Understand AI
What is an AI system, a workflow and a loop?
Learn the difference between a system, a workflow, an automation, an agent and a feedback loop using two everyday business examples, and map one of your own on a worksheet.
For: AI beginners and the people who run day-to-day operations, who keep hearing these words used as if they mean the same thing.
- Format
- Illustrated guide + fill-in workflow worksheet
- Published
- September 9, 2026
- Last checked
- September 9, 2026
- Written by
- Dryvn AI editorial
- Edited by
- Dryvn AI editorial
- Reviewed by
- Not yet reviewed by a named person
- Read as
- Editorial (Dryvn's practical method)
Why the words get mixed up
If you have sat through a software demo in the last two years, you have heard "system", "workflow", "automation", "agent" and "loop" used in the same sentence, often as if they meant the same thing. They don't. Each one describes a different amount of decision-making, and knowing which one you are being sold tells you what it will do when something unusual happens.
This guide defines each word with a business example, not a technical one. If you only remember one thing, remember this: a workflow follows steps, and a feedback loop checks results and changes what happens next.
Five words, five different amounts of judgment
| Word | What it is | Who or what decides the next step | Everyday example |
|---|---|---|---|
| Workflow | A set of steps in a fixed order. Can be done by people, software or both. | Nobody, once it starts. The order is the decision. | New customer form arrives, office creates a record, sends a welcome email, books a call. |
| Automation | A workflow that software runs without a person pressing the buttons. | Nobody. It does the same thing every time. | When an invoice is marked paid, the accounting package emails a receipt. |
| Feedback loop | A workflow that checks the result of a step and repeats, changes or stops based on what it finds. | A rule, a person, or a model, depending on how it is built. | Send a reminder; if no reply in three days, send a different one; if still nothing, hand it to a person. |
| Agent | Software that uses an AI model to choose its own steps and tools toward a goal, rather than following a fixed list [S11]. | The model, within whatever limits it was given. | "Find out why this customer's job is late and draft an update": the software reads the record, checks the schedule, then writes. |
| System | All of it together: people, tools, rules, data, checks and the person accountable for the outcome. | Shared. A well-designed system says who decides what. | Your whole customer-request process, from the first call to the closed file. |
Anthropic's engineering guide draws the line between workflows, where the steps are set out in advance, and agents, where a model decides its own steps as it goes, and it recommends using the simplest design that does the job [S11]. That advice is worth carrying into your business. Most useful business "AI" is a workflow with a model doing one step well, and there is nothing wrong with that.
A workflow: steps in order
Here is a plain workflow for an estimate. There is no AI in it and no loop. Each step follows the last, and when the last step is done, the workflow is finished whether or not the customer ever answers.
Text version of this diagram
- 1. Site visit (measure, photos) → next step
- 2. Write estimate (price, scope, terms) → next step
- 3. Send to customer (email or text) → next step
- 4. Mark as sent (done, as far as the workflow knows)
This is where a lot of businesses actually are. The estimate went out. The workflow considers itself complete. Whether it turned into work depends entirely on someone remembering to ask.
A feedback loop: check the result, then decide
Now add one thing: after sending, check whether the customer replied, and let that answer decide the next action. That single check turns the workflow into a loop.
Text version of this diagram
- 1. Send estimate (or a follow-up)
- 2. Wait (e.g. four working days)
- 3. Check for a reply (the evidence step)
- 4. Decide (reply, decline, silence?)
- Back to step 1: No reply yet? Send a different follow-up, or hand it to a person after the third one.
Notice what makes it a loop: the check. Without a check there is nothing to feed back. The loop also needs a stop condition. It stops when the customer replies, declines, asks not to be contacted, or the estimate is replaced by a new version. A loop with no stop condition is just a nuisance with a schedule.
Illustrative example · fictional
Estimate follow-up at a fictional flooring company
A fictional flooring installer sends about fifteen estimates a week. The owner sets up a simple loop: the day an estimate goes out, a follow-up is scheduled for four working days later. When that day comes, the system checks the estimate record. If the customer has replied, the follow-up is cancelled and the owner sees the reply. If not, a short check-in message goes out and a second follow-up is scheduled a week later. After the second one, the estimate goes onto a list the owner reviews every Friday, and the loop stops on its own.
There is no AI in this loop. It is rules, a calendar and a record of replies. The owner could later add a model to draft the check-in message in a friendlier tone, or to read the customer's reply and flag "they asked about a payment plan". That would be AI assisting one step. The loop itself would be unchanged.
Illustrative example · fictional
A customer request at a fictional accounting practice
A client emails a fictional bookkeeping practice: "Can you send me last year's T4 summary?" The practice's loop looks like this. The request is logged with an owner and a due date. The bookkeeper finds the document and sends it. Two days later, the system checks: did the client confirm they received it, or open the secure link? If yes, the request closes. If not, a gentle nudge goes out. If the client replies with a new question instead, the loop does not close; it creates a new request with the same owner and starts again.
The interesting part is the last branch. A reply is not the same as a resolution. The loop is only allowed to close when the original request has actually been met, and the system knows the difference because someone decided in advance what "met" means.
A failed loop: when "sent" gets treated as "done"
Illustrative example · fictional
The loop that closed itself too early
A fictional property manager sets up an automated reminder to tenants about a scheduled water shut-off. The tool sends the message, records "reminder sent" against each unit, and marks the task complete. On the day, four tenants are surprised and one has a home dialysis appointment that had to be moved. When the manager looks, all forty reminders show as complete. Three of the phone numbers were old. One tenant had replied "this date doesn't work for me" and nobody saw it, because the loop had already closed.
The tool did exactly what it was built to do. The failure was in what counted as evidence. "Sent" is a fact about the sender. "Received and acknowledged" is a fact about the outcome. A loop that stops at the first one is not checking anything; it is a workflow wearing a loop's clothes. The fix is not more automation. It is deciding that the evidence of completion is a reply or a confirmed read, and routing anything else to a person before the day.
Where AI fits, and where it doesn't have to
- A loop does not need AI. Reminders, escalations and reorder points are loops built from rules. They are cheap, predictable and easy to explain to staff.
- AI is useful inside a step that needs judgment. Reading a customer's reply and sorting it into "approved", "question" or "declined" is that kind of step. Writing a message in a natural tone is another.
- An agent is useful when the steps can't be listed in advance. "Work out why this job slipped and draft an update" is a goal, not a list. That is agent territory, and it needs clear limits on what it may do [S11].
- The system is what makes any of it safe. Owners, approvals, exception routes and evidence are system decisions. No tool supplies them for you. The approval boundaries guide covers the limits part.
Map one of your own
Take one recurring piece of work and fill in the nine fields below. If a field is blank, that is the finding. The two most commonly blank fields are the exception route and the evidence of completion, and those are the two that turn a workflow into a loop that actually closes.
Worksheet
Workflow and loop worksheet
One row per piece of work. Fill it in with what happens today, not what you wish happened.
- Event
- What starts this? A call, a form, a date, a payment.
- Intended outcome
- What has to be true for this to be finished?
- Owner
- One named person responsible for the outcome.
- Information
- What the step needs, and where it comes from.
- Action
- What is done: by a person, software or a model.
- Approval
- Which decisions inside this need a person's yes, and whose.
- Response
- What is checked after the action, and when.
- Exception
- What happens when the check fails or the case is unusual.
- Evidence of completion
- What proves the outcome, not the action.
Downloads · no email required
- Download .md
Workflow and loop worksheet
The nine-field worksheet as a blank fill-in template, with instructions and a separate fictional example. Opens in any text editor or notes app.
Questions people ask
- What is the difference between an AI system, a workflow and a loop?
- A workflow is steps in order. A loop is a workflow that checks its own results and decides the next action from them. A system is the whole arrangement of people, tools, rules, checks and accountability around the work. AI can sit inside any of the three, but none of them requires it.
- Can you explain an AI feedback loop using a small-business example?
- A customer request comes in and is logged with an owner. A model drafts a reply for the owner to approve. After it is sent, the loop waits, then checks whether the customer confirmed the issue is resolved. If yes, it closes. If the customer replies with something new, the loop opens a fresh request. If there is silence, it nudges once and then puts the item in front of a person. The AI is doing the drafting and the sorting; the loop is doing the checking.
- Is an agent better than a workflow?
- Not by default. An agent decides its own steps, which is useful when the steps cannot be listed in advance and risky when they can. For most recurring business work the steps are known, and a workflow with a model in one step is easier to trust and explain.
- How do I know if my automation is actually a loop?
- Find the check. If nothing looks at the result of the action and changes what happens next, it is an automation. Then find the stop condition and the exception route. If both exist, you have a loop that can close.
Related resources
- Estimate follow-up: the complete process, with message templatesRun estimate follow-up as a process with an owner, a schedule, stop conditions and a tracker, using five original messages you can copy and adapt.
- Customer requests: track the work until it is actually resolvedRun every customer request through the same nine stages, so nothing is marked done until the customer's problem is gone and you can show it.
- How multi-model AI systems work, and when one model is enoughSee what happens when several AI models share a job, why disagreement is useful, what it costs, and how to write down roles and checks before you build.
- AI assistant, automation, agent or business software: which do you need?Tell the four categories apart by the job each one does, and pick the right one for a specific task in your business, including the case where what you already own is enough.
- BlogWhat’s Your Differentiator? The Question That Made Me Rethink DryvnI am not trying to build another dashboard that gives a business owner more information to monitor. I am building Dryvn toward a harder outcome: reducing the daily effort of coordinating the business, including the effort of operating the software itself.
- BlogThe AI Fabric: A Research Thesis on Loops, Graphs and Everyday CoordinationBrent develops a proposed AI fabric connecting people, context, permissions and ongoing work. The paper examines its research foundations, limits and a method for testing whether it reduces human coordination effort.
Sources
Dates are when each source was last checked by the editor. Sources support specific claims; they are not endorsements.
- S11Building Effective AI Agents · Anthropic · checked September 9, 2026Primary source for the distinction between a workflow (predefined steps) and an agent (a model directing its own steps and tools), and for the advice to use the simplest design that does the job. Used for definitions only, not for any claim about Dryvn.
R02 · Published September 9, 2026 · Next scheduled review December 9, 2026 · Teaches process management; not legal, warranty, safety or engineering advice. Examples are fictional unless stated. Part of the Dryvn resource library (17 resources).