AI amplifies whatever system you point it at, including a broken one. Someone has to build the motion once, correctly. Then AI scales it. That order is not negotiable, and almost nobody selling AI to revenue teams mentions it.
This is the piece I'd most want a revenue leader to read before signing an agentic AI contract this year.
Same HubSpot session, August 2026. Talking about AI agents doing outbound, the presenter made a point about the AI SDR being only as good as the website it reads. If the information there is thin or wrong, the answers it gives won't be good.
Simple, almost throwaway. But it generalizes to everything.
An agent drafting outreach from job titles that were accurate two years ago is worse than a generic blast, because it sounds personal while being wrong. Automation built on duplicate records doesn't clean them, it propagates them, faster, to more people, with more confidence. A scoring model trained on a lead definition nobody agreed on will produce a confident answer to a question you never asked.
The model isn't the variable. The thing underneath it is.
This isn't a contrarian opinion anymore, it's the consensus of everyone measuring outcomes.
Gartner forecasts that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Gartner's own analysts have been more specific about the mechanism: an earlier prediction held that organizations would abandon 60% of AI projects that weren't supported by AI-ready data, and Gartner now recommends that the majority of an AI project's budget, at least 60%, go to data engineering rather than model selection.
McKinsey found that while nearly two-thirds of enterprises have experimented with AI agents, fewer than 10% have scaled them to measurable value, with eight in ten citing data limitations as the barrier. RAND researchers studying AI project failure put the rate at over 80%, more than double that of non-AI technology projects, and found the models were rarely at fault.
Specific to our world: Gartner expects 40% of agentic AI CRM projects to fail or stall by 2028, and in almost every case the cause is data quality rather than the AI technology.
Read those together and a pattern shows up. The projects that die don't die of insufficient intelligence. They die of insufficient foundation.
Here's the part that I don't think is priced in yet.
Bad data used to produce obviously bad output. A mail merge with a broken field said "Hi {FIRST_NAME}" and everyone knew something was wrong. The error announced itself.
AI removes the tell. It writes fluent, plausible, personalized-sounding messages from wrong premises. The output looks correct, so nobody checks, so the error scales silently across thousands of contacts. Meanwhile the underlying churn continues: research on CRM decay in 2026 found that roughly 70% of business contacts change roles, companies or responsibilities within twelve months, and nearly 43% of phone numbers become invalid within a year.
So you're pointing an amplifier at a dataset that degrades by about a fifth annually, and the amplifier is good enough to hide the degradation. That's the actual 2026 risk, and it's a process risk, not a technology one.
The other half of this is discovery, and it's the same problem wearing different clothes.
Buyers increasingly resolve questions without visiting anyone's site. US zero-click rates reached 58.5% of searches in 2025, and around 83% on queries where an AI Overview appears. On the B2B side, 6sense's buyer research found 94% of B2B buyers used generative AI tools during their purchase process.
Which changes what a website is for. It's no longer only a page someone reads. It's a source something quotes. And what gets quoted is what's clearly written and attributable. Bain's analysis found that 89% of citations for unbranded B2B questions come from third-party sources rather than the brand's own website, so this isn't a matter of optimizing your homepage. It's a matter of being described accurately and consistently wherever you appear.
The traffic that does arrive this way behaves differently. Semrush found AI search visitors converting at 4.4 times the rate of traditional organic visitors, and when a model surfaces a vendor a buyer hadn't heard of, 51% go straight to that vendor's site. Fewer visits, much higher intent.
And the part I find genuinely useful: the clean, clearly written site that a buyer's AI can quote is the same site your own agents read when they go do outreach. One piece of work, two payoffs. That's rare enough to be worth prioritizing.
If I compressed everything above into a sequence, it would be this.
Define the motion in writing first. What counts as a qualified lead here, who owns each stage, what happens on each outcome. If a human can't state it, an agent can't execute it.
Fix the data the agent will read. Not all of it. The specific fields the agent depends on. Titles, ownership, stage, source. Gartner's 60%-to-data-engineering guidance exists for a reason.
Write your own information plainly, in one place. For the human buyer, for their AI, and for your agents. Plain sentences, no superlatives, no claims that can't be quoted.
Then automate, narrowly, with a baseline. One motion, one measurable before-number, a fixed window to reach production. Projects without a defined success condition are the ones that become cancellation statistics.
Keep a human on the decisions. Agents are good at execution and bad at judgment. That division of labour is the whole design.
I don't think anyone has this fully figured out yet, us included. But the order is the part I'm confident about, and it's cheap to get right compared to the cost of getting it wrong at scale.
Why do agentic AI projects fail?
Rarely because of the model. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027 due to cost, unclear value and weak risk controls, and puts data quality at the centre of CRM-specific failures. McKinsey found eight in ten enterprises name data limitations as the barrier to scaling.
Can AI fix bad CRM data?
It can help clean it, but it won't fix a process that produces bad data. Pointed at a decaying database, AI produces fluent, confident, wrong output at scale, which is harder to catch than obviously broken output. Fix the fields the agent depends on first.
Do I need to optimize my website for AI search?
If B2B buyers matter to you, yes, though not the way schema vendors suggest. Ahrefs found no significant citation lift from adding JSON-LD across 1,885 pages. What works is clear, quotable, accurate writing, plus consistent third-party descriptions of your company.
What should come first, AI tooling or process work?
Process. Define the motion, assign ownership, clean the specific data the agent reads, then automate one narrow use case with a measurable baseline. Gartner recommends at least 60% of an AI project's budget go to data engineering rather than model selection.