Are Agencies Eating the Software Companies?

So here's something nobody in the ecommerce SaaS industry wants to hear: the biggest threat to their business isn't another SaaS company. It's their own customers.

Pattern — an ecommerce agency, not a software company — just walked onto a stage in Salt Lake City and unveiled an AI execution engine called Pi that makes a meaningful chunk of the existing ecommerce tool stack look like it was built for a world that no longer exists. And they're not alone. The holding groups are doing the same thing, only louder and with more zeros on the investment slide.

We should probably talk about what this means for anyone who sells things online.

When Your Agency Starts Acting Like Your Software

Pattern manages brands across more than 70 global marketplaces. They are, by any traditional definition, a service business. But Pi — Pattern Intelligence — behaves like a platform. It runs on what the company claims are 77 trillion proprietary data points (with another 800 billion added weekly, because apparently trillions just aren't what they used to be). The system monitors advertising, pricing, content, inventory, and featured offers continuously. When it spots a problem — a lost buy box, a pricing anomaly, a content issue — it can fix it autonomously. No human logging in. No dashboard to squint at. No ticket to file.

Here's the thing: the distinction that matters here isn't between "agency" and "software company." It's between systems that tell you about problems and systems that actually fix them. Pi detects a buy box loss, diagnoses the cause, executes the recovery, timestamps the action, and gives the brand a searchable audit trail. Your current tool stack probably sends you a Slack alert and wishes you luck.

And Pattern built this because they had to. Their chief revenue officer framed it during the keynote as an "executive dilemma" that anyone selling on marketplaces will recognise immediately: 70% of North American brand leaders have fewer than ten people dedicated to ecommerce. Half have teams of fewer than five. Those teams are simultaneously managing advertising across multiple marketplaces, monitoring pricing, optimising content for both humans and AI discovery surfaces, coordinating inventory, and now — because the universe has a sense of humour — figuring out how to show up in Alexa for Shopping conversations.

Five people. Doing all of that. The traditional answer was "buy more SaaS tools." The emerging answer is "build a system that does the work itself."

The Holding Groups Aren't Waiting Either

Pattern would be an interesting case study on its own. But the reason this matters beyond their client base is that the same pattern (no pun intended, but also not not intended) is playing out across the entire advertising and commerce ecosystem.

WPP launched its Agent Hub in January — an internal marketplace of AI agents built on nearly 30 years of proprietary data, including brand valuation research, behavioural science frameworks from Ogilvy, and cross-industry creative intelligence. They've since layered on WPP Open Pro, an agentic platform that lets marketers plan, create, and activate campaigns without formally engaging an agency. The reported investment: $400 million. In AI infrastructure. From an advertising holding group.

Publicis Groupe is spending €300 million over three years on something called CoreAI, built on an internal AI platform they've been developing since 2018 — which, for context, was before most of us had heard the word "agentic" used in a sentence without air quotes. They've integrated identity data from Epsilon and digital transformation capabilities from Publicis Sapient, and in April they expanded a partnership with Microsoft to build a full-stack agentic marketing platform.

Omnicom, not to be outdone, unveiled its next-generation Omni platform powered by 2.6 billion verified global identity records. But the genuinely wild part? They've already executed real client media buys using agent-to-agent AI. Not a demo. Not a proof of concept. Actual money flowing through an agentic pipeline. They've also developed something called AdCP — Ad Context Protocol — an open standard where AI agents autonomously discover inventory, purchase media, build creative assets, and activate audiences. (If you're a media buyer reading this, now might be a good time to sit down.)

Dentsu launched dentsu.Connect 4.0 in April, positioning it as an agentic operating system. Their China operation partnered with iFLYTEK to integrate agentic capabilities into generative engine optimisation specifically — which means they're already building for the world where AI agents mediate product discovery, not just search ads.

The common thread across all of these: the organisations with the deepest proprietary data, the longest operational histories, and the most direct execution responsibilities are building internal AI ecosystems that absorb the functions previously outsourced to vertical SaaS tools. They're not buying software to manage campaigns. They're building systems that manage campaigns, create content, buy media, and optimise pricing autonomously, using their own data as the moat.

The Uncomfortable Question for SaaS

Here's where the economics get properly uncomfortable for ecommerce software companies.

The traditional SaaS value proposition rested on two pillars: integration complexity (connecting to marketplace APIs is hard) and aggregated intelligence (managing campaigns for thousands of brands produces cross-category insights no individual brand has). Both pillars are eroding simultaneously.

Model Context Protocol and similar standards are collapsing integration complexity toward zero. When any AI agent can talk to Amazon's advertising system natively through a standardised protocol, the software layer that existed to mediate that connection loses its foundational justification. The "we connect to the APIs so you don't have to" pitch doesn't land the same way when connecting to APIs is becoming trivially easy.

Meanwhile, the aggregated intelligence advantage is being outgunned. WPP has three decades of brand research. Publicis has eight years of identity data infrastructure through Epsilon. Omnicom has 2.6 billion identity records. Pattern claims 77 trillion data points from 13 years of marketplace operations. Before AI, having that much data was mostly an infrastructure headache — impressive on a conference slide, practically limited by human capacity to analyse it. Now an autonomous system can continuously monitor it, identify patterns across millions of signals, and execute responses in real time. The data moat only becomes a functional advantage when paired with the execution capability to act on it. Which is precisely what these organisations are building.

So what's left for the vertical SaaS tools? Governance. Approval workflows. Audit trails. Multi-user permissions. Budget controls. Enterprise brands are not going to let autonomous systems make unchecked decisions on $50 million advertising accounts (at least, not yet — give it eighteen months). But governance alone is a thin foundation for a product category that has commanded premium pricing on the basis of intelligence and optimisation.

The Conversational Commerce Wrinkle

All of the above assumes the current model of commerce — keyword search, ad auctions, product grids — remains the primary surface. It won't.

Amazon's replacement of Rufus with Alexa for Shopping wasn't a rebrand. It was a structural change to how the platform's dominant customer base discovers and purchases products. In a traditional search interface, optimisation is a gradient — bid up, bid down, adjust placement, test keywords. In a conversational interface, the outcome is increasingly binary: the AI recommends your product, or it doesn't. There is no "page two" in a conversation.

That binary dynamic changes everything about what "optimisation" means. The signals that matter in conversational commerce — review sentiment, catalogue completeness, pricing consistency, fulfilment reliability, structured data quality — are not the same signals most vertical SaaS tools were built to measure or improve. Their entire architecture is oriented around the keyword auction model. Pattern's Pi already includes GEO and Alexa for Shopping scorecards showing how products rank across AI shopping surfaces. Dentsu is building agentic GEO capabilities in China. The holding groups are building for the commerce environment that's emerging. Most of the dedicated tools are still optimising for the one that's receding.

What This Actually Means If You Sell Things Online

The implication isn't that every brand should fire their tool stack tomorrow and call a holding group. (The holding groups would love that, obviously, but reality is more nuanced.) Here's what's actually worth paying attention to:

The gap between execution partners and reporting tools is becoming strategically important. A system that alerts you to a buy box loss is useful. A system that detects it, diagnoses it, fixes it, and logs the whole thing for your review occupies a fundamentally different position in the value chain. When evaluating technology partners, the question isn't "how pretty is the dashboard?" It's "does this system do work, or does it create work?"

The data advantage is real, but so are the incentive misalignments. Agencies and holding groups have deep proprietary data, but they also have their own commercial incentives. Pattern's approach — letting brands upload their own rules, tone, and preferences into a knowledge management system — suggests they're at least aware of the tension. Any brand evaluating an AI execution partner should be asking whether the system's autonomous actions are optimised for the brand's interests or the platform's.

And the GEO readiness gap is widening fast. Brands not yet measuring how their products perform in AI shopping environments — across Alexa for Shopping, Google AI Mode, ChatGPT shopping — are building strategy on an incomplete picture. The number of tools and partners that can provide this visibility today is small. It won't stay small for long, but right now it's a genuine differentiator.

The Bottom Line

The companies defining the next phase of ecommerce technology aren't necessarily the ones building better dashboards. They're the ones turning operational experience and proprietary data into systems that act — continuously, autonomously, and with an audit trail. Pattern's Pi is one version of that thesis. The holding groups are building their own, with substantially larger cheques.

For the SaaS tools that have served ecommerce brands for the past decade, the question isn't existential (yet). Enterprise governance needs aren't going anywhere. But the value proposition is being renegotiated, and the negotiating position of the incumbents is weaker than it was twelve months ago.

The agencies were supposed to be the customers. Turns out, they've been taking notes.

• Running a query fan-out once gives you a sample, not the full topical space. Run the same prompt repeatedly until no new sub-queries appear, cluster the results, and map empty clusters to content gaps.

• Benedict Evans released his biannual macro trends deck, 'AI Eats the World.' No content preview, but his past decks have landed at Alphabet, Amazon, L'Oréal, LVMH, and Nasdaq.

• TikTok data shows 64% of users buy after discovering on platform. Discovery and purchase now collapse into the same moment. Brands still measuring via last-click ROAS are systematically undervaluing it.

• Google launched Universal Cart, aggregating items from Search, Gemini, YouTube and Gmail, plus AP2, a protocol letting AI agents complete purchases autonomously within user-set spending limits.

• Google Search can now call local businesses on your behalf for select bookings. Combined with information agents and custom mini-app dashboards, it's shifting from search engine to operating system.

🎨 Soona - AI-powered creative tools for e-commerce and UGC.

⌨️ Quicky AI - Use AI on any website instantly just by selecting text

⚙️ Gumloop - Automate any workflow with AI.

🌐 Browser Use - We enable AI to control your browser

🤖 Doclime - Get answers from your documents

😍It was great, all I needed

🙂Good, not great

🙄It sucked