AWS's David Dorf On Why LLMs Have Peaked & What's Coming Next In 2026

Foreign.

Joining us once again for five insightful minutes is one of our favorite recurring guests, David Dorff.

David is the head of Retail Industry Solutions at aws and today he is going to regale us with his AI predictions for 2026.

And if we go a little long today, I will take full responsibility because you will want to hear what what David has to say.

David, welcome back to Omni Talk and let's get started with this.

You said in a recent article that 2025 was the year of LLM overhype.

To paraphrase, that's a bold statement given all the noise in the market, and also one that I happen to agree with for those that know me well.

So what makes you think we've hit a plateau and where do you see real gains happening from here on out?

So it was just a couple years ago people were talking about artificial general intelligence being right around the corner and I think we got a little bit over our skis on that.

LLMs are tremendously powerful, but the idea that we can continue to just scale up the data doesn't actually lead to better results.

So one of the things is that LLMs will incrementally improve, but there won't be exponential gains with the current technology.

There needs to be a new architecture like the Google Transformer that started the whole LLM thing.

Something like maybe what's going on with world models.

I'm not a scientist so I don't go that deep, but there's some cool stuff brewing out there that could help us in the long run.

But in the meantime I think it's better to focus on answers at lower costs.

And so we need to think about things like domain specific models.

So these are models that have a special purpose.

We've already seen some of these in like finance and healthcare and I'm waiting for some, some of these to come out for retail.

One of the things that we just announced at our RE Invent conference last week was this product called novaforge.

It's pretty unique.

We take a half trained frontier model and hand it over to the customer and then the customer can finish off the training with all their specific data.

So think about using that for something like a product catalog where you have a really tight LLM that really knows product catalog super super well, knows your brand, all that good stuff.

That's kind of where I think we're heading for the next thing.

David, we saw this flood of agent e commerce announcements at the end of 2025.

I'm curious what your take was on all the announcements and what needs to happen for these answer engines to really become a meaningful channel for retailers rather than just another place that demand shifts over to.

First of all, I think Flood is being kind.

It was more of a tsunami, actually.

Yeah.

Wow.

I was.

I was just trying to keep track of all the things that were going on.

There was so much great innovation.

It was really great.

But I would sum up things by saying I think it was fantastic progress, but let's put the cake back in the oven for a little longer.

Some of the usages were a little bit clumsy, but kudos to Walmart and Target for getting their implementations done with Peak.

We three have all been part of retail and we know November is a crushing month for them to get those things done.

I really know that was tough.

But I do think that agenic commerce is going to continue to make inroads.

But digital and physical commerce are certainly not going anywhere.

And yes, there will someday be an online retailer that has no website and just uses MCP servers to sell things, but that's an exception.

Most are going to really want to cater to both humans and agents.

And by the way, bots now outnumber the humans on the Internet.

As of last month, they crossed the line.

So just as it's really important for your products to be listed in search results, you'll also want to be listed in answer engines.

And the key question for us is, did the answer engine get her to buy when she would not have otherwise?

And today I would say no.

Answer engines are substantially similar to search engines, and that's because of a disjointed experience.

But if you look at Amazon, for example, they're seeing an uplift in sales from shoppers that use Rufus over shoppers that don't.

So I kind of feel like if you get the experience right, then it could deliver incremental sales.

Got it, David.

And that brings up.

Actually brings up the next question I was going to ask you because, you know, sometimes when I sit back and I think about everything that's been going on, I just wonder if we're just reimagining search first and foremost.

Which brings up a natural tension to me in my mind.

Like, there's always a tension in search between advertising and trust.

So how do you predict this reimagining of search, so to speak, will evolve?

And how should retailers think about this tension between profiting from but also figuring out how to participate with the answer engines?

Yeah, it's very expensive to run an answer engine, and most answer engines can't survive on consumer subscriptions alone.

So advertising is coming and sponsored prompts are the first shot that I've seen.

I know Amazon's doing that and Walmart's a fast follower in that regard.

Ads are coming and it's going to result in a loss of trust for some savvy consumers.

But we've been dealing that for a while and as you mentioned, Chris, this is an echo of se.

It's not just the ads.

Also, it's answer engine optimization.

Companies figuring out how to manipulate LLMs as well.

Right.

Just like they did with search engines.

So if you think about it, ChatGPT is notoriously known for loving top 10 lists.

So retailers have figured out, hey, if I publish a bunch of blogs with top 10 lists, I put my products at the top.

More chance that I'll be mentioned by an answer engine.

So there's already a lot of manipulation going on.

But there's.

But it's not new.

We've dealt with this with search engines and consumers will figure it out.

This is an ongoing battle that dates back to Mad Men, right?

Yeah.

So, David, how optimistic then are you about shopping agents actually delivering time savings and how do you envision them giving or gaining consumer trust, really, and consumer adoption, if ever?

That's where we diverge a little bit from answer engines to kind of standalone shopping agents.

And I definitely predict that third parties are going to start offering the shopping agent subscriptions that are going to automate some select shopping tasks.

And I think it's going to start with delivery companies like Instacart and DoorDash, because they already look across multiple sources for their products.

Right.

And they already have the delivery infrastructure.

So they might start with that answer engine integration like Instacart has done.

But I think they will eventually offer their own agents that are personalized to your tastes, and that's the big value add.

So they'll know your favorite restaurants, your favorite orders, the grocery items you typically buy.

They'll find the best prices and they'll use coupons.

That's where a lot of that time savings is going to come.

I also think there's going to be more focus on outcomes over products.

So you're going to say things like plan a birthday party with a pirate theme and shop for the items and it'll do that for you.

Or I want a dinner party for eight with a French bistro theme or something like that, and the agent will go off and grab all those items and have them delivered to you.

That's interesting.

I just heard Today actually that UBS is predicting that grocery will be the first industry hit by agentic commerce, particularly from a commerce perspective, which, you know, is kind of the flip side of how we saw E commerce develop overall.

So, so all right, so let's get you out here on this.

We've talked a lot about the consumer facing side of generative AI already, but what's your take on how agents will transform internal retail operations in 2026?

Yeah, here's where I think the crystal ball gets a little clearer because you know, consumers are pretty fickle, but companies are all about increasing efficiency and that's kind of the sweet spot for AI.

So retailers are definitely going to start deploying these agents that leverage reasoning to assist with things like merchandising, marketing and supply chain processes.

We'll start with a lot of human in the loop as we gain confidence, which makes a lot of sense.

And one example is last week at Re Invent we announced three New Frontier agents covering developers, security and cloud ops.

And that's kind of the first start.

I think we're going to start to see a lot of new agents and it's inevitable that they'll start with horizontal agents that work across lots of industries.

But eventually we'll get to industry specific agents that can really help people in places like merchandising.

And I wouldn't be surprised if eventually your SaaS subscription for some application comes with agents that help you run that application.

That that could be coming up soon in the future as well.

And then I also want to mention we just announced a partnership with Visa around agentic payments and I think, yeah, this works for B2B and B2C.

I think B2B is going to be the big impact because I can imagine the whole purchase order, inventory, receiving, invoice, match payment, all that stuff can be easily handled by an agent that's checking all the policies, making sure everything's been done right and then making a payment without having to do checks and things like that.

So I think, I think that's going to have a big impact on the back end as well.

Amazing stuff.

David, what are your plans for NRF before we let you really go?

Yeah, so I'll be in the booth and hanging around on the floor.

We're in booth 4438 this year and then also want to give a shout out to Super Saturday.

So the retail ROI folks have their special conference on the Saturday before NRF that's free for retailers to attend.

It's be great for people to go there.

As well.

I'm a big supporter.

Did a trip to Honduras with the gang and handed out Tom's shoes once.

And it's a it's a great time to really learn about retail tech, but also hear about some people that are less fortunate.

Such a great capstone to 2026 for us, too.

Really appreciate you coming on.