyone knows which concept is actually strongest, injecting rework risk. While the concept artifact is expensive, we are forced to make these compromises. When these artifacts become much less expensive to produce, the need to compromise falls away.

The figures BBS has been able to obtain about how much costs are falling are not yet extensive but the directionality is clear. An AI-built animatic was delivered for Mazda South Africa in two days against a normal minimum of ten, published by the testing firm System1. Lester Platt of STLL Studios put a two-minute proof of concept together in roughly a day against what he described as months of work. Scott Kemper, in an article BBS published, describes refining design ideas, animation ideas and animatics for significantly less cost and with greater speed. The figures demonstrate direction and rough magnitude, and they all point the same way.[1]

We found additional evidence of improvement beyond animatics and within generating video frames. Kingsley Harden, a creative director and VFX supervisor at The Jogger Studios, describes initial generations that look 80 to 90 percent final in 10 percent of the time, and a previs phase in which the edit consists of full-motion photoreal video with actors, sets and lighting represented. Are these frames really that convincing, though? Ipsos, working with Syracuse University's Newhouse School, tested twenty ads in May 2026 — ten made conventionally and ten AI counterparts with no human post-production at all — with 3,000 US consumers. Only 25 percent of viewers correctly identified which were which. Andy Lambert of Dark Energy Films gives the most precise description BBS has found: photorealistic, but essentially animated. And the directors Will and Tim Truesdell, on an Oakley campaign, had what they called a rough cut of the film before they set foot on set. As they put it — the most expensive part of filmmaking is not knowing.[2]

What not knowing actually costs a production is set out in the AICP's own guidelines. Under the National Live Action Guidelines, a change in project specifications is an overage billed as a change order, and the production company is entitled to its production fee on top of that overage. A client-caused postponement inside ten business days costs incurred expenses plus the full director's fee and the full production fee as bid, whether or not anything was shot. Ouch. Changing direction late is priced almost independently of how much work the change saves. Against that, the BetterBriefs Project found 90 percent of marketers saying their brief changes after briefing, with rebriefing routine for roughly seven in ten on both the client and agency side. Those two facts are the case for spending more at the concept stage, both true before generative tools arrived. Generative tools are capable of reducing what that additional concepting costs.[3]

So BBS's call is to recheck the budget to see if the development of more concepts under serious consideration can be expanded, within reason, to reduce risk. The arithmetic has changed and you may have new options. An advertiser can evaluate a wider range of ideas that, in the past, may have felt too risky. A concept within the consideration set that previously went without an animatic may now be able to have one.

This may also increase creative quality. When a concept artifact is expensive, the ideas that reach a client are inevitably influenced by how cheaply they can be demonstrated. A simple idea with one location and one performer could always be shown for less than an idea calling for three environments, a crowd and a visual effect, so the complicated idea lost ground before anyone judged it on merit. As the cost of demonstrating complexity falls, more ambitious ideas may remain in the consideration set.

The level of finish for concepting artifacts becomes a second decision, and just because photorealism (or "high-fidelity") is more possible now, it shouldn't be a default. Vendors selling AI animatics argue that rough stimulus forces a respondent to mentally construct the finished ad, which introduces bias. Pedr Howard, Head of Creative Excellence at Ipsos US, who sells testing rather than animatics, says close to the opposite: animatics often score higher than fully produced films, and the best performing animatics he has seen are the most basic and the cheapest. As more art than science, we know that a controlled comparison here is an impossibility. What is new is having various finish levels almost equally available.[4]

Our own experience is that lower fidelity (non-photorealism) serves important purposes beyond cost. It separates decisions, so that each one gets proper focus. A rough previs lets a client approve the idea itself without being asked to judge script, performance, lighting and camera angles at the same time — all things a high-fidelity artifact has to commit to whether or not those decisions are ready. A boardomatic restrains the conversation to staging and story. Harden, the VFX supervisor at The Jogger Studios, still prefers illustrated boards with lighting and color over photoreal video for mood work, to leave room for interpretation. Runway, which makes one of the generative video tools in common use, says of its own product that rough is often better for disposable previs because it keeps attention on staging rather than finish.[5]

Generative tools change fidelity from whatever the budget allowed into a choice available at each stage. That choice now has to be made deliberately, because high fidelity has become the inexpensive option, and inexpensive options get applied without anyone deciding to.

Looking into the future: if initial generations can already arrive at 80 to 90 percent of final quality, a step appears between the animatic and the finished spot that has not existed before — an artifact developed enough to direct from, and on some jobs nearly close enough to run. At BBS, we have started to see concept artifacts approach quality that can be considered for running as paid advertising, in very limited circumstances, of course. As the quality of generative output increases and the specific areas where it struggles can be avoided while producing the right final product, large savings in production can be the result.

Higher fidelity certainly has its risks. A recent study looked at a specific danger called early aesthetic closure. When presenting AI-generated images that look nearly finished, clients treated them as the expected final product, which crowded out other ideas and locked in expectations.[6]

In our own experience, most advertising concepts already reach a client at high fidelity. And three concepts shown together seem less likely to produce fixation on any single detail than one concept shown alone. Matching the right artifacts to a given situation is an art form in itself.

And there is always the "design by committee" risk as the options multiply: pulling pieces from multiple ideas and asking for a combination. It becomes easier to ask for when the options on the table are more developed.

What effect does this have on an agency contract or retainer? There may be some modifications required to better integrate potential benefits of generative concepting. Your creative team still concepts the ideas, but a director may need to be brought in earlier to guide the animatics or the prototype. That director may need to be an added allocation. And, of course, there is the generative development team working alongside creative and the director — perhaps allocated but could be project-based as well. Pulling the director in earlier can enable fidelity with directorial judgment, increasing its value. Many agencies already have favored directors effectively on call or on staff, so it may be a matter of securing access contractually. The work of defining an idea in enough detail to generate it, especially across more concepts than before, may actually increase retainer costs. Getting better creative is the trade-off.

BBS's call on all of this: see if your budget will now cover funding deeper development of additional ideas. Specify the fidelity of each artifact as a scope deliverable rather than negotiating it mid-flight. But be careful not to substitute perceived high fidelity for a final product. And the high-fidelity version may not always be the most helpful way to make decisions.

AI Review Notes Save Cost

We've all been in a meeting wondering why the feedback we provided in a prior meeting doesn't seem to have made it into this subsequent round. AI/Tech tools may be able to help catch lapses. Missed changes get ever more expensive. Under AICP's guidelines a change in project specifications can be an overage billed as a change order with the production fee owed on top of it, so a round that has to be run a second time may cost double. And heaven forbid if it moves an air date or compresses post time leading to overtime and rush charges.

As the number of ideas being reviewed increases, every one of them carries its own notes that have to survive into its own next version. It becomes far too easy to start losing track.

AI tools have the potential to help here, and the layer they would work on is already bought and already everywhere. Microsoft reported more than 30 million paid Microsoft 365 Copilot seats in the fourth quarter of its 2026 fiscal year. Zoom's AI Companion generated more than a million meeting summaries across 125,000 accounts inside two months of launch, and includes it with paid accounts at no extra charge. Three years on from that, meeting capture, transcription and summarization are standard equipment. Both clients and agencies can benefit from adopting these assistive technologies to keep track of the details and catch errors and omissions earlier.[7]

Dedicated review and approval platforms offer this as well. Ziflow's review assistant evaluates a proof against a pre-set checklist, marks each item pass or fail with a rationale, suggests revisions, and tracks its own reliability on each item over time, running inside the proofing interface rather than as a separate tool. The capability that would confirm a later version actually incorporated earlier feedback is one Ziflow says it is actively building. Adobe introduced Brand Intelligence in April, described as learning from review cycle feedback, annotations, rejections and approvals rather than from static brand guidelines — but it has not been released yet. We can keep an eye out for these in the future.[8]

How any of this gets used is going to be personal, and it is going to vary. Someone will ask an assistant to summarize the last review before walking into the next one. Someone will record both sessions and run a comparison. Someone will try analysis live in the room. We do not expect a single best practice to arrive from a vendor, and the thing that will decide which approach wins is whether it is seamless. A practice that happens by default will always have better adoption.

There is a contract question underneath this. A brand's accumulated review history — what it approved, what it rejected, and the notes it gave along the way — is the advertiser's own data, and arguably among its more valuable assets. Where that record is held, and whether the learning drawn from it transfers when the advertiser changes agencies or platforms, is easier to agree before the record accumulates than after.

BBS's call on this one: AI capture tools are already paid for and already deployed, so there is no need to wait for a review platform to deliver verification before putting them to work on creative review. Figure out what works best for you and enjoy the benefits. And maybe ask about where the review record is held on the agency side and who owns the learning that comes out of it.

More Concepts Can Reach Real Audiences

When it costs less to create testable assets, more ideas can be tested with people who can provide a measured response to them. Existing solutions such as Kantar's LINK+ evaluate up to twelve early-stage creatives in a single study, mood films included, positioned for exploring ideas before production budget is committed. And their volume is growing: Kantar's early-stage creative tests rose 17 percent between December 2025 and July 2026. We have also seen at least one documented case of an AI-built animatic going out to nationally representative consumer testing before the shoot — which shows the new pipeline can work.[9]

There is a limit on what a low-fidelity test can tell you, and it has been observed for thirty-five years across separate studies using different methods.

In 1991, a study of 186 finished and 186 rough commercials across ten categories found that rough tested equivalently on message, news and credibility, and significantly lower on mood and entertainment. Executions built on star talent or strong action were penalized the most. In 2026, matched pairs of AI and real food images produced no difference at all in perceived healthiness or calorie content, while realism and willingness to eat both came out significantly lower. And in the Ipsos work described earlier, AI ads held up on functional, problem-solution briefs and degraded on emotional ones, with the conventionally made versions scoring roughly 14 percent stronger on short-term sales impact and 17 percent on long-term brand equity.[10]

Rough or synthetic stimulus gauges the rational response faithfully but understates the emotional one. For an advertiser that is practical rather than academic: a low-fidelity test can tell you whether an idea communicates but not if it inspires. Studying whether the work moves someone requires higher fidelity stimuli.

Pedr Howard of Ipsos adds a condition that is easy to lose. The whole process depends on the finished ad staying close to what was tested, and he is blunt about a common error — elements that tested well getting dropped from the finished film. A test only means something if the thing tested is the thing that airs. That cuts against changing the work after it tests well, which requires uncommon discipline.

When looking at using AI to predict or simulate responses, the research clearly shows that it's still a non-starter. Synthetic respondents — AI-generated stand-ins for actual survey participants — averaged 12.4 points of error. On nearly half the questions there was an answer that no synthetic respondent chose at all. A real audience sometimes picks the option nobody predicted, which is a large part of why anyone asks them. We do not recommend synthetic respondents as a substitute for people, and we think you should be wary if a research partner is proposing to use them.[11]

BBS's call on this one: as the cost of creating testing inputs comes down, the number of concepts reaching real people can increase. And the level of display realism (fidelity) can be dialed up or down more efficiently and deliberately to allow finer control of what is being reviewed and decided. The audience itself is the part that still has to be real.

Gen AI Can Create Environments, Heroes Still Require a Master

The range of work generative tools handle well keeps widening, and the categories where they still fail are specific enough to plan around. This is also what sets the ceiling on everything in our first section: a concept artifact can only approach final quality where final quality is achievable in the first place.

Environment and mood work is deliverable today. An opening location fly-over, a city at dusk, weather effects such as rain/mist/fog/snow, a textured backdrop, crowd ambience behind a scene. Generative output is already standing on merit in real work in these areas.

Hero product is where it still fails. An instrumented study compared generated motion against motion-captured real footage and found issues — generated things move in the correct way but at the wrong speed. The worst categories were impact, fast-moving fabric, and liquids or soft materials changing shape. We regularly see those motions in beverage, fabric and food. A pour. A garment moving on a body. A fork going into something. The model maker's own release notes disclose the same weakness, which validates our observed experiences. The study measured deviation from captured footage rather than whether a viewer notices, and there are certainly shots where nobody would notice. That gap has not been explored enough and we do not recommend exploring it with important hero shots just yet.[12] Our clients have noticed that packaging text and details are often unacceptable. And AI vocal pronunciation of product names can be inconsistent and wrong — a disaster when it’s your company’s product.

Somewhat perplexingly for something so seemingly precision-driven, brand color accuracy is a big problem. Generative models (23 were tested) managed a score of only 2.9 out of 5 for hitting a specified color. Video models (six tested) do better with color coordination. Scored on whether colors simply looked good together it ranked as 4.86 out of 5. Models appear to have been trained to produce color that people respond to, which often pushes everything brighter. For a brand with a style guide, which is all of them, that gap is a major stumbling block.[13]

We have reliably moved past some hurdles that were immediate show-stoppers. At least one model can render text correctly about 99 percent of the time, and the six-fingered hands are solved.[14]

We are seeing more use of a combination of Gen AI and traditional production — often the product rendered or photographed conventionally, the background generated. That is where we are seeing savings right now without a quality compromise, and it is the first place we would expect to expand in 2027.

There is also a consumer side to the quality question. Kantar's facial coding across hundreds of ads involving generative AI in its LINK database found that seamless use performed well, with more than 40 percent reaching the top tier for branded cut-through, while ads carrying distracting or unnatural visuals performed worse. Clearly, integrating Gen AI well is non-negotiable if you are going to use it. There is also a gap in receptivity for disclosed use of Gen AI: 41 percent of consumers say AI-generated ads bother them, against 29 percent of marketers.[15]

Looking forward, every model maker is working to push past current limits, so the boundary is constantly moving. The state of the art will need to be rechecked frequently. A category that was off the table in 2026 may be on it in 2027, and set policies will age poorly. It's worth asking for periodic boundary tests.

BBS's call on this one: environment and mood work can utilize Gen AI now. Hero product will likely need to stay in conventional production another cycle. Harder but innovative shots can be funded as a test, with a test's budget and a test's expectations. To avoid relying on a lucky drift, repeatability should be part of the testing.

Summary

For a brand team, the question underneath all of this is whether the process of picking the best creative option can be improved. We think it can. More directions can be developed far enough to be judged rather than described. The level of finish can be chosen to suit what a particular situation needs to see rather than being solely budget-constrained. More of those directions can be tested with actual audiences before the money is committed. The cautions are equally practical: flashy prototypes invite premature lock-in, more options on the table makes a committee mash-up easier to ask for, and the finished film that deviates from the version that tested well may invalidate the research.

For procurement, we see opportunity in contract construction to prepare for success. Is there an allocation for a director early enough to guide concepting rather than only to shoot? Is generative capability retained, or billed per project — because if it is per project, the economic gains may get lost in planning. Is level of photorealism/fidelity for concepting defined in the work description, or left to be negotiated mid-flight when the polished favorite tends to win by default? Do concept assets carry rights and formats that let them survive into production? And where is the review record held, and who owns the learning drawn from it?

None of this is settled, and we said at the outset that our crystal ball is no better than yours. What we are reasonably confident of is the direction. The cost of showing an idea before making it has fallen far enough to change which ideas get shown, and that is a change to the economics of advertising production enabled by our developing toolset.

Notes

1. Boundless for Mazda South Africa, via System1 — undated, vendor-published. Lester Platt, STLL Studios, LBB, 6 May 2026. Scott Kemper, executive producer at Droga5 and a BBS partner, in BBS's own article.

2. Kingsley Harden, The Jogger Studios, LBB, 29 April 2026. The detection test is Ipsos with the Syracuse University Newhouse School, May 2026 — twenty ads, 3,000 US consumers, AI versions used exactly as generated. Andy Lambert, Dark Energy Films, LBB, 17 December 2025. Will and Tim Truesdell, Stept Studios, on Oakley, LBB, 9 July 2025.

3. AICP National Live Action Guidelines, revised December 2019. BetterBriefs Project, October 2021, more than 1,700 respondents.

4. Pedr Howard, Head of Creative Excellence, Ipsos US, 9 July 2025.

5. Harden, as in note 2. Runway, 28 August 2026.

6. Ella Hafermalz and Marleen Huysman (VU Amsterdam) with Jana Retkowsky (Erasmus University), "Harnessing a 'Spirited Technology': How Working with Generative AI Collapses the Creative Process," Academy of Management Discoveries, 2026 — doi:10.5465/amd.2024.0280. A year inside the creative department at RTL Nederland, analyzing 207 AI-generated images. "Early aesthetic closure" is the researchers' own term.

7. Satya Nadella, Microsoft fourth-quarter fiscal 2026 earnings call, 30 July 2026. Zoom investor release, 30 October 2023.

8. Ziflow, 1 April 2026. Adobe, 20 April 2026.

9. Kantar LINK+, 4 March 2025. The 17 percent rise is Kristi Rogers of Kantar, in Digiday, 2 October 2026.

10. Magier, Quirk's, 1991, from a presentation to an ARF Copy Research Workshop. Diel and colleagues, Scientific Reports, 2026. Ipsos figures as in note 2.

11. Pew Research Center Data Labs, 30 September 2026.

12. GAUGE, arXiv 2608.05948, 6 August 2026, Shanghai AI Laboratory — preprint, not peer-reviewed. ByteDance's own release notes describe the same weakness.

13. The 2.9 figure is a benchmark of 23 image models asked to return a specified hex value. The 4.86 is WorldJen (moonmath.ai), March 2026, preprint — six video models scored across sixteen dimensions.

14. GPT Image 2, from Artificial Analysis's image arena rebuild, August 2026. On hands: Julie Seal, Republic of Imagination, in The Drum, 1 May 2026.

15. Kantar, 4 November 2025 — facial coding across its LINK database.

Search