FOXNUT

Comparison

AI agent, workflow or automation: which should you buy?

An automation follows rules, an AI workflow uses a model inside fixed steps, and an AI agent picks its own steps. Buy the most predictable that works.

By the Foxnut team · Updated

What is the difference between an AI agent, a workflow and an automation?

The difference between an AI agent, an AI workflow and an automation is where the next-step decision lives. They are not sizes of the same thing. In an automation, a rule decides every step in advance: when X happens, do Y, identically every time. In an AI workflow, the path is fixed but individual steps use a model to handle input a rule cannot, such as reading a document, drafting a reply or classifying a message, and the output flows back into the fixed path.

In an agent, the model itself chooses the next step. It is given a goal and a set of tools, and decides at run time what to do, in what order, and when it is finished.

We are Foxnut Studios, a senior-only practice in Bengaluru and Paris, and we build these systems for clients, so we are a seller in this category. Our defence is that no row below favours the most expensive option. There’s a link to talk to us at the end. Let’s first work out which of the three you are actually buying.

How do agents, workflows and automations compare?

They compare on the things you pay for later: who decides, what it costs to run, how it fails, and what it takes to test and hand over. The vocabulary matters less than these rows, so buy the row behaviour and not the label.

AxisAutomationAI workflowAI agent
Who decides the next stepA rule, written in advanceThe path is fixed; a model works inside stepsThe model, at run time
PredictabilityIdentical every runBounded: same path, variable step outputLowest: the path itself varies between runs
What you authorise up frontExactly what the rule doesThe credentials its named steps needEvery tool any possible path might need
Running cost shapeNear-zero and flatPer-step model calls; roughly linear with volumeOpen-ended: the system decides how much model to use
How it failsVisibly - a rule mismatch breaks the runQuietly inside a step, while the path completesQuietly and structurally - a plausible wrong path, finished with confidence
Testing itOrdinary software testsAn eval set per model stepAn eval set plus run-level review; the hardest to test
Handing it overA runbookPrompts, evals and a runbookAll of that, plus judgment about when not to trust it

Read the table as a ladder of what it costs to be wrong. Each step to the right adds capability and subtracts predictability, and you pay for the added capability in testing and oversight, not in the licence fee.

When should you choose each one?

Each category wins real situations. The failure modes below are the ones practitioners report, not hypotheticals.

When should you choose an automation?

Choose an automation when a person can write the rule and the inputs do not drift. It is the correct buy and the cheapest thing in this comparison to own. Most of the reliable value in most businesses still lives here, which nobody selling the shinier categories is paid to say.

Failure mode: brittleness at the edges. The rule silently stops matching reality, and nobody was assigned to notice.

When should you choose an AI workflow?

Choose an AI workflow when the steps need judgment but the path does not. Most briefs that arrive saying “agent” end up here: the process is known and repeatable, and what was missing was a way to handle fuzzy input inside it, the kind of somewhat-repeatable judgment work that used to need a person. Practitioners report their plainest wins here, such as a measured “11 times faster” on a defined process.

Failure mode: silent decline inside a step. The path completes, the output is worse, and only an eval set can tell you. This is why the eval set, and not the prompt file, is the artefact to insist on.

When should you choose an AI agent?

Choose an agent when the path genuinely cannot be fixed, because the situations are too varied to list and the system must decide what to look at next based on what it just found. It is the honest fit there, and the only category that can do the job. It is also the minority case: practitioners who build these systems for a living report that most briefs asking for agents do not need them.

Failure mode: compounding. You inherit an open-ended cost shape, unpredictable paths and the hardest testing burden on day one, whether or not the problem required them.

What do you have to authorise for an AI agent?

You have to authorise every tool any possible path might need, which is more than the agent uses on any one run. The third row of the table carries this distinction, and it decides who has to sign. An automation’s authority can be listed in advance and matches its behaviour: it writes the one field or sends the one message its rule names, so the specification and the permission request are the same document.

An AI workflow widens that a little. The model reads and writes inside named steps, and the credentials it holds are the ones those steps need, which is still a list somebody can review line by line.

An agent breaks that match. Because the path is chosen at run time, the tool set has to cover every path the system might take, so what you authorise is the union of everything it could do. The specification stops describing behaviour and starts describing a boundary.

Two things follow, and neither shows up in a demo. First, an agent is a security review where the other two are a change request, and the reviewer has to reason about that union and not about the one run they were shown.

Second, the question to put to a seller is what the system is permitted to do that nobody is planning for. If the answer is a short, fixed list, the seller has built a workflow and called it an agent. That is the most common thing in this market, and not the worst outcome for a buyer.

What does this comparison usually get wrong?

It is argued as a maturity ladder, with the agent as the destination and everything else as a stepping stone. That framing is a sales artefact. The categories are tools with different failure economics, and “which is most advanced” is a vendor’s question where “where does a wrong decision cost me least” is the buyer’s.

The vocabulary makes this worse. The adjectives sellers attach to “agent” have no settled definition, and buyers in practitioner discussions now read them as a warning sign and not a credential, to the point where the word itself has become a purchase blocker.

So read the table as a direction: start at the most predictable category that solves the problem, and step right only when a measured limitation forces it. Whichever category is built, insist on the artefact list a complete AI handover carries, because the handover burden in the table’s last row is where the territory on what a consultant leaves behind says the real price of each category shows up.

What has Foxnut Studios built in each category?

We have built and shipped two agents, and our record runs in both directions. We state it precisely, because a shipped agent is easy to file under the client’s industry and not under the work that was done.

The first is a customer-support agent, built for a SaaS company whose product is accounting software sold to small businesses in California. The accounting is what the client sells, and the engagement was support conversations.

The second is a project-manager “whip” agent for a design agency. It turns meeting discussions into actionables, breaks them into tasks, assigns them to the right projects and tracks them until they are actually done. The subject there is one agency’s own coordination.

Neither is a finance-process build and neither is a supply-chain build. Reading them as either is reading a client’s industry as our experience, which is how a proof claim gets overstated without anyone intending it.

The record runs the other way too, and we have recommended the least fashionable category of all when that was the honest answer. A small industrial machinery manufacturer in Texas asked for AI to track working capital, and we told them not to use AI at all: a one-time-license software met every requirement at a fraction of the cost.

That is the actual test, applied three times in three directions: which category the problem in front of you calls for, whether that is an agent, a workflow, an off-the-shelf automation or nothing built at all.

We are a one-tier practice, so the person telling you that your brief needs a rule and not an agent is the same person who would have built the agent. We launched our own products before we advised anyone on theirs. A free audit call comes first, with no pitch deck, and our consulting page covers the rest of what we do.

If you are not sure which of these your situation needs, work that out before committing to any of them. Tell us what you are trying to solve and we will say plainly which category it actually is.

The part most pages leave out

When not to choose Foxnut Studios

Situations where another option is the better call, and where we say so in the first conversation rather than the fourth.

  • You have chosen the category and want a product: which vendor's platform to license. Your own team, trialling two or three products on a real workflow for a week. Tool selection is not judgment work worth consulting rates, and a studio that took it on would be selling you its reading list.
  • The work is high-volume integration across legacy systems: classic large-scale enterprise automation. An RPA or systems-integration firm with a delivery bench. That is a mature category with real specialists, and a two-person judgment studio is the wrong shape for it.
  • You want to explore what is technically possible at the frontier, without a named workflow or a measurable outcome attached. Your own engineers, a research partner, or a lab. Exploration is legitimate work, but it is not an operations engagement, and billing it like one is how pilots that were never going to ship get funded.
  • Procurement requires named client case studies in your category before anyone signs. A firm with a public client list. Foxnut Studios has no publishable named client case study yet, and we would sooner say so here than have you discover it in procurement.

Foxnut Studios takes on work like this from Bengaluru and Paris. If you want to see how we do it before you talk to anyone, here is what an AI engagement covers and what you keep.