Leaving every conference this year, but especially the Cannes Lions in June, I was struck by how totally AI excitement has proliferated across the verticals PurpleLab works with (Life Sciences, Advertising, Payer Provider, and more). It’s been a trending topic the last five years of course, but the earlier waves of AI, like image focused generative ai, appeared too early along for real creative applications, and tangential at best to the rest of business processes. This seems to have fundamentally changed this year- the generative AI Adobe is showcasing is vastly better than before, genuinely threatening the content creation side of advertising. At PurpleLab, my clients have more frequently come from the strategy, media planning, and omnichannel teams, and the major development has not come from a single app, but from a new approach to using AI: MCP. If you’re not familiar with it, you’re in good company, MCP is a methodology powering the much more well known new kid on the AI block, agents.
So what is it? MCP stands for Model Context Protocol, it’s a standard created by one of the leaders in the AI race (I have no dog in that match) that enables an LLM to interact with different databases and services. If that’s making you yawn, let me connect the dots. This is huge for agents, because the goal of an agent, doing things on your behalf independently, requires the LLM to interact with a lot of things external to it. Connect your LLM to a calendar, and it can schedule appoints for you. Connect it to your databases and IDE coding framework, and it can start writing code for you! Variations on this theme have existed for a while, for instance no-code platforms, but their range of function has always been far more constrained, because LLMs performance have advanced so significantly in the past couple of years.
I’m a historian by training, so for a brief pause- the current age of AI enthusiasm was preceded by a time known as “AI winter”, there have been two but the more recent and interesting one for my purposes happened the last time we all tried to make money off AI, in the late 80s. In that case you had a wave of what were called expert systems that encoded peoples knowledge and used that to make rules based decisions. This was a very promising field, and subsequently has gone on to produce the field of RPA, robotic process automation which I’m a great fan of, but was ultimately subject to a limitation. With the limitation of “the time and cost of paying experts”, it turned out not to be possible to encode every iteration of what might be asked of an expert system into the software. The volume of edge cases is simply too large.
A great example of this is autonomous driving- I could very easily use an expert system to encode a set of rules, say the state of Pennsylvania’s drivers ed manual, into software. The ability to actually drive down a street, knowing how fast to drive in different weather conditions, dealing with pedestrians and other drivers, creates infinite permutations I cannot write an instruction for each instance of. It’d take the next generation of AI, neural networks and RLHF plus oceans of recorded driving data to get to where we are today.
Even at that, the AI space is an intimidating space to get into. As of just a week ago a reported 95% of AI pilots drive revenue growth. Google and Meta and the rest are throwing billions into the race. So my first piece of advice- do not build your own model. They are spending more on it than you, and the bitter lesson of AI is that the “best” AI is a function of how much data and compute you’ve got, which is a checkbook race. Instead, build the applications around someone else's model that make it useful to a job function or vertical, and connect it with proprietary data that differentiates it’s abilities. That’s part of where PurpleLab comes in incidentally, as a cloud native in the Datavant token ecosystem we’re positioned as one of the easiest integrations to MCP enabled AI platforms.
Investing in technology for your enterprise is a stressful activity, and I have been involved in purchase decisions at all sorts of scales, from buying a couple licenses of a sales software (10-20k), to deciding how much to put into a conference (don’t ask me about Cannes), to multiyear data and software that support our advertising business (that make Cannes a rounding error). It’s always easiest to be wrong about the software licenses, if you’re wrong about the others they’ll haunt you- if ghosts could send invoices! So when I think about investing company dollars in new projects, I anchor my expectations to two things: time to value, and ROI. These amount to the so-what? of building products.
MCP and agents are interesting for that very reason- they amount to an orchestration layer between the user and the things the user needs to do in their jobs, embedding AI rather than having it flop freely without any particular application. What does this look like at PurpleLab? I could tell you, but I’d spoil a lot of what our product team is hard at work on. So instead I’ll show you by highlighting a few great businesses I’ve seen putting this into practice. Another reason I appreciate this category of company, I’ve long advocated for AI as a tool to augment, not replace the role of the agency, the media planner, the marketer. Much as expert systems were a significant advancement with a ceiling, the best versions of AI tech out there today make teams faster and stronger, not redundant, because the human edge remains.
https://www.akkio.com/ – I was deeply impressed with Akkio’s audience agent layer, which leverages MCP protocol to understand the attributes of a first party CRM or database to develop different audience strategies and ship them. https://axonal.ai/ – It takes one to know one, and the founder of Axonal Larry Mickelberg, a MM+M Top Healthcare Marketing Influencer, knows agency needs really well. When I think about value creation and ROI in this phase of the cycle, the value an agentic organization brings, it’s moat, is how fit to purpose the agent is for the client's needs. https://www.newtonresearch.ai/ – Perhaps it’s my ex-LiveRamp bias showing, but founder John Hoctor built a fantastic solution for TV measurement with Data+Math, and is now encoding his experience into agents to apply all kinds of measurement models, uncoupled from the human hours model that has traditionally dictated measurement’s speed.
There are no barriers to entry in the AI space right now, you can go sign up to Replit today and start “vibe-coding” to your hearts content. The right question when there is no barrier, is what is your moat to differentiate? At Purplelab we’ve got a world of ideas to share, the raw material to build with, and we understand founders visions. Drop us a line.
Originally published on LinkedIn.