Key takeaways
- AI tools deepened key person risk at agencies because useful output requires account context that only one person holds.
- When an account manager leaves, they take the relationship and the reasoning behind every structural decision, unless context is captured at the agency level.
- Documentation does not fix the problem. Context must be captured automatically and kept current to be reliable.
- Versioned, client-scoped context means onboarding compresses and accounts survive resignations intact.
- The fix separates context from the person who holds it, so the agency owns the knowledge rather than renting it from employees.
When agencies started adopting AI tools in earnest, the pitch was straightforward: reduce dependency on any single person, spread institutional knowledge across the team, make accounts more resilient. Two years in, most agency founders I have spoken to report the opposite happened. The AI concentrated key person risk rather than distributing it.
Here is why that happened, and what a structural fix actually looks like.
Why AI Made the Problem Worse
The core issue is not the quality of the AI tools. Most of them are genuinely capable. The issue is what they require to produce useful output on any specific account.
To get a good answer from a general-purpose AI, someone has to supply the context first. Why the account is structured the way it is. What tests were run and killed last quarter. What the client decided on the Monday call. What creative claims Legal has already rejected. None of that lives in the AI. It lives in the person who has been on the account for 18 months.
So the workflow that emerged at most agencies looks like this: the senior account manager types in the background, gets the AI output, then checks whether the output is actually true for this client. The AI is doing the drafting. The human is doing everything that requires knowing the account.
That is not a bad workflow. But it means the AI is only as useful as the person operating it. And that person is exactly who you were hoping not to depend on so heavily.
The Leave Problem
The symptom most agency founders describe first is the leave problem. When the account lead is out, nobody else can work the account effectively. Not because the rest of the team is less capable in general, but because they do not hold the context. The AI cannot help them fill that gap because the AI does not hold the context either.
A two-week vacation becomes a two-week slowdown on a client relationship that took years to build. A medical leave becomes a genuine operational risk. The tool that was supposed to make the team more resilient made it more fragile.
The Resignation Problem
The leave problem is temporary. The resignation problem is permanent.
When a strong account manager leaves, they take three things: the client relationship, the reasoning behind every structural decision in the account, and the only prompt library that produced reliable output for that client. The first one you expect to lose. The second and third ones you did not realize were at risk until they were gone.
The new hire spends weeks asking why things are set up the way they are. Nobody knows. The account history exists in meeting recordings nobody has time to watch, Slack threads that scroll back 18 months, and email chains that were never organized. The AI cannot synthesize any of it because it was never structured or retained anywhere the AI can reach.
Estimates vary, but onboarding a replacement account manager on a complex paid media account typically takes four to eight weeks before they are operating at full effectiveness. During that period, the client notices.
What the Fix Is Not
The instinctive response is documentation. Write everything down. Build a wiki. Require account leads to maintain a running brief.
This does not work, and most agency founders already know it from experience. Documentation requires discipline that competes directly with billable hours. It gets done once, then goes stale. The budget that shifted on Monday does not make it into the wiki until someone remembers to update it, which is usually after something has already gone wrong. And even a well-maintained wiki is a static artifact. It does not tell the AI what matters most this week versus what is historical background.
The other common response is to build redundancy into the team structure: two people on every account, mandatory knowledge transfer sessions, paired account management. This helps at the margin but does not solve the underlying problem. It also increases headcount per account, which is the opposite direction from where agency economics need to go.
What the Fix Actually Requires
The fix requires separating context from the person who holds it, and doing that continuously rather than as a manual task.
Three properties are necessary for this to work in practice.
Context has to be captured automatically. If capturing context requires a human to do it, it will not happen consistently. The account lead will capture it when they have time, which means it will be incomplete, delayed, and biased toward the decisions they remember as significant. The meeting recordings, Slack threads, email, and ad platform data need to feed into a system that structures and retains them without anyone writing a brief.
Context has to stay current. A static document is not context. It is a snapshot. The account's actual context changes every week: new tests, shifted budgets, updated creative direction, revised KPIs from the last client call. The system needs to update as those changes happen, not when someone remembers to log them.
Context has to be scoped to the client and owned by the agency. This is the part most tools miss entirely. The context needs to live at the agency level, isolated per client, versioned so you can audit what the system knew and when. When an account manager leaves, they take the relationship. They do not take the account's reasoning, history, or institutional knowledge.
How Versioned, Client-Scoped Context Changes the Operational Reality
When context is captured continuously, kept current, and held at the agency level, a few things change in practice.
Coverage during leave is functional, not heroic
Another team member can work the account because the context they need is accessible. They are not starting from zero or relying on a handoff document that was written in a hurry. The AI can answer questions about why the account is structured a certain way, what was tested and when, and what the current priorities are, because that information is in the system rather than in one person's memory.
Onboarding time compresses
A new hire on an account with versioned context is not starting from scratch. The reasoning behind structural decisions is documented and accessible. The test history is retained. The things that were tried and killed are on record with the reasoning attached. Onboarding that used to take six weeks can realistically compress to two or three, because the new person is learning from a system rather than interviewing colleagues who are already stretched.
The account survives the resignation
When someone leaves, the client relationship may follow them. That is a real risk and no technology fully eliminates it. But the account itself, the structure, the history, the reasoning, the things already ruled out, stays with the agency. The next person on the account is not rebuilding from memory. They are inheriting a documented, versioned record of everything that happened and why.
Where Almond Fits
Almond's Business Context Engine was built to solve exactly this structural problem. It connects to the sources where account context actually lives: ad platforms, Slack, email, meeting recordings, Notion, Sheets, and CRM. It captures context automatically as it happens, without requiring anyone to write a brief or maintain a document.
Every piece of context is versioned and retained, so the test you ran and killed in March is on record with the outcome and the reasoning. Every client's context is isolated and scoped at the agency level, so one client's strategy cannot surface in another client's recommendations, and the context belongs to the agency rather than to the account manager who held it.
When someone is on leave, the rest of the team can work the account because the context is in the system. When someone resigns, the account history stays. The Canvas dashboard and Workflow Builder both draw on this context layer, so automations act on current account reality rather than stale assumptions.
Almond 2.0 early access is open at almond.ai. Every team that joins gets a detailed 1:1 Meta and Google ad account audit with the CEO, recorded. Waitlist requires a work email.
The goal is straightforward: your expertise becomes something the agency owns, rather than something you hope does not resign.
Frequently asked questions
- Why does AI increase key person risk at agencies instead of reducing it?
- AI tools require context to produce useful output on any specific account. That context lives with the person who has worked the account longest. So the AI is only as useful as the person operating it, which deepens rather than distributes the dependency on that individual.
- How do agencies retain institutional knowledge when account managers leave?
- The most durable approach is capturing context automatically from the sources where it is created: meetings, Slack, email, and ad platforms. When context is versioned and held at the agency level rather than in a person's memory, a new hire inherits the account's history rather than rebuilding it from scratch.
- What is client-scoped context and why does it matter for agencies?
- Client-scoped context means each client's data, decisions, and history are isolated and retained separately at the agency level. It matters because it ensures one client's strategy cannot surface in another's recommendations, and it means the context belongs to the agency rather than to the individual who managed the account.
- How long does it take to onboard a new account manager on a paid media account?
- Without structured knowledge retention, onboarding typically takes four to eight weeks before a new account manager operates at full effectiveness on a complex paid media account. With versioned context that captures test history and decision reasoning, that window can compress to two to three weeks.
- Does documentation solve the agency key person risk problem?
- Not reliably. Documentation requires consistent discipline that competes with billable hours, goes stale quickly, and is never complete. It also does not tell the AI what is most relevant this week versus historical background. Automatic, continuous context capture is more durable than any manual documentation process.
Almond 2.0 early access
Get a 1:1 Meta and Google audit with the CEO.
Every waitlist member gets a 1:1 audit of their ad accounts. Work email only.