Intentsify this month published a white paper arguing that the central failure mode of AI agents in B2B sales and marketing is not model capability but data quality. The document, authored by Chief Technology Officer Marco Lagi and prefaced by President Charlie Allieri, frames intent data as the binding constraint on agentic system performance and traces the technical decisions Intentsify has made over seven years to address that constraint.

The paper arrives against a backdrop of mounting industry anxiety over agentic AI deployments. According to Gartner, over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, and inadequate risk controls. That research, announced on June 25, 2025, by Senior Director Analyst Anushree Verma, described most current agentic projects as early-stage experiments driven by hype. Intentsify's white paper takes a complementary but more specific position: the Gartner cancellation rate reflects a failure of inputs, not intelligence.

"An agent operating on incomplete or imprecise signals isn't just underperforming," Allieri writes in the foreword. "It's confidently wrong, and in an autonomous workflow, confident errors scale."

Three eras, one escalating problem

The white paper structures its argument through a historical frame, dividing AI's application in sales and marketing into three periods. The first, beginning in the 2000s, was what it calls the age of prediction - machine learning systems scoring leads, prioritizing accounts, and predicting conversion using behavioral signals such as email opens, link clicks, and form fills. Those systems changed how revenue teams operated. According to Intentsify, they gave marketers a systematic, data-driven mechanism for focusing attention for the first time. Their limit was also plain: they could rank and score, but they could not process language.

The second period began in 2017 when the Transformer architecture was introduced. The key technical innovation was self-attention, a mechanism that allows a model to relate any word to any other word in a sequence globally, rather than processing language as a local chain of neighboring tokens. Training at web scale on next-token prediction forced these models to internalize the statistical structure of language itself - grammar, syntax, semantic associations, and the patterns that encode knowledge and reasoning. The practical result was two complementary capabilities. Encoding converts text into dense numerical vectors, known as embeddings, that capture meaning and context. Decoding turns those vectors and prompts back into text with coherent learned structure.

ChatGPT's release in November 2022 was the inflection point. Long, coherent, context-aware generation became reliably possible, and according to Intentsify, the industry shifted from skepticism to adoption. But that shift surfaced a new problem: if text generation is no longer the bottleneck, the bottleneck moves to precision. Generating the right text, calibrated to deal stage, channel, timing, and buyer persona, requires something the models themselves do not carry - grounded, current, account-specific context.

The third period is 2026 and beyond: the agent era. According to Intentsify, an AI agent is a system that, given an objective, takes actions in an environment to maximize the likelihood of achieving it. The control loop runs through four steps: perception, in which the agent reads its environment such as the current state of a CRM; update, in which it refreshes its memory and beliefs based on new information; decision, in which it selects the next action under uncertainty; and intervention, in which it executes that action through available tools and APIs. After each intervention the loop restarts.

Most production agents do not learn autonomously. Their model weights are fixed. What makes them powerful is their architecture - an LLM for reasoning and generation, tool access for action, a memory layer for maintaining state, and a planner that decomposes complex tasks into steps. Intentsify describes them as the next evolution of workflow automation: more flexible and adaptive than scripts, but not yet self-improving. That architecture carries one constraint no amount of reasoning can overcome. According to Intentsify, "agents can only work with the information they receive."

How Intentsify built its data layer

The second section of the paper describes how Intentsify positioned itself technically at each of those three inflection points. The company, founded in 2018 and headquartered in Westwood, Massachusetts, began building on large language models before the term had broad industry currency.

When the Transformer architecture emerged in 2017, most of the intent data industry remained committed to keyword matching and rigid taxonomies. Those techniques could identify what words appeared on a page but could not determine what a buyer was actually trying to do. Intentsify illustrates the limitation through a concrete example. A buyer at a telecom company researching ways to reduce subscriber churn triggers keyword flags on words like "churn," "retention," or "customer success software." But the same buying process involves a procurement manager reading analyst reports on customer lifetime value modeling, an IT lead comparing integration architectures, and a CFO reviewing ROI benchmarks. None of those activities necessarily contain the flagged keywords. The research pattern is present; the matching words are not.

Intentsify's architectural choice was to treat intent as a language understanding problem. With LLMs, intent becomes a probabilistic, evidence-based interpretation of behavior rather than a list of keywords to match. The model infers what a buyer is trying to accomplish based on the full context of their research. By training on web-scale corpora - vast datasets of text sourced from across the internet, comprising billions of documents - and fine-tuning on proprietary B2B data, Intentsify's patented models learned domain-specific vocabulary and the language patterns of its customers' buyers.

According to Intentsify, those models process billions of text signals each month across intent streams, corporate websites, marketing and sales collateral, and social content. From that corpus, the system does three things: surfaces accounts whose research most closely aligns with a customer's value proposition, infers where those accounts likely sit in the buyer journey, and estimates the strength of each intent signal with calibrated confidence.

That capability solved one problem and exposed a harder one. Knowing what is being researched at the account level does not identify who is doing the research.

The 2024 acquisition and persona-level resolution

Traditional intent data resolution collapses all research activity at the account level, typically using IP-based identification. In a company of dozens, that compression is workable. In a company of thousands, it becomes dangerous. Engineers, procurement teams, executives, and interns browse for different reasons. When all of their activity is attributed to "the account," the signal becomes noise. According to Intentsify, that noise is expensive: it drives spend toward accounts not actually in market while obscuring the buying group members who are.

To resolve this, Intentsify acquired Five By Five (5x5) and its Identity Graph in 2024. The rationale was not simply additional data volume. It was the mechanism the 5x5 data co-op uses to keep data alive. Built around a cooperative model where participants both consume and contribute, the system creates compounding verification loops rather than the slow decay characteristic of static databases. The result is an Identity Graph linking contacts, companies, and digital identifiers at scale.

With that infrastructure in place, intent signals can be resolved from "an account is interested" into "a specific persona is showing behavior consistent with active evaluation." For sales and marketing teams, the shift is material: it enables reaching the right member of a buying committee at the right moment rather than broadcasting to an account and hoping the message finds someone relevant.

That infrastructure received external validation in The Forrester Wave: Intent Data Providers For B2B, Q1 2025, which placed Intentsify at the top of the strength-of-offering hierarchy. The report, published on February 27, 2025, evaluated 15 providers across 21 criteria. According to the Forrester report, "The introduction of the Intentsify identity graph in 2024 ranks among the most significant innovations in the space over the past two years, driving persona-based analysis and improved buying group prediction." Intentsify received the highest overall score for Current Offering and the highest possible score across 12 of the 21 categories. PPC Land covered the expansion of Intentsify's data infrastructure when the company reported more than 20% year-over-year revenue growth and a significant international product expansion on April 15, 2026.

Over the same period, Intentsify was included in the Inc. 5000 Fastest-Growing Private Companies in America for three consecutive years and named Best Interaction Data Analysis Solution in the 2024 MarTech Breakthrough Awards.

What happens when agents get bad data

The third section of the white paper moves from historical architecture to concrete failure mechanics. The central claim is that an AI agent is only as good as the information it receives before it acts. This is not, according to Intentsify, a limitation of any particular model or vendor. It is a constraint from information theory: a system cannot reason its way to accurate outputs from inaccurate inputs.

The paper illustrates the point with a comparison drawn from a backtest conducted by a Fortune 500 telecom. Given the same historical accounts and the same time window, Intentsify identified three times as many close-won opportunities as the incumbent intent provider. The signal quality difference, not the model architecture, produced that gap.

The paper shows what happens when a capable LLM is asked to prioritize accounts without grounded intent data. In the screenshot included in the document, a user presents ChatGPT with a target account list of 100,000 companies and asks which ones to pursue. The model responds that the file contains only one field - the domain name - and proposes prioritizing based on domain characteristics such as commercial top-level domains and business-relevant keywords. The output, as the paper notes, sounds smart. It is generic and ungrounded. The model is reasoning in the absence of data, which makes it "just an agent with an opinion."

Add intent, engagement, timing, and buying stage signals, and the output shifts from generic recommendation to grounded, actionable intelligence. That shift is the stated purpose of Intentsify's data layer.

The white paper's technical section also addresses the longer-term dynamics of bad input data. Since outcomes are only observed on the subset of accounts selected for action, the dataset an agent learns from is inherently subject to selection bias. If input signals are weak, the agent faces two failure paths. It either discards intent as non-informative and stops using it entirely, or it enters a self-reinforcing loop where biased selection produces biased labels, which amplify the original error and accelerate financial damage over time. Either outcome degrades the agent's ability to improve. According to Intentsify, the longer the system runs on bad data, the worse the compounding becomes.

The short-run costs are framed in terms of false negatives and false positives. Higher false negatives reduce gross profit by missing real opportunities. Higher false positives inflate spend by chasing accounts that are not actually in market. Those losses are recoverable. The long-run dynamics are harder to unwind because they corrupt the agent's training signal itself.

Why this matters for B2B marketing teams

The broader context of this white paper sits inside a significant structural shift in how programmatic advertising and B2B marketing are being built. PPC Land has tracked the agentic AI buildout across platforms since at least early 2026, documenting how companies from LiveRamp to PubMatic to Adobe have introduced autonomous agents into campaign management and audience activation workflows. LiveRamp launched agentic orchestration in October 2025, giving autonomous agents governed access to its identity resolution and activation tools. Adobe released three B2B AI agents in October 2025 targeting the complex purchasing cycles characteristic of enterprise B2B. PubMatic launched an agentic advertising operating system in January 2026.

As those agent frameworks multiply, the question of what data feeds them becomes increasingly consequential. The Intentsify white paper is notable not for presenting an entirely new argument, but for grounding it technically. Most public discussion of agentic AI in advertising has focused on agent architecture, workflow automation, and interface design. Intentsify's paper redirects attention upstream, to the data infrastructure that determines what those agents can actually know.

Salesforce research published in June 2025 found that enterprise AI agents fail 65% of multitask conversational tasks, and the Intentsify paper suggests that the problem is at least partially one of data quality rather than model capability. A March 2026 survey by Agentcy found that eight in ten B2B marketing leaders described AI visibility as a blind spot, with only 10% able to link AI touchpoints to revenue.

The white paper positions Intentsify's technology stack - LLM-based intent models, the Identity Graph acquired through Five By Five in 2024, and the infrastructure to process billions of signals a month - as the missing data layer that determines whether agents produce insight or error. According to the company, that foundation was built not in response to the agent era but as preparation for it.

According to CTO Marco Lagi, who holds a PhD in physical chemistry, has conducted research at MIT, and has authored more than 40 academic papers, the technical architecture reflects choices made at each inflection point in the history of AI applied to marketing. His career at the intersection of AI and commercial application began with Kemvi, an AI startup he founded in 2013 that was later acquired by HubSpot. At HubSpot, he built the company's machine learning modeling capabilities from the ground up before moving to Intentsify.

The conclusion of the white paper returns to the fundamental constraint. "Agents will compete on interfaces, workflows, and reasoning capability," it states. "But their recommendation quality will always be bounded by the context they can access. A competitor supplying incomplete or misaligned signals won't be saved by a better model. The model's intelligence can't compensate for what it was never given."

Timeline

  • 2000s - The age of prediction: machine learning applied to sales and marketing produces lead scoring, propensity modeling, recommendation engines, and ad targeting systems based on behavioral signals
  • 2013 - Marco Lagi founds Kemvi, an AI startup later acquired by HubSpot, where he builds the company's machine learning modeling capabilities
  • 2017 - Transformer architecture introduced, enabling self-attention mechanisms that allow models to relate any word to any other word in a sequence globally
  • 2018 - Intentsify founded in Westwood, Massachusetts; begins building LLM-based intent models before the term is widely used in the industry
  • 2022 - ChatGPT released in November 2022, marking the inflection point at which long, coherent, context-aware generation becomes reliably possible; Intentsify establishes LLM foundation stage in its architecture
  • 2024 - Intentsify acquires Five By Five (5x5) and its Identity Graph, enabling persona-level intent resolution and buying group prediction; PPC Land covered Intentsify's B2B audience infrastructure expansion
  • February 27, 2025 - The Forrester Wave: Intent Data Providers For B2B, Q1 2025 published; Intentsify named a Leader with the highest Current Offering score and highest possible scores across 12 of 21 criteria
  • June 25, 2025 - Gartner predicts over 40% of agentic AI projects will be canceled by 2027, citing costs, unclear business value, and inadequate risk controls
  • June 29, 2025 - Salesforce research finds enterprise AI agents fail 65% of multiturn tasks in a new CRM benchmark study
  • October 1, 2025 - LiveRamp launches agentic orchestration, giving autonomous agents governed access to identity resolution and activation tools
  • October 9, 2025 - Adobe releases three B2B AI agents targeting audience building, journey optimization, and data insights for enterprise buying cycles
  • December 10, 2025 - Intentsify executive forecasts B2B targeting shift as search dollars move toward paid social, CTV, and DOOH channels
  • January 5, 2026 - PubMatic launches AgenticOS, an agentic advertising operating system with live campaign management
  • January 27, 2026 - Intentsify acquires Salutary Data, adding a database of 156 million verified B2B contact profiles and AI-enhanced data validation capabilities
  • March 5, 2026 - Agentcy survey finds 80% of B2B marketing leaders cannot link AI touchpoints to revenue
  • April 15, 2026 - Intentsify reports more than 20% year-over-year revenue growth and expands to 700+ pre-built B2B audience segments in The Trade Desk and DV360, with distribution to 90+ countries via LiveRamp

Summary

Who: Intentsify, a B2B intent data and signal-based marketing company headquartered in Westwood, Massachusetts, through its Chief Technology Officer Marco Lagi and President Charlie Allieri.

What: Publication of a white paper titled "Confident, Automated, and Wrong: Why AI Agents Fail Without the Right Context," arguing that AI agent failure in B2B go-to-market systems stems from poor-quality input data rather than model architecture weakness, and presenting Intentsify's technical approach to solving that problem through LLM-based intent modeling and persona-level identity resolution.

When: Published on May 31, 2026, drawing on architectural decisions and product milestones spanning from 2017 to the present, including Intentsify's acquisition of Five By Five in 2024 and recognition in The Forrester Wave: Intent Data Providers For B2B, Q1 2025, published on February 27, 2025.

Where: The white paper is global in scope, addressing enterprise sales and marketing teams deploying AI agents in go-to-market workflows, with Intentsify's infrastructure built to operate at production scale across billions of monthly signals.

Why: As agentic AI frameworks proliferate across ad tech and B2B marketing - from LiveRamp and Adobe to PubMatic and The Trade Desk - the question of what data feeds those agents becomes commercially decisive. Intentsify's paper makes the case that intent data quality, not model sophistication, determines whether autonomous agents compound business advantage or compound business error.

Share this article
The link has been copied!