Why AI Did Not Fix Agency Margins: Two Years In, Most Founders Saw the Opposite

Two years after most digital agencies adopted AI tools, headcount per account is roughly where it was, a new cost line appeared on the P&L, and margins for many founders got tighter, not wider. The tools are not the problem. The problem is structural.

What everyone expected in 2024

The pitch was simple: leaner teams, faster turnarounds, same client roster handled by fewer people. Agencies were going to run a bit more like software companies. Gross margins would expand. Senior time would shift from execution to strategy.

None of that happened at scale. Output got faster and often genuinely better. The economics did not follow.

Three structural reasons keep coming up when agency founders talk honestly about what went wrong.

Reason one: context is messy and AI cannot hold it

AI works best when the problem is well defined and the answer is either right or it is not. That is a large part of why it landed so hard in software development. The spec is written down, the test passes or fails, and the relevant context lives in the repo.

Performance marketing is not that.

Almost every decision in an agency day is a context call. Why attribution is set up the way it is. Why that creative got pulled two months ago. What the team tested last quarter and killed. What the client actually decided on Monday's call. What Legal rejected. What the brand positioning shift means for the media plan.

Almost none of that is written down anywhere a tool can find it. It lives in meeting recordings nobody indexed, Slack threads that scroll away, email chains, and the heads of two or three senior people on the account.

So when a marketer opens an AI tool and asks a real account question, they spend fifteen to twenty minutes typing background before they can ask the actual question. Or they skip the context and get an answer that is confident, well formatted, and wrong for this client. The tool proposed the exact test the team ran and killed in March. It wrote ad copy around a claim Legal rejected last month. It is still optimizing for a KPI the client dropped at the last quarterly business review.

The senior person still has to check every output against account reality. That review step is not optional, and it does not get faster just because the first draft arrived in ten seconds.

Reason two: reliability keeps senior people in the loop

Would you let AI send the performance report that goes straight to a client's CMO without a senior review? Would you let it configure a campaign with $200,000 in monthly spend behind it and press go?

Most agency founders would not. Not yet. The output is good enough to accelerate a skilled person, but not good enough to replace the judgment call at the end.

That means the review cycle stays. The same senior account manager who was in the loop before is still in the loop now. The AI produced a faster draft. The headcount requirement did not change.

This is not a criticism of the tools. It is a description of where reliability actually sits in 2025 and 2026 for work where a mistake costs a client relationship, not just a failed test. The gap between "useful assistant" and "trusted autonomous actor" is still wide enough that agencies are staffing for the former while hoping for the latter.

Reason three: in-house tools added a second permanent cost line

A meaningful number of larger agencies responded to the AI moment by building internal tools. Custom prompt libraries, internal dashboards, proprietary workflows. Reasonable instinct: if the tool is the moat, own the tool.

The problem is that software has two costs. Building it and maintaining it. Both are permanent. So now those agencies are paying engineers to build, paying engineers to maintain, paying for the underlying AI APIs, and paying for whatever third-party tools they kept. Two cost lines instead of one, neither of which existed two years ago.

For agencies that did not build in-house, the cost structure is simpler but the problem is the same: a new AI spend line appeared without a corresponding headcount reduction to offset it.

What would actually move margin

The three problems above share a root cause. The context that makes AI output usable for a specific account lives inside people, not systems. As long as that is true, every AI workflow has a human tax attached to it: someone to supply the context at the start, someone to verify the output at the end.

Reducing that tax is what would actually move margin. Not faster output generation. Not more integrations. Not better dashboards.

Specifically, three things would change the economics:

Context that captures itself

If the system reads meetings, Slack, email, ad platform data, and client decisions as they happen, the fifteen-minute briefing before every AI query goes away. The context is already there. The senior person spends time on judgment, not on re-entering information the team already produced.

Context that stays current without human maintenance

The brief loaded into most AI tools is from three months ago. The budget shifted on Monday. The competitor launched last week. The brand position changed at the last QBR. A context layer that updates automatically as decisions happen means the AI is working from this week's reality, not last quarter's snapshot.

Context that lives in the agency, not in the account manager

When the best account manager on a client leaves, they take the relationship and the reasoning. Every test result, every judgment call, every thing the team already ruled out, gone. If that institutional knowledge is captured, versioned, and held at the agency level, the key-person risk collapses. A new person on the account can be productive in days rather than weeks. The agency owns the expertise instead of renting it from whoever currently holds it.

These three things are also the core of what would let an agency demonstrate to a client why their AI is different from the $20-a-month tool the client opened last Tuesday. Right now, most agencies cannot show that difference. They can assert it. They cannot prove it.

The honest state of the market

Several tools are moving toward parts of this. Platforms like Madgicx, Adalysis, and others position themselves as AI-powered ad management and optimization layers, based on their public materials. Verify current feature sets and pricing on their sites. Most of them are strong on the ad platform side, which is the 30 percent of performance marketing that is already well served by AI.

The 70 percent, the business context that explains why the account is structured the way it is, what was tested and killed, what leadership decided on Monday, is largely unaddressed by tools that were built around ad platform data.

The agencies that will expand margin over the next two years are the ones that solve the context problem, not the ones that add another dashboard.

Where Almond fits

Almond is built specifically around this gap. The Business Context Engine reads ad accounts, Slack, Notion, Sheets, email, meeting notes, and CRM in read-only mode, captures every test, result, and decision as it happens, keeps it current, and prioritizes it by relevance, recency, and business impact. Every client context store is isolated, versioned, and auditable. The agency owns the context, not the account manager.

The products built on the engine, Canvas for AI-powered visual reporting, Workflow Builder for judgment-based Google and Meta automations, and open integrations including MCP, are designed to draw on that context layer rather than require a human to supply it each time.

Almond 2.0 early access is open at almond.ai. Every member gets a detailed 1:1 Meta and Google ad account audit with the CEO, recorded. Waitlist requires a work email.