In the AI era, the in-house-versus-agency decision is no longer mainly about cost. Keep the layers that compound — brand judgement, first-party data, and the client memory that makes AI improve over time — and use agencies for elastic production and specialist execution. The deciding question has moved from "what is cheaper?" to "what gets better the more we do it?" That is a different test, and it points to a different split.
I first wrote about in-housing versus outsourcing in 2007, when the argument was almost purely economic: hire the capability or rent it. The economics still matter. But AI changes what "the capability" is, because the value of a team now depends partly on what it accumulates. That is the part most in-housing debates miss.
What changed since the cost debate
The old calculus was simple. An agency has scale, so it can spread specialist cost across clients and charge you less than a full-time hire. You outsource when the need is intermittent or specialist; you in-house when the need is continuous and central. That logic still describes production capacity and channel expertise.
What AI adds is a second axis: some work gets more valuable the more context you feed it. A model that has seen two years of your briefs, results, brand rules, and past decisions produces better work than a fresh one. That improvement accumulates where the context lives. So the question is not only "who can do this cheaper?" but "where do we want the learning to pool?" — and that answer favors keeping the compounding layers close.
The deciding question: compounding or elastic
Every capability you can either keep or outsource falls on a spectrum between compounding and elastic. Compounding work gets better as it accumulates context you own: brand judgement, audience understanding, performance history, the record of what you decided and why. Elastic work is capacity you need sometimes and whose value does not depend on your specific history: production volume, overflow, a specialist channel launch.
Compounding capabilities belong in-house, because outsourcing them means renting out the very thing that makes you better. Elastic capabilities belong with a partner, because building permanent capacity for intermittent work is waste. Most mistakes come from confusing the two — keeping elastic production in-house out of habit, or handing a compounding layer to an agency and then wondering why the work never improved.
The capability map
| Capability | Keep in-house | Agency | Why |
|---|---|---|---|
| Brand and category judgement | Yes | — | Compounding, proprietary, hard to brief |
| First-party data and memory | Yes | — | The learning pools here; you must own it |
| Strategy and measurement design | Yes | Limited | Sets the standard everything else inherits |
| Media planning and buying | Shared | Shared | Judgement in-house, execution capacity rented |
| Creative production at volume | — | Yes | Elastic; the tool made it cheap and fast |
| Localization and transcreation | — | Mostly | Needs native nuance at scale |
| Specialist channels | — | Yes | Intermittent expertise, hard to staff permanently |
| Reporting and intelligence | Yes | Shared | Interpretation is compounding; assembly is elastic |
| Governance and policy | Yes | — | You are accountable; it cannot be delegated |
The table is a starting position, not a rule. A small brand may reasonably outsource more than a large one; a large brand may in-house production because it has the volume to justify it. The axis that matters is compounding, and each row is decided by where the learning needs to land.
Why brand judgement stays in
Brand judgement is the clearest in-house call, and AI makes it clearer. A model can generate on-brand work only if it knows what "on-brand" means for you — which is a body of taste, precedent, and context that a partner never fully holds. The more options a model produces, the more valuable the person who can tell which one is right becomes.
This is the trap in outsourcing brand judgement to an agency: you can delegate the work but not the knowledge of your own brand. Every round of feedback you give is context you are teaching someone else's model. Over time, the improving asset sits with them, not you. If brand judgement is how you compete, that is the wrong place for the learning to pool.
Why data and memory stay in
First-party data and the memory layer around it are the compounding layer that matters most, and the one most at risk of being outsourced by accident. If an agency builds the context layer that makes your campaigns smarter — the briefs, the results, the decision log — then the compounding benefit accrues to the agency relationship, and you lose it the day the relationship ends.
I go deeper on this in building a client-memory data moat. The short version: the model is a commodity you can rent; the memory is the asset that makes the rented model useful. Keep the memory. Rent the model. Governance is the same story — accountability cannot be outsourced, so the rules about what the model may see and who owns the output stay with you. The data-governance piece covers that in detail.
What to outsource
Outsource the elastic layer. Production volume, overflow, campaign localization, a specialist channel you launch twice a year — all good partner work. These are capabilities where the value does not depend on your specific history, so there is no compounding loss in renting them, and real cost in building them permanently.
The AI-native partner here is genuinely useful, because a workflow-driven shop can produce and adapt work at a speed and cost your in-house team cannot match for intermittent needs. The point is to buy capacity without buying the learning. Keep the briefs, the standards, and the decision record on your side of the wall.
The transition, if you are re-splitting
Most teams do not start from a clean slate; they already have a mix, and it drifted rather than got designed. The re-split works best as a one-time audit. List every capability you currently buy or build, mark each one compounding or elastic, and you will usually find two or three that are on the wrong side. The common finds are an outsourced measurement function that should be interpreting your own data, and an in-house production team doing elastic work a partner could absorb.
Move one capability at a time. Pulling a compounding layer in-house is a project — the context has to be migrated and the memory rebuilt on your side — so do it deliberately, not in a panic when a contract is up for renewal. And be honest about what the partner built: if the workflow contains your context, extracting it is part of the move, not an afterthought.
The contract implication
The split changes what the contract should say. If the compounding layers stay in-house, the partner agreement should be explicit about data: what the agency may retain, what must be returned or deleted at the end, and who owns anything derived from your context. The default in most contracts is silent on this, which quietly hands the compounding asset to the party that built the workflow.
Ask three questions of any partner. What context do you keep after the engagement ends? Can we export the memory layer you built? Who owns the derived data? If the answers are vague, that is the real risk in the deal — more than the rate card.
FAQ
Should a brand in-house its AI tooling or use an agency's?
Always in-house the context the AI runs on — your data, brand rules, and decision history — even if you use a partner's tools. Tools are replaceable; the context is the asset. You can rent the model and the workflow while owning the memory that makes them useful.
Is it ever right to outsource strategy?
For a specific launch or a market you do not know, yes, as a temporary input. But strategy that compounds — the standard for how decisions get made — belongs with the team accountable for the results. Borrow a point of view; do not outsource the accountability.
What if we cannot afford a senior in-house team?
That is where the elastic model earns its place. Rent senior judgement for specific decisions and use a partner for production, but keep the context and the accountability. The goal is to own the compounding layer even at small scale.
How do we decide service by service?
Ask whether the work gets more valuable as it accumulates context you own. If yes, keep it in-house. If its value is independent of your history, outsource it. That single test resolves most cases without a long debate.
The short version
In the AI era, keep the compounding layers in-house — brand judgement, first-party data, and the memory that makes AI improve — and use agencies for elastic production and specialist capacity. The old cost question still applies, but it now sits under a bigger one: where do we want the learning to pool? Owning that answer is the whole game. The Agency-Model guide covers the rest of the split, and the for in-house teams track works it through governance and measurement.
If you have run this split on your own team, I would like to hear which layer turned out to be harder to keep than you expected.
Cheers, Chandler