Why 99% of Marketers See No ROI from AI Yet, and What the 1% Do Differently

Most marketing teams are using AI. Almost none of them are seeing it in their margins. That is not a tool problem. It is a context problem, and the gap between the 99 percent and the 1 percent is almost entirely explained by how each group handles it.

The field signal is hard to ignore

Survey after survey in 2024 and 2025 showed AI adoption in marketing climbing past 70, 80, even 90 percent. Usage metrics look great. ROI metrics do not. When you ask performance marketers and agency founders whether AI has actually reduced headcount per account, shortened review cycles, or improved margin, the answer is almost always no.

The confusing part, as one agency founder put it: the tools are not bad. Output is fast. Copy is decent. Briefs come together faster than they used to. And yet the P&L looks roughly the same as it did before, except now there is a new line for AI subscriptions and API spend.

Something is eating the gains before they reach the bottom line.

Three places the ROI disappears

The senior review tax

Every AI output in a performance marketing context eventually reaches a moment where a senior person has to decide whether it is actually true for this account, this client, this week.

Would you let an AI send the performance report that goes directly to a client's CMO? Would you let it launch a campaign with $200,000 behind it without a human checking the targeting logic against decisions made last Monday?

Most teams would not. So the senior reviewer stays in the loop on every consequential output. That review cycle is the same headcount you were hoping not to need. The AI made the first draft faster. It did not remove the person who validates the first draft.

The result: you are priced as if you are automated and staffed as if you are not.

Context typed by hand, every single time

AI tools do not arrive knowing your account. They arrive knowing the internet. Before you can ask a useful question, someone has to supply the background: why the account is structured the way it is, what the current campaign objective is, what leadership decided at the last QBR, what tests are running, what Legal killed last month.

One media buyer described it this way: "I spent twenty minutes typing background before I could ask the actual question." Multiply that across a team of ten, across fifty accounts, across every week of the year, and you have not saved labor. You have moved labor from execution into prompt preparation.

This is the context tax. It does not show up on any AI vendor's case study, but it shows up on your team's calendar.

Stale briefs and wrong objectives

Even teams that do the work of writing briefs and uploading brand guidelines hit a second problem: the context goes stale the moment it is saved.

Budget shifts on Monday. The brief loaded into the AI tool is from March. The KPI changed at the last quarterly business review. The competitor launched a new product and the team is talking about nothing else, but the agent has no idea it exists.

The consequence is AI recommendations that are confidently wrong in ways that are hard to catch quickly. "It's still optimizing for a KPI we dropped." "It wrote copy around a claim Legal killed last month." "It proposed the exact test we ran and killed in Q1."

Each of those errors requires a senior person to catch it, which brings you back to the review tax. The staleness and the review tax compound each other.

What the 1 percent do differently

Teams that actually see ROI from AI are not using better models. They are solving the context problem in a way that makes the senior review tax shrink rather than persist.

Three things separate them.

They give the AI history, not just a prompt

Instead of typing background into each session, these teams have built or adopted systems where the AI already holds the account's history: past tests and their outcomes, structural decisions and the reasoning behind them, brand and legal constraints, seasonal patterns from prior years.

The practical effect is that the AI stops proposing things the team already ruled out. "You tested this in March and it failed on creative, not targeting. Here is the variant worth running" is a materially different output than "here are fourteen things you could test."

History is not the same as data. Data is impressions and clicks. History is the reasoning layer on top of the data: why you made the decisions you made, what you learned, what you decided to stop doing.

They keep the context current automatically

The teams seeing ROI are not re-uploading briefs. They have connected the AI to the places where context is actually created: meeting recordings, Slack, ad platform signals, CRM updates. When the budget shifts on Monday, the AI knows on Monday. When a campaign is sunset in a team call, the automation built on top of that AI does not scale it on Tuesday.

This is the difference between a static knowledge base and a live context layer. The static knowledge base requires a human to update it, which means it is always at least one decision behind. The live context layer updates as the business moves.

They let the AI act on judgment, with guardrails, not just generate drafts

The final difference is the hardest to replicate. The 99 percent use AI to produce outputs that humans then review and act on. The 1 percent have found specific, bounded places where the AI can act directly, because the context layer gives it enough grounding to be trusted.

This is not about removing all human oversight. It is about being precise about where the senior review is actually necessary and where it is just a habit from a time when the AI did not have enough context to be trusted.

A workflow that automatically adjusts bids based on a rule is a trigger. A workflow that adjusts bids based on a rule but also knows that the team decided to protect brand terms this week regardless of efficiency, because a competitor just moved into the space, is judgment. The first requires a human to catch the edge cases. The second requires a human only when something genuinely novel happens.

Why most teams cannot get there from here

The honest answer is that building this kind of context layer is not trivial. The context that matters in performance marketing is created continuously, in meetings and Slack threads and decisions that never make it into a doc. It is partially captured, manually, and usually late. And even when it is captured, it is rarely ranked by what actually matters this week versus what is just noise.

Connecting more data sources does not solve it. Several teams have found that connecting everything made the outputs worse, not better. The agent started weighing a throwaway Slack message the same as a strategy call with the CEO. More inputs without prioritization is just more noise.

The problem is not access to data. It is classification, recency weighting, and business impact ranking applied before the agent ever sees the context.

Where Almond fits

Almond is built specifically around this problem. It positions itself as a Business Context Engine for performance marketing, reading ad accounts, Slack, Notion, Sheets, email, meeting notes, and CRM in read-only mode, and classifying everything it captures by relevance, recency, and business impact before surfacing it to any agent or workflow.

The products built on top of it, Canvas for visual dashboards and Workflow Builder for Google and Meta automations, are designed to draw on that context layer rather than operate independently of it. The Workflow Builder runs on judgment rather than triggers, so it will not execute an action the team already decided against in a Monday meeting.

For agencies specifically, the context is scoped per client, which means one client's strategy cannot surface in another client's recommendations, and when an account manager leaves, the account history stays with the agency.

Almond is in early access. Every team that joins gets a detailed 1:1 audit of their Meta and Google ad accounts with the CEO, recorded. The waitlist is open at almond.ai and requires a work email.

The 1 percent are not using smarter AI. They are using AI that knows more about their business, keeps that knowledge current, and acts on it with enough guardrails that the senior review tax finally starts to shrink. That is the gap, and it is almost entirely a context problem.