In October, The Martech Weekly published its Enterprise Martech Outlook, which looked at 417 enterprise stacks and surveyed 127 brand-side leaders in the UK, the US, and Australia. Two numbers from it belong on the same slide.
95.3 per cent of them have AI agents on the roadmap. Yet of the 461 agent initiatives the researchers tracked, only 10.8 per cent had reached full production.
The gap between those two figures isn't a technology problem. The models work fine. The gap is everything the model needs to know before you can trust it to do anything, and nobody wants to budget for that part.
What the Failed Projects Had in Common
As Scott Brinker put it in CMSWire in May, "The big gotcha is that AI doesn't make martech complexity go away. It exposes it."
Real Story Group said something similar in a piece on agentic ROI a year ago: "ROI comes from the systems and structures feeding the models, not the models themselves." Without a disciplined content supply chain, they added, "GenAI just generates noise faster."
The Martech Weekly's review of Brinker's Martech for 2026 report shows where that noise comes from. 62.1 per cent of organisations weren't connecting their data warehouse to their agents at all. Only 20.4 per cent were feeding their agents transaction history, and only 22.3 per cent were feeding them support tickets. That means the agent that's supposed to help a customer knows nothing about what the customer actually bought or complained about. Five of the top nine implementation problems in that study were about organisational readiness, not software.
Brinker's word of the year for 2026 is context, meaning all the customer, company, and system knowledge an AI needs before its output means anything. Or as he puts it less politely: "Garbage in, garbage out hasn't been repealed. It's just that the garbage is now beautifully written."
The Final Boss
Juan Mendoza, who publishes The Martech Weekly, calls data fragmentation "martech's final boss". At one of his conferences this year, 45 per cent of the room named it their biggest problem, and his research shows that 58 per cent of enterprises have only a siloed or partial view of their customer. He describes the typical enterprise as a Jenga tower, where every team has added a block of customer data for its own purpose, and nobody is allowed to take one out.
His conclusion is both uncomfortable and right: "There's no incrementality, no POC, no experiment or sprint that can solve data fragmentation problems in large organisations." Instead, it takes an executive who is willing to fund the unglamorous work and, sometimes, to slow something else down while it happens.
The Pressure Is the Opportunity
Here's the useful part, though. Boards want AI, they want it quickly, and right now they aren't asking hard questions about where the money goes. As Mendoza sums up the current mood, there is "plenty of money for a hyped thing, and not enough for the important thing."
Tony Byrne describes marketing leaders as being caught in an "AI sandwich", with a board above expecting multiples, and a team below finding that the tools are inaccurate and hard to teach. The skills data backs him up. In a NewtonX survey of 500 marketers, 69 per cent had been strongly encouraged or ordered to use AI, yet 43 per cent had received no formal training, and 81 per cent said their company overstates its own AI proficiency. Meanwhile, Gartner found that only 15 per cent of chief executives believe their CMO has the AI knowledge the role now needs.
The practical move, which Mendoza makes explicit, is to put the foundation work under the AI heading. Customer data unification, a proper content model, metric definitions, consent, and governance have all been on the to-do list for a decade without ever getting funded. Now they're called "AI readiness", and AI readiness gets signed off.
The agent is the cheap part. What it needs to know is the expensive part, and that's been sitting on the backlog for ten years.
A Two-Speed Plan
Real Story Group recommends a two-speed plan, running near-term wins alongside the foundation work. That's the right shape, and here's how I'd draw the lanes.
The fast lane is for low-risk, high-value uses that pay for themselves inside a quarter and don't touch the customer. When Brinker surveyed marketers at software companies who build their own tools with AI, he found the work clustering around workflow automation (57 per cent), integrations (50 per cent), and data cleaning (48 per cent). Only 17 per cent were building customer-facing apps, which is the right ratio for now. Byrne's view is that while generative AI gives meaningful, mostly incremental gains, the real scale is in decisioning, and decisioning is where governance bites hardest. Brinker has a test he uses for a different question, whether to build or buy, that works just as well here: could a bad output become a screenshot? If it could, keep a person in the loop.
The slow lane is the foundation, which in RSG's framing has four layers: customer data and identity, structured content, decision rules and orchestration, and metadata, taxonomy, and governance. None of these produces a flashy demo for the board, but together they decide whether the demos ever become real products.
The trick is that the two lanes fund each other. The fast lane buys you credibility and a little time, while the slow lane builds what the next three years of AI spend will actually stand on.
Five Questions Before Signing Off an AI Programme
What will this agent read, and is it one source or six?
If the answer involves exporting a spreadsheet, the project is just a prototype, whatever the slide says.
What happens when it's wrong?
Who approves an action, what gets logged, can it be reversed, and is the spend capped? RSG's vendor evaluations now ask all four of those questions.
What does it cost in production, rather than in the pilot?
As RSG points out, "An inexpensive pilot can reveal almost nothing about production cost." Pricing now mixes seats, credits, tokens, and minimums, and The Martech Weekly reports that 85 per cent of companies can't forecast their AI costs to within 10 per cent.
Could we turn the vendor's AI off and put our own in?
Byrne suggests putting both parts of that question into every RFP. If a platform's AI can't be switched off, you've bought the vendor's judgement along with its software.
What foundation does this project leave behind?
You want it to leave a cleaner customer record, a better content model, or a shared set of definitions. If the answer is nothing, it was just a demo.
The Short Version
Nine agent projects in ten never reach production, and the cause is almost always data, content, and rules that were never fixed. The board's current appetite for AI is the first real budget those fixes have ever had. So run a fast lane of safe, internal uses to buy yourself time, and spend the rest on the foundation. By the time it's ready, the agents themselves will be cheap.