- il y a 21 heures
The Future of AI in Enterprise
Catégorie
🤖
TechnologieTranscription
00:03All right, thanks for coming.
00:05I'm Scott Likens with PwC, and I'm the chief AI engineer.
00:09And in that role, I have to not only do research and development around AI, which is fairly exhausting,
00:15but I have to think about what it means for the enterprise.
00:18So I'm really happy to be joined by Martin today of Cohere.
00:21And I think that is your focus, AI for the enterprise.
00:24And I like to joke is, what's taking so long?
00:27I mean, AI has been around forever, but tell us about Cohere, and what's your mission for the enterprise?
00:33Yeah, well, first of all, thanks for that loud cheer as I walked in, although I think maybe it was
00:38for you.
00:40So great to be here.
00:41I'm Martin Kahn, the president and COO of Cohere.
00:44So we are the leading secure AI for business company, and we take that very seriously.
00:52We do not and will never have a consumer chatbot.
00:55We only serve enterprise and, of course, government ministries and organizations like that.
01:02It's completely different to serve consumer markets and business markets, as I think we all know.
01:08There are very few companies of any size, let alone a startup or a scale-up, who can serve consumer
01:13enterprise equally well.
01:14And so what that means is there are really sort of four main areas of differentiation.
01:19Number one is we are independent.
01:22We are not controlled by any big U.S. tech company.
01:25We are Canadian in our roots, but we have a global company.
01:29But we deploy in any data environment.
01:32We bring our technology to your data, whether that's Azure, AWS, OCI, GCP.
01:37We have great relationships with all of them.
01:40But 90% of what we do is private deployment.
01:43So either in a VPC, on-prem, all the way to air-gapped on-prem.
01:47So we have been living, breathing data security, privacy, sovereignty since day one.
01:53We also have developed applications, not just the raw models.
01:56We create our own tech from scratch.
01:59There are six other companies in the world that do that.
02:01But all of them are controlled, the other five, by U.S. big tech.
02:04And all of them are primarily focused on consumer.
02:06We're just focused on enterprise.
02:08But we now have applications like Compass, an end-to-end search application, or North, which is an agentic AI
02:13platform,
02:14that brings together all the components that you need so you don't have to build them all yourselves.
02:19Then we help customize and configure so that each of our enterprise, government, partners, customers have something that is unique
02:26to them,
02:27gives them alpha, gives them differentiation.
02:29It's not a generic capability that 500 million people use.
02:33And then lastly, we've been focusing since day one on multilingual, the real focus on multilingual.
02:38As an enterprise, you can't choose to speak Brazilian Portuguese if you have a subsidiary in Brazil.
02:43And so our generative capabilities work across the top 23 business languages, and our search works across 109.
02:51And so those are four things that we say that enterprises really care about and that we focus on as
02:56a company.
02:56And that's a really different approach.
02:57Which, as you said, a lot of it's come out of the consumer technology trying to then apply into an
03:02enterprise.
03:03What's been the challenge for you to build this and then thinking about the barriers for the enterprise,
03:08which I'm well aware of coming with the big tech consumer-based technology.
03:13But what's been the journey to create this?
03:15It sounds like the full stack, the full solution, and solving some of the problems enterprise needs.
03:19But walk us through that journey and some of the barriers that you're seeing in the enterprise.
03:24Yeah.
03:25Well, we work very closely with companies like yours and with you specifically who are giving advice to companies.
03:31Okay, what do we now do?
03:33Someone I'm actually speaking with a little bit later today said on a panel we were on,
03:37if you're still doing strategy, you're already too late, meaning start building stuff.
03:43That's why I think it's great to be with you, an engineer.
03:45I'm a reformed strategy consultant, unfortunately.
03:49The barriers are just getting on and building things in production at scale, not proofs of concept.
03:57Kill that word.
03:59Kill that word.
04:01There's no concept left to prove.
04:03If you're still proving concepts, you're going to be too late.
04:05I liken this to 30 years ago, unfortunately, I'm old enough to have been working then,
04:11when the internet came along in the mid-90s and everyone would put up an HTML page
04:16that was designed by some company with a picture of the CEO and her phone number and we're digital.
04:22Well, no, you're not.
04:23We're online.
04:24You're not digital.
04:24You're not online.
04:25It changed everything.
04:27It changed everything internally.
04:28It changed how you serve your customers.
04:30It changed your supply chain.
04:31And I think the same is true now.
04:33And so the key we say for enterprises is get something into production at scale.
04:38And so with us, we deploy our tech in your data environment, connected to your data sources,
04:43not using some chat bot, some API, playing around with a demo, and really having the conviction.
04:52And it always comes from the very top of the enterprise to say, we're moving into production.
04:57Great example is RBC.
05:00They're public about this.
05:01We have a very tight partnership with RBC.
05:02They have a phenomenal team.
05:04They were building a lot of stuff internally.
05:06They have 14 risk departments, 14, and the federal regulator.
05:11So they had 192 ideas they wanted to deploy.
05:14The issue wasn't, do we have the cool tech?
05:17Do we have enough ideas?
05:18Do we have enough strategy consultants or whatever?
05:20The issue was one of those 14 risk departments would say no, because it wasn't compliant with this.
05:26There's an issue with data leakage, with hallucinations, whatever the case is.
05:29So we're partnering closely to deploy North for Banking in their data environment, hooked up to their data sources.
05:36We don't even see what they're doing with it.
05:38So that they can then go and focus on the applications, the solutions that will create real value for all
05:4491,000 of their employees to serve their customers better.
05:48So it sounds like you've kind of attacked the technology platform, and now it becomes this business challenge, right?
05:54How do I organize it?
05:55And you mentioned the Internet.
05:57If you heard the last panel, Mohamed, our global chair, mentioned Mosaic, but I actually worked there.
06:02So I remember the days, too.
06:04So we're same vintage.
06:06It seems to be moving a lot faster, though.
06:08Back in those days, you get your website up, and that was it for the next two years.
06:12Is the pace part of the challenge with the enterprise?
06:14You mentioned 147 ideas.
06:17Those are all great concepts.
06:18How do we get them all in production at once?
06:20How do we make those 14 risk departments talk?
06:23Do you think the pace is causing some of the challenge?
06:25Or, you know, what are those other barriers in the business?
06:27Are they just stuck in paralysis?
06:30It's a great question.
06:31I think part of it is a bit of paralysis.
06:34I've heard someone say to me, POC prison, you know, proof of concept prison, or death by 1,000 POCs,
06:43you know,
06:43death by 1,000 proofs of concept, doing something to impress the board or leadership rather than getting into production.
06:50And so I think there's been a lot of focus on how do we find the right use cases?
06:55How do we stack rank them?
06:57How do we prioritize them?
06:58And it's lots of months and months and months of analysis.
07:01What we've been saying, the real leaders are just pick something.
07:04Who cares what it is?
07:05Employee onboarding.
07:06But deploy that in your environment.
07:09Figure out how much compute you need.
07:11Figure out how to hook up to the data sources, your control systems, your data access privileges,
07:17you know, all of the different yucky things that no one writes articles about.
07:21Once you've done that once, twice, it makes it so much easier to do the next 5, 10, 50, 100
07:27capabilities.
07:28And that's where we developed North.
07:30So once you deploy North, which brings together our generative models, embeddings, re-rankers, chunking, parsing, indexing, vector database,
07:39all the connectors to your data sources, GUI, all of that stuff is turnkey.
07:44You don't have to worry about it.
07:45Once it's deployed inside your environment, every single person in your company can create their own agents like this
07:51without any tech capabilities, complete no code, because all of the hard work has already been done up front.
07:58And so if you pick something, maybe not your most risky thing, that forces you to do all the yucky
08:04work to hook it up with your engineers,
08:07with your CTO, with your CISO, with your general counsel, et cetera, et cetera, et cetera,
08:11then you've got something that gives you huge advantage over your competitors.
08:14Because then when you have a great idea, not on a PowerPoint deck, but someone in their working life says,
08:21why don't we have an agent that does X, you can very rapidly deploy that because all that heavy lifting
08:25has been done.
08:26And you'll see there are some companies who are just being more ambitious at doing that.
08:31A great example, we recently announced a very deep partnership with Dell.
08:36Our CEO and co-founder Aiden was on stage with Jeff and Arthur to talk about this.
08:41They are customer zero. They're deploying Cohere internally inside Dell.
08:46But then they're going to take this and ship it with their devices for on-prem deployments,
08:51where it's already pre-configured for their own customers.
08:54But Michael Dell has basically said, our entire company is going to become an AI company in how they do
09:00business.
09:01And it's coming from the top. And John Rose, the CTO, there's some incredible stuff that they've rolled out
09:07because he's saying, we are going to be AI first.
09:10And I think to your point about things speeding up, when competitors see their competitors or companies see their competitors
09:17moving more quickly,
09:19then they realize, my goodness, you know, are we going to be like, I don't know, Olivetti or Nortel or
09:25whoever does it back in the Internet days
09:27that literally do not exist anymore, you know, blockbuster, if we don't hurry up and catch up ourselves.
09:32And so I think companies like RBC, we have a deep partnership with STC and Saudi.
09:38It's the biggest telco in the Gulf. They're deploying north for telco and also customizing capabilities,
09:44our capabilities to be very performant in 12 different Arabic dialects.
09:49That's important to them. If they do that and their competitor doesn't, they're going to have alpha.
09:54They're going to do better. And so I think that's what is speeding up.
09:59We've really focused on working for the leaders in various regions, Fujitsu in Japan, LG in Korea,
10:06or sectors like RBC in banking, like Dell, obviously, in technology, like STC and telco,
10:12where they want to lead and be on the vanguard.
10:15I think now others are saying, my goodness, we better hurry up or we're going to be left behind.
10:19Yeah, so speed matters. I agree. And I love that concept of client zero.
10:23I like to say drinking our own champagne.
10:25It's probably a unique time in technology where you're building something for yourself,
10:29but also for your customers. Normally, you're building for your customers.
10:33Can you talk a little bit about where you've mentioned a few industries like banking and tech,
10:37but some of the industries, they're seeing an advantage already.
10:41I think last year was your point. It was experimentation. It was POCs everywhere.
10:46It feels like the shift is happening to enablement to get into the enterprise,
10:50but can you talk about some of the sectors you're working with?
10:52I think it's really unique with your approach.
10:54You've solved that security and data and privacy and the whole problem.
10:59So maybe some examples of other industries.
11:03I'll be cheeky, and I'll say every sector.
11:06I mean, if you think about the Internet, it changed everything, absolutely everything.
11:13Now, I will answer your question, though, properly, because where we focus as a company
11:17is on sectors where there's a particular focus on data privacy, data security, differentiation,
11:24the ability to create alpha, the ability to do something that gives you an advantage of your competitors.
11:30And those are generally critical infrastructure sectors, often highly regulated sectors,
11:36banking, insurance, asset management, health care, energy and utilities, manufacturing and supply chain,
11:43telco, as I mentioned.
11:44And so you can see a lot of the companies we work closely with are in those areas.
11:49And another great example, I can say this now publicly because it was in the press this morning,
11:56we're working with a company called Ensemble Health Partners.
11:59They're owned by private equity, but they're a multibillion-dollar revenue cycle management company
12:04in the U.S., connecting payers, insurers, hospitals, providers, et cetera.
12:09And it's in the article, so you can read it, their CEO.
12:13The work that we're doing is deploying our North platform, not to augment their employees,
12:19but to automate entire processes.
12:21And he has said, and we've signed up and committed to this,
12:25we're going to take 40% of entire process and automate, 40%, four is zero.
12:33And so if you think about the impact, if they can do that and either offer better pricing
12:38or spend more time avoiding issues or innovating or returning money to shareholders
12:44or whatever it is that's important to their company, that's enormous if they can do that
12:49and their competitors can't.
12:52So I think you could apply that to anything.
12:54It's basically revenue cycle management, automating entire process, very similar for insurance,
12:59very similar for banking, very similar for manufacturing, supply chain, transport, you name it.
13:07But very importantly, this is healthcare.
13:10They have healthcare data.
13:11They can't send that out through an API to, you know, some public chatbot
13:16or to a company that is, you know, exclusive to one of the cloud providers.
13:21And so that's where the security and privacy, the applications where North enables them
13:27to very quickly deploy and customize and configure for them.
13:31And in that example, the first results are going to be seen in six months, within six months.
13:36It can be very fast because we've done a lot of the heavy lifting to start, you know,
13:41I say start the marathon a mile 25.
13:44We're in Paris, so I'll say kilometer 41.
13:47I'm also Canadian, but we use miles for marathons there too.
13:51And that is a real advantage for them.
13:54I think we talked about Royal Bank of Canada, RBC.
13:57They have masses of processes that they can serve their high net worth individuals better
14:02by giving them better information because they can access masses of internal and external data
14:08to give better advice or make better loan decisions or do better analysis about equity investments.
14:15You know, the list is kind of endless.
14:16But I think those critical infrastructure sectors where there may be regulators,
14:21data privacy, data sovereignty is very important.
14:24and customization configuration for your language, for your dialect, for your values, very important.
14:32The values that, you know, a French company or a German company or a Canadian company may have
14:36and want to embolden may not be the same as where the technology that is publicly available is from.
14:44You may want to make sure that it's going to give the right kind of guidance, advice,
14:49look at the right sources, whatever the case may be.
14:52So in a lot of the conversations I have, I actually walk away thinking AI is not actually the hard
14:57part.
14:58You're hitting is here.
14:59Do you think enterprises are prepared with their data?
15:01You mentioned security and privacy, you know, understanding where their data is at.
15:06Can they access it?
15:07Are they allowed to legally use it?
15:09Where do you think the enterprises are in that journey?
15:11We've been in the big data era for probably 20 years
15:14and it almost feels like we're starting over to some degree for what AI needs.
15:17But any thoughts around what enterprises should be thinking about around the data to power this AI
15:22and get that 40% process automation?
15:27Yeah, I mean, data is very important and data privacy and security is very important.
15:33Maybe not for all enterprises, but certainly the ones that I've been talking about in those sectors,
15:37certainly in regulated sectors.
15:38I think we're seeing now also geopolitically there's a lot of focus on, well, you know,
15:43where is our data going and, you know, are we sure that, you know,
15:46we're going to be able to access when we need it if something goes wrong or whatever.
15:50Data sovereignty.
15:52You know, we've thought about data sovereignty not just in terms of nationally,
15:55but actually for the companies themselves.
15:57But certainly it fits perfectly with what many companies
16:01and many governments are pushing for on data sovereignty these days.
16:07So our philosophy is pretty simple.
16:10Like I kind of mentioned is that we ship our tech to your data.
16:14We do not ask you to ship your data to our tech.
16:17And many companies, maybe most companies, are actually multi-cloud or even hybrid cloud,
16:23meaning they have some stuff on-prem, they have some stuff with one provider,
16:26some stuff with another provider.
16:28Maybe they have, you know, their SaaS providers.
16:30We power over 100 features inside Oracle's SaaS capabilities that will be deployed on OCI.
16:37Maybe they use SAP.
16:39Maybe they use Salesforce.
16:42And they want to be able to access that data to make better decisions.
16:49And so quite simply, we just want to make our tech available wherever their data is.
16:53So you don't have to think about, you know, do we shift to GCP if we want to use Gemini?
16:58If their MLEs, their machine learning engineers, decide that Model X is a really good one
17:04because they played with it, but it's only available on a different cloud.
17:07And we want to change it next week when the new one comes out.
17:09And then next week they want to adopt something else.
17:11And, oh, we just signed a, you know, $400 million cloud contract.
17:15We don't think that's what enterprises really want.
17:18They don't want to be tied in.
17:19So it's really about flexibility.
17:21I think the other thing important for us, because technology is changing very rapidly,
17:26you know, we have a very fantastic relationship with NVIDIA.
17:30They're investors.
17:31But so do we with AMD, who are also investors.
17:34So Cohere is available not just on all different clouds and on-prem,
17:38but we work for serving these models on NVIDIA, on AMD, on Cerebris, on Grok with a Q.
17:46The Q, yeah.
17:47And if a customer says, hey, we really want to be available on, you know, something else, you know, we
17:54will look at that as well.
17:55Because we don't want this to be, you know, a siloed game.
18:01We think that our customers and partners want to deploy wherever their data is.
18:06I couldn't agree more.
18:07That's my architectural philosophy is flexibility over longevity.
18:11Things are just changing much faster than we've seen before.
18:14So in the final minute or so, you know, you only said the A word agent once, which I love
18:19because I'm sick of hearing the word agent.
18:21But where do you see things going?
18:23And I'm not going to ask five years, but in the next 12 months, maybe from your own perspective at
18:27Cohere,
18:28but like in the industry for enterprise, what do you think the big things we have to solve for in
18:32the next 12 months?
18:34Deploying in production at scale.
18:36Get it out there.
18:37We're just boring.
18:38We're a boring, boring company that just wants to get stuff out there.
18:42We don't believe in hyperbole.
18:44Our co-founders don't talk about crazy things happening in the future.
18:50It's how do we help our partners and customers get out there, deploy at scale in production?
18:57And I think I am a sort of partially reformed consultant myself.
19:00So I joke that my old colleagues, you know, were metaverse experts and NFT experts.
19:06Then don't say the M word, you know, then cyber experts, AI experts.
19:11Now everyone's an agent expert.
19:13Agents are really exciting, but you can't just jump all the way there.
19:17So I'll leave you with something.
19:19We're very fortunate to have incredible technical leaders that are literally inventing this technology,
19:24starting with the Transformer, our co-founder and CEO, Aidan Gomez,
19:28is one of the inventors of the Transformer.
19:30We have a chap called Patrick Lewis, who leads our RAG,
19:33Retrieval Augmented Generation Capabilities.
19:35He's the lead author on that paper.
19:36Yeah, it's pretty good.
19:37He coined the phrase RAG.
19:38And what he said, and of course, what our chief scientists, Phil and Aidan and Nick and Ivan,
19:42our co-founders believe, is everything starts with incredible search and retrieval capabilities.
19:47If you do not have great search and retrieval capabilities, everything else is a waste, garbage in, garbage out.
19:53And so we focused as much, if not more, on search and retrieval.
19:56You need to have an incredible RAG foundation to pull information that the models were not trained on.
20:03You probably would never want them to be trained on sensitive, up-to-date information, whatever.
20:08Then secondly, when you do that, you can make them really good at tool calling, tool use.
20:15LLMs are not great at math.
20:16So don't try and make them great at math.
20:18Just give them a calculator.
20:20Give them a Gmail, you know, an email program.
20:23Give them access to write, to Salesforce, et cetera.
20:26And then when you have the, you can have agents who tie a lot of those things together and do
20:31things in a more automated way to actually, as I mentioned before, in the case of Ensemble, automate entire processes.
20:37But unless your capabilities, your tech, is really good at tool use and tool calling and really good at RAG
20:44to start, you can't just use the word agent and know what you're doing.
20:47And we've been developing that for the last, well, I guess, five years as a company, certainly the last three
20:53years to go down that path rather than just find, replace, because it's now the, you know, the word of
21:00the day.
21:00You need to build those foundational capabilities and then customize them and configure them for the specific context of your
21:07enterprise, not the generic context that 500 million people use and consumer,
21:12but something that is designed for your enterprise context, helps you find needles in haystacks, helps you make better decisions
21:20in your competition.
21:21Well, it was a pleasure talking to you.
21:23I think you have a unique approach for the enterprise.
21:25It was very compelling to me, and I agree with all of your architectural strategies, even as a former strategy
21:30consultant.
21:31So, Martin, thank you very much, and I appreciate your time today.
21:34Thank you.
Commentaires