Alp SuleimanGrowth Partner

Which Growth Work Goes to an Agent? Choosing Between a Prompt, a Workflow and an Agent, Stage by Stage

“Let's give this to an agent” sounds like a technical choice. It is really a decision about authority and economics. At each stage of the funnel, which work belongs to a plain prompt, a workflow or an agent?

Growth PartnerArticle 01 / 018 min read

Published

There is a sentence you now hear often in growth meetings: “Let's give this to an agent.” The segment list, the email variants, churn follow-up, even the pricing test. It sounds like a technical preference, as if the question were which tool to use.

The real question is a different one. Handing work to an agent means deciding who you leave the decisions inside that work to, and what you pay for it. “Agent” is sold as a product feature; for a growth team it is a decision about authority and economics. How much autonomy, who measures it, what does each successful task cost, and who notices when it goes wrong?

This series is about making that decision from inside the growth system. The first question is the most basic one: which growth work goes to an agent, and which doesn't?

A workflow is not an agent

Anthropic's “Building effective agents”, published in December 2024, clears up most of the confusion with one distinction. A workflow is a system where the language model and tools are orchestrated through predefined code paths. An agent is a system where the model dynamically directs its own process and tool use, and decides for itself how to get the work done.

The advice in the same piece is just as clear: find the simplest solution possible, and only increase complexity when needed. The authors say outright that this might mean not building agentic systems at all, and add that agentic systems often trade latency and cost for better task performance, so you should consider when that trade makes sense.

Carry that distinction into growth work and you get four levels:

  • Plain prompt. A person asks the model something, reads the answer, uses it or throws it away. The decision stays entirely with the person.
  • Workflow. The steps are written in advance: data comes in, the model classifies it, the result goes to a set place. The model doesn't choose the path.
  • Human-approved agent. The model plans its own steps and uses tools, but a person approves every action that reaches the outside world.
  • Autonomous agent. The model plans, acts and reports back. A person looks afterwards.

These levels are not a maturity ladder. The top rung is not automatically the best one for a given job, and as a rule of thumb, for many jobs the right answer sits on the second.

Five criteria

Five questions will tell you which level a piece of growth work belongs to.

1. Can it be undone? OpenAI's “A practical guide to building agents” names two primary triggers for human intervention: exceeding failure thresholds (the agent fails to understand a customer's intent after several attempts) and high-risk actions. It defines the second as actions that are sensitive, irreversible or high-stakes, which should trigger human oversight until confidence in the agent's reliability grows. Its examples will be familiar to any growth team: cancelling orders, authorising large refunds, making payments. The same guide suggests rating every tool low, medium or high risk based on read-only versus write access, reversibility, the account permissions it needs and its financial impact.

2. Is a mistake expensive, and who notices? A wrong internal report gets corrected in a meeting. A wrong discount code that has reached thousands of customers is a second mistake in the customer's eyes when you correct it. The cost of an error grows with the time it takes to notice.

3. Can it be measured? Can you check automatically that the work was done right? A segment list either matches its definition or it doesn't, and a query will tell you which. “A good campaign idea” can only be judged once the results are in. Autonomy over work you can't check is a way of accumulating mistakes. In AI this automatic check is called an eval, and it is the closest thing to a growth team's experiment culture.

4. How much context does it need, and how sensitive is that context? The OpenAI guide groups the places where agents add value under three headings: complex decisions involving nuanced judgment and exceptions, rulesets that have become hard to maintain, and heavy reliance on unstructured data (natural language, documents, conversation). That tells you where an agent might help. How dangerous the context can become is described by what Simon Willison calls the lethal trifecta: access to your private data, exposure to untrusted content, and the ability to communicate externally. Put the three together and an instruction hidden inside some outside text can persuade the agent to carry your data out. Willison's conclusion is blunt: the only way to stay safe is to avoid that combination entirely. He is also suspicious of guardrail products that claim to catch “95% of attacks”, pointing out that in web application security 95% is very much a failing grade.

5. How long is the task? In March 2025, METR proposed measuring AI agents by task length: the length of task, measured by how long it takes human experts, that a model can complete with 50 percent probability. At the time, models succeeded almost 100 percent of the time on tasks that take humans less than 4 minutes, and less than 10 percent of the time on tasks that take more than about 4 hours. Claude 3.7 Sonnet's 50 percent horizon was about an hour. Yet the same post says the best models could only reliably complete tasks of up to a few minutes. These numbers come from software tasks and the models of that date; today's values are different. The lasting lesson for a growth leader is this: 50 percent is a coin toss. It is not the reliability you would accept on work that reaches a customer. The longer the task and the more steps it has, the earlier human approval should come.

Stage by stage

The jobs below are examples, not taken from any particular company. The levels are a judgement: the five criteria applied to the funnel.

Acquisition. Pulling a target segment list from existing customer data fits a workflow: the definition is clear, a query can check the result, and a wrong list is simply regenerated. Writing ad copy and email variants is a plain prompt or a workflow, but publishing should be human-approved, because published copy speaks for the brand and stays seen even after it's pulled. An agent that raises budget on a new channel by itself can produce a mistake that is both expensive and noticed late; moving budget fits what OpenAI calls high-risk actions (sensitive, irreversible or high-stakes).

Activation. An assistant that answers a user's questions during onboarding from the product documentation is a reasonable agent candidate: the work is full of unstructured language and hard to capture in rules. But its access should stay read-only. The moment it starts changing settings or upgrading plans in the user's account, the level of the job changes.

Retention. Flagging churn risk, putting accounts with falling usage on a list every week, is a workflow. It is measurable, reversible and cheap when wrong: a wrongly flagged account costs at most an unnecessary call. An agent that writes to at-risk customers and sends the messages itself completes Willison's trifecta: it touches customer data, reads customer email (untrusted content) and sends messages out. The sensible level here is a human-approved agent: the agent drafts, a person presses send.

Revenue and expansion. Price changes, discounts, refunds and contract terms don't go to an autonomous agent. They can't be undone, or undoing them costs customer trust; errors are expensive, and the effect shows up months later. An agent can help with the analysis, for instance a report on which cohorts are price-sensitive. The decision and the execution stay with people.

The growth team's own work. Summarising experiment results, preparing the weekly metrics report, gathering data for a hypothesis: this is where agents are most at home. The output stays inside, a person reads it, and a mistake is caught at the next meeting. An agent that starts and stops experiments by itself is a different level altogether: at that point it isn't writing a report, it is changing the product customers see.

Economics: an agent is an expensive hire

Cost rises with autonomy, and Anthropic's data shows by how much. In June 2025, describing its own multi-agent research system, Anthropic reported that in its data agents typically use about 4 times more tokens than chat interactions, and multi-agent systems about 15 times more. The post is plain about the consequence: for economic viability, multi-agent systems require tasks where the value of the task is high enough to pay for the increased performance.

A growth team already knows this arithmetic. You judge a channel by CAC; you should judge an agent by cost per successful task. Failed attempts, the time people spend fixing output, and the delay of work waiting for approval all belong in that cost. If building an agent for a segment list pulled once a week costs more than a workflow and doesn't produce a better list, the agent is a nice feature and a bad investment.

Build your own decision table

What this piece suggests is not a tool but a page. Write every job your team is thinking of handing to AI on its own row and fill in five columns:

  • Job: concrete, with a verb (“every Monday, list accounts whose usage fell 30 percent”).
  • Reversibility: can the result be undone, and will a customer have seen it by then?
  • Cost of error and time to notice: what is lost if it goes wrong, and who notices, when?
  • Measurement: can you check automatically that it was done right?
  • Context and authority: do private data, outside content and external communication meet in this job?

In the last column write a level: plain prompt, workflow, human-approved agent or autonomous agent. One test question tells you the table is done: if this job is done wrong, who notices, when, and can it be undone? If the answer is “a customer, a week later, no”, don't write autonomous agent on that row.

Our expectation: many teams that fill in the table may find most of the list settling on workflows. That shouldn't count as a disappointment. It is exactly what Anthropic's “simplest solution” advice looks like inside a growth team.

Take care.

Sources