TL;DR
- The pattern that works is preparation, not delegation. Claude Code handles research, first drafts, scoring, and assembly. People handle relationships, judgement, and the final call. Teams that invert this produce confident output that is wrong.
- Four building blocks do most of the work: a project instruction file, reusable skills, subagents for parallel work, and plan mode for anything with side effects.
- Measure cycle time, not output volume. More drafts produced is not a result. The same quality of work produced with a fraction of the preparation time is.
- Departments differ enormously in payoff. Research-heavy functions such as search and sales preparation compress dramatically. Relationship-heavy functions barely move.
- The failure mode is not bad grammar, it is confident specificity that happens to be wrong. Every figure that reaches a client or a prospect needs a human who checked it.
Contents
- Where the leverage actually is
- The four building blocks
- Department by department
- The division of labour that holds up
- A six week rollout
- What to measure and what to distrust
- FAQ: Running an Agency on Claude Code
Where the leverage actually is
Most agency work decomposes into four activities: gathering context, producing a first version, judging whether it is any good, and having the conversation that turns it into revenue.
The first two are mechanical and expensive in hours. The third is judgement. The fourth is relationship. A coding agent with filesystem access, tool use, and long context is extremely good at the first two, unreliable at the third, and irrelevant to the fourth.
That is the whole thesis, and it explains why the results vary so widely between teams. Agencies that deploy it against context gathering and first drafts report large, durable time savings. Agencies that deploy it against judgement produce plausible work that a client eventually catches, which costs more than the time it saved.
| Activity | Share of typical agency hours | Automatable | Risk if wrong |
|---|---|---|---|
| Gathering context | High | Yes | Low, caught in review |
| First draft | High | Yes | Low, caught in review |
| Judging quality | Moderate | No | High |
| Client conversation | Moderate | No | Very high |
The honest expectation is that research tasks compress from hours to tens of minutes. That is a substantial and real gain. It is not the same as replacing a function, and framing it that way internally guarantees disappointment and resistance.
The four building blocks
Everything below is built from four primitives. Learn these before deploying anything.
The project instruction file
A file at the root of each project that the agent reads automatically. It carries the standing context: what the client sells, who they sell to, the house style, the constraints, the things that must never appear in output.
This is the highest-leverage artefact in the entire setup and the one most teams underinvest in. A thorough instruction file removes the need to re-explain context every session, and it is the mechanism by which output quality becomes consistent across people. Treat it as a living document owned by whoever runs the account.
Skills
Reusable, named procedures for tasks you perform repeatedly: build a research object on an account, audit a sequence against deliverability rules, produce a first draft brief in house format.
A skill is the difference between a capable assistant and a repeatable process. The test of whether something should become a skill is simple: if two people would do it differently and both ways are acceptable, it should be a skill so that it stops being a coin flip.
Subagents
Parallel workers handling separate parts of a task with their own context. Useful when a job splits cleanly, for example researching twenty accounts at once, or reviewing a long document against several different checklists simultaneously.
The constraint is coordination. Subagents work well on independent tasks and badly on tasks where each step depends on the previous one, which is the same rule that applies to delegating to people.
Plan mode
Requiring the agent to produce a plan for approval before acting. Mandatory for anything with side effects: sending, publishing, deleting, or changing client-facing assets.
The rule worth adopting: read operations run freely, write operations require a plan. This single convention prevents the majority of the incidents teams report.
Department by department
Ordered roughly by payoff, which is not the order most agencies start in.
Search and content research
Highest payoff of any function. Competitive analysis, gap identification, source gathering, outline construction, and fact assembly are almost entirely context work. Teams report research phases compressing from a few hours to tens of minutes, with a corresponding multiple on articles shipped.
What stays human: the angle, the argument, and the decision about what is actually worth saying. A model produces the median of what already exists, which is precisely what content differentiation must avoid.
Sales and go-to-market preparation
Very high payoff. Building the research object for an account, drafting sequence variants, scoring accounts against a written profile, and preparing call briefs. A full campaign preparation cycle that took most of a day can run in well under an hour.
What stays human: who is worth pursuing, the final wording of anything a prospect sees, and every conversation. Also every number: a fabricated benchmark in an outbound email is a credibility loss you cannot recover.
Advertising operations
High payoff. Account structure audits, naming convention enforcement, budget pacing checks, creative variant generation, and cross-platform reporting assembly. Managing substantial monthly spend across several platforms is largely a data reconciliation problem, and reconciliation is mechanical.
What stays human: budget allocation decisions and anything involving what to say rather than how to structure it.
Design and production
Moderate to high payoff, mostly in the preparation layer rather than the creative one. Brief construction, asset inventory, copy blocking for layouts, and producing the variants of a finished concept for different placements.
What stays human: the concept. Every time.
Lead magnets and tooling
High payoff, and underrated. Small interactive tools, calculators, and assessments are the most effective lead magnets available and were historically expensive to build. An agent with code execution collapses the cost of a simple tool from a development project to an afternoon.
What stays human: deciding which tool is worth building, which is a positioning question rather than a technical one.
Closing and account management
Low payoff, and appropriately so. Preparation for calls, follow-up drafting, and proposal assembly all benefit. The conversation itself does not, and clients notice immediately when correspondence stops sounding like a person.
What stays human: effectively all of it.
Recruiting
Moderate payoff. Role description drafting, structured screening against written criteria, and interview question preparation. Genuine caution required here: automated screening against candidate data raises fairness and legal questions that vary by jurisdiction, and the fact that a tool can rank people does not mean it should.
What stays human: every decision about a person.
| Department | Payoff | Primary use | Main risk |
|---|---|---|---|
| Search and content research | Very high | Research and outlines | Median output |
| Sales and GTM preparation | Very high | Research objects, drafts | Fabricated specifics |
| Advertising operations | High | Audits and reconciliation | Misread data |
| Lead magnets and tooling | High | Building small tools | Building the wrong one |
| Design and production | Moderate | Briefs and variants | Concept dilution |
| Recruiting | Moderate | Descriptions and screening | Fairness and legal |
| Closing and accounts | Low | Call preparation | Losing the human voice |
The division of labour that holds up
One sentence, and it is worth writing on a wall: the agent prepares, the person decides.
Everything upstream of a judgement is a candidate for automation. The judgement itself is not, and neither is anything downstream where a person is on the other end.
Three corollaries.
Never ship unreviewed output to a client or a prospect. Not because the writing is poor, but because the failure mode is confident specificity that happens to be wrong. A fabricated statistic reads exactly like a real one, and the reader will assume you checked.
Every figure traces to a source someone verified. Make this a documented rule with a named owner rather than an assumption. Research output is a lead to verify, not a fact.
Positioning stays entirely human. Models generate the centre of the distribution of what already exists. Positioning is the deliberate act of not being there.
A six week rollout
Agencies that succeed with this roll it out narrowly and deepen. Agencies that fail deploy it everywhere at once and generate distrust in week two.
Weeks one and two: one function, one person. Pick the highest-payoff research-heavy function, usually content research or sales preparation. One capable and interested person builds the project instruction file and works normally with the agent alongside. Objective is a working instruction file and an honest sense of where it helps.
Weeks three and four: turn repetition into skills. Whatever that person did more than three times becomes a named skill with a written procedure. This is where the compounding starts, because a skill is transferable and a habit is not.
Week five: second function, second person. The first person teaches the second. Crucially, the second function should be different in character so you learn where the approach does not transfer.
Week six: write the rules and the boundaries. What may run unattended, what requires plan mode, what never ships without review, who owns verification of figures, and what data may not be shared with the tool at all. Then extend.
| Phase | Weeks | Objective | Failure signal |
|---|---|---|---|
| Single function pilot | 1 to 2 | Working instruction file | Nobody uses it after week one |
| Codify repetition | 3 to 4 | Named, reusable skills | Every session starts from scratch |
| Transfer | 5 | Second person productive | Only one person can operate it |
| Governance | 6 | Written rules and limits | Incidents before rules exist |
What to measure and what to distrust
Measure cycle time, not output volume. More drafts produced is not a result and is frequently a symptom of a review bottleneck forming. The metric that matters is time from brief to approved deliverable.
Measure review burden. If output volume triples and review time triples with it, you moved the constraint rather than removing it. A healthy deployment shows review time per deliverable falling as the instruction file improves.
Measure error escape rate. How often something factually wrong reaches a client or a prospect. This should be zero, and if it is not, the verification rule is not being enforced rather than the tool being unsuitable.
Distrust output-per-person claims. Published figures about multiples of output are usually counting drafts rather than approved deliverables, and they rarely account for the review time absorbed elsewhere.
Distrust vendor performance figures generally. Claims about lead volume increases and cost reductions from AI adoption originate largely in surveys with obvious commercial interests. The defensible claim is narrower: research compresses from hours to minutes, and that alone justifies the investment.
| Metric | Direction | Why it matters |
|---|---|---|
| Time from brief to approved | Down | The real productivity measure |
| Review time per deliverable | Down | Shows the instruction file is working |
| Error escape rate | Zero | Non-negotiable |
| Drafts produced | Ignore | Measures activity, not output |
AI vs SDR Cost Comparison Template: Download Free
Get the practical framework for applying this article to your GTM system.
Download the Cost Comparison Template →
FAQ: Running an Agency on Claude Code
What can Claude Code actually do for an agency?
It is strongest at gathering context, producing first drafts, scoring against written criteria, and assembling reports, which together account for a large share of agency hours. Research-heavy functions such as content research and sales preparation compress most, with research phases moving from a few hours to tens of minutes. It is unreliable at judging whether work is good and irrelevant to client conversations, so the working rule is that the agent prepares and the person decides.
Which departments benefit most?
Search and content research and sales preparation benefit most, because both are dominated by context gathering. Advertising operations follows closely, since managing spend across several platforms is largely a reconciliation problem. Lead magnet and small tool building is underrated, because interactive calculators and assessments were historically expensive and now take an afternoon. Closing and account management benefit least, and appropriately so, since clients notice immediately when correspondence stops sounding like a person.
What are the building blocks I need to set up?
Four. A project instruction file at the root of each project carrying standing context, house style, and prohibitions, which is the highest-leverage artefact and the most underinvested. Named skills for procedures you repeat, so they stop varying by person. Subagents for genuinely parallel work such as researching many accounts at once. And plan mode for anything with side effects, under the simple convention that read operations run freely while write operations require an approved plan.
How do I stop it producing factually wrong output?
Treat every research output as a lead to verify rather than a fact, and make verification a documented rule with a named owner rather than an assumption. The failure mode is not bad grammar, it is confident specificity: a fabricated statistic reads exactly like a real one and the reader will assume you checked. Nothing should reach a client or a prospect unreviewed, and every figure in published or sent material should trace to a source a person actually opened.
How should I roll this out across a team?
Narrowly, then deepen. Spend two weeks with one capable person on one research-heavy function, building the project instruction file. Spend the next two turning anything done more than three times into a named skill, which is where compounding begins. In week five transfer to a second person on a deliberately different function so you learn where the approach does not carry. In week six write the governance rules covering what runs unattended, what needs approval, and what data may never be shared.
What should I measure to know whether it is working?
Time from brief to approved deliverable, review time per deliverable, and error escape rate. Cycle time falling is the real productivity signal; drafts produced is activity rather than output and often indicates a review bottleneck forming. Review time per deliverable should fall as the instruction file improves, and if output triples while review time triples too, you moved the constraint rather than removing it. Error escape rate should be zero, and a non-zero figure indicates the verification rule is not being enforced.
Yananai A. Chiwuta
Author
Yananai Chiwuta is a Go-To-Market Architect and founder of Forma Nôrden. He builds Go-To-Market growth systems for B2B companies selling into enterprise and mid-market accounts.
Celine Sky
Reviewer
Celine ensures technical accuracy and strategic alignment across all Forma Nôrden playbooks and resources.
