Understand AI
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.
For: Buyers and beginners trying to work out what kind of tool a problem actually needs before they shop for one.
- Format
- Decision guide + comparison table + decision sheet
- 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 labels stopped helping
Two years ago the words meant something. An assistant was a chat window. Automation was a rule in your scheduling app. An agent was a research idea. Business software was the thing with the login. Now every vendor uses all four words on the same page, and an owner trying to solve one problem, say quotes that never get chased, cannot tell which kind of thing to buy.
The way out is to stop asking 'which category is this product' and start asking 'what job does this task need done, and by what kind of thing'. The categories still describe distinct jobs. Here is what each one does when it is doing its own job.
The four categories, compared
| AI assistant | Fixed automation | AI agent | Business software | |
|---|---|---|---|---|
| Job it performs | Helps a person think, draft, summarize and answer questions. | Repeats a defined action when a defined trigger happens. | Works toward a goal by choosing its own steps and tools as it goes. | Holds your records and runs your standard processes. |
| Where the business records live | Usually not in the assistant. It reads what you give it or connect to it. | In whichever tool the rule runs in. | Wherever it is connected. An agent without a system of record has nowhere to write. | Here. This is the system of record. |
| Repeatability | Low. Each answer can differ. | Very high. Same input, same action. | Medium. It adapts, which is the point and the risk. | High for the processes it was built for. |
| Human authority | Total. The person decides what to do with the output. | Set once, in the rule. Nobody looks again unless it breaks. | Depends entirely on the permissions you set. | Built into roles and approvals, if configured. |
| Maintenance | Low. Prompts and habits. | Low until the tools change, then rules break silently. | Highest. Goals, tools, permissions and checks all need tending. | Moderate. Configuration, updates, training. |
| Best fit | Messy input, judgment needed, one-off or varied work. | Clear trigger, clear action, no judgment. | Multi-step work where the path varies and the goal is clear. | Standard processes many people run the same way. |
Anthropic's engineering guidance draws the line between the middle two columns: a workflow follows steps you laid out in advance, while an agent directs its own process and tool use. The same guidance recommends the simplest design that does the job [S11]. That is the whole argument of this page in one sentence.
Where they overlap, and why that matters
- An assistant with connected tools can act. General assistants can be connected to live data and controlled tools through official plugin and connector systems [S17]. So 'assistants can only write' is not true. The real question is whether you want the assistant acting, under what permissions, and where it writes the result.
- Ordinary software can automate. Your calendar app already sends reminders. Your accounting package already chases invoices. Before buying an automation layer, check what the tools you own can do with a setting turned on [S10].
- An agent needs a system of record. An agent that books an appointment has to write it somewhere your team will see. Without business software underneath, the agent's work evaporates into a chat log.
- AI is not always needed. If the trigger is clear and the action is the same every time, a rule is cheaper, more predictable and easier to check than a model. Reserve AI for the parts where the input is messy or a judgment is required.
When is an AI assistant enough?
A general AI assistant may be enough when it can reach the tools the job needs and complete the task with the permissions and oversight you want. ChatGPT, Claude and Gemini are the familiar examples. What each can do depends on the configured product, the permissions you grant it and the actions it supports: connections to live data and controlled tools [S17], and in some products scheduled, recurring or event-triggered tasks, which the vendors document with plan and account conditions [S35] [S36]. Nothing on this page assumes an assistant can only draft.
For repeatable business work, four checks decide it. They are the same checks a connected assistant, a simple automation or a business operating system would have to pass, so run them before choosing between those, not after.
- Where do the records live? If the assistant can read and write the system your team already works from, its result can land where the work happens. That is necessary, and it is not proof the task is done: the record shows that a message went out or a field changed, not that the customer replied or the money arrived. If the result lives only in a chat window, someone still has to carry it across.
- Who owns the next step? Work that has to move on a schedule, or when something happens, needs a named owner and a mechanism that runs it. Some assistant products can run recurring or event-triggered tasks when the plan, the app and the connected account support it [S35] [S36]; others act only when asked. Check what the product you actually have offers for this job, and decide who notices when a run does not happen.
- What happens after an error? A wrong reply, a failed connection, a customer who answers unexpectedly. If the recovery is 'the owner notices', the assistant is enough only for work the owner is already watching.
- How is completion verified? Decide what proves the task is done (the reply, the payment, the booking on the calendar) and who checks it. If the answer is 'nobody, we trust it', keep the assistant in an advisory role for that task.
Four honest outcomes come out of those checks: the software you already own, configured properly; a connected assistant used within the permissions you set; a fixed automation for the repeatable part; or a business operating system where the records, the ownership and the checks need to sit in one place. The decision sheet below records which one you chose and why. Where the assistant is enough, use it; a vendor page, this one included, is not evidence that it is not.
Four questions that sort most decisions
- 01What is the job?Name it as a verb and an outcome: 'chase unanswered quotes until the customer replies or the quote expires'. If you cannot state the outcome, no category will help yet.
- 02Who holds the record today?If the answer is 'a spreadsheet and Dana's memory', the first need is business software or a proper record, not an agent. Tools act on records; they cannot act on memory.
- 03How repeatable is it?Same trigger, same action, every time: fixed automation. Same goal, different path each time: an agent or a person. Varied, judgment-heavy, one-off: an assistant helping a person.
- 04Who must approve the result?If a person must approve each outcome (money, commitments, complaints), an assistant that drafts or an agent that proposes is the ceiling. Use the approval boundaries worksheet to write that down.
Four worked decisions
Illustrative example · fictional
Decision one: appointment reminders at a physiotherapy clinic
A fictional clinic sends the same reminder text two days before every appointment. Trigger is clear, action is identical, no judgment. The record already lives in the booking system. Verdict: fixed automation, and the booking software they already own has the setting. Turning it on took ten minutes. No AI, no new tool. This is the decision vendors rarely help you reach.
Illustrative example · fictional
Decision two: replying to messy inbound enquiries at a landscaping company
Enquiries arrive as photos, voice notes and half-sentences. Each needs a different reply and often a question back. The owner wants faster first replies but wants to read them before they go. Verdict: an AI assistant, assistance-only. It drafts the reply and a list of missing details; the owner sends. The customer record still lives in the CRM. Automating the send would remove the judgment that makes the replies good.
Illustrative example · fictional
Decision three: chasing unanswered estimates at an electrical contractor
The goal is clear (a decision from the customer) but the path varies: some reply and ask questions, some go quiet, some approve by text at 9 p.m. The work spans the CRM, messaging and the calendar. Verdict: an agent, inside business software, with a narrow approved scope. It may send the standard follow-ups and log replies; it may not change prices, promise dates or send a third message without a person's yes. The estimate follow-up workflow shows what that scope looks like on paper.
Illustrative example · fictional
Decision four: a bookkeeping firm with no shared client record
Two partners track thirty clients in personal spreadsheets and email. They want 'an AI to keep everything organized'. The four questions stop at the second one: nobody holds the record. Verdict: business software first. A shared client record with statuses and dates. Once that exists, an assistant can summarize it and automation can remind on it. Buying an agent first would have given them a very capable tool with nothing to act on.
Use the decision sheet
One sheet per task. If the 'category chosen' line is 'what we already have', that is a good result. Print it or copy it and keep it with the task.
Worksheet
Decision sheet
Fill this in before you look at any product page.
- The job
- Verb and outcome, e.g. 'chase quotes until answered or expired'
- Who holds the record today
- Which tool, or which person's memory
- How repeatable
- Same every time / same goal, different path / varied and judgment-heavy
- Who must approve the result
- Nobody / a named person per item / a named person for exceptions
- Who maintains it
- The person who fixes it when a tool changes
- Is an assistant enough?
- Records reachable / next step owned / error recovery / completion check: yes or no for each
- Category chosen
- Assistant / fixed automation / agent / business software / what we already own
- Why
- One or two sentences that someone else could challenge
Questions people ask
- When is ordinary automation better than an AI agent?
- When the trigger is clear, the action is the same every time and no judgment is needed. Rules are cheaper, predictable and easy to check. An agent earns its cost only when the path to the goal genuinely varies.
- Can an AI assistant actually do things, or only write things?
- It can do things if it is connected to tools and given permission [S17]. Whether it should is your decision, and the approval boundaries guide is where to make it.
- I already own business software. Do I need anything else?
- Often not. Check what your existing tools automate with a setting first [S10]. Add an assistant where the input is messy, and consider an agent only for multi-step work with a clear goal and a varied path.
- What should I automate first?
- Use the what to automate first guide to rank recurring work by frequency, clarity and consequence, then run the four questions here on the top item.
Related resources
- What should I automate first in a small business?Compare recurring work on six practical measures, sort each task into the right lane, and pick one bounded experiment with a baseline.
- What should AI be allowed to do without asking?Decide, action by action, what an AI assistant or automation may do on its own, what it may do inside an approved scope, and what always waits for a person.
- 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.
- AI models and tools: how to choose for the jobKnow what you are comparing across OpenAI, Claude, Gemini, Grok, Llama and Perplexity, and pick a tool for one business job with a trial plan, not a ranking.
- BlogShould You Hire an Office Manager or Use AI? A Decision Guide for Trades BusinessesHire a person when the business needs judgment, relationships, ownership and capacity across unpredictable work. Use software or AI when the work is repeatable, the record is clear and the result can be checked. If the process is not defined, fix that before doing either.
- BlogAn answer is not an outcome: what has to be true before AI can finish work for youAssistants now answer well and act inside limits. The distance between a good answer and finished work is made of records, authority, checks and follow-up, and most of that is on the business, not the model.
Sources
Dates are when each source was last checked by the editor. Sources support specific claims; they are not endorsements.
- S10How to complete a digital transformation in the age of AI · BDC · checked September 9, 2026Supports the position that proven core software is a valid, often sufficient, starting point and that not every task needs AI.
- S11Building Effective AI Agents · Anthropic · checked September 9, 2026Source for the distinction between predefined workflows and agents that direct their own process and tools, and for the advice to use the simplest design that does the job.
- S17Plugins · OpenAI · checked September 9, 2026Official documentation showing that general assistants can be connected to live data and controlled tools. Used to avoid the false claim that assistants can only draft. It does not describe any Dryvn capability.
- S35Scheduled tasks in ChatGPT · OpenAI Help Center · checked September 9, 2026States that ChatGPT can run one-time or recurring tasks and, for eligible paid plans with a connected Gmail, Slack or GitHub account, event-triggered tasks; availability depends on account, app and app version. Used only to show that scheduled and triggered work exists in an assistant product under stated conditions.
- S36Plan ahead with scheduled actions in the Gemini app · Google (The Keyword blog) · checked September 9, 2026Announcement (6 June 2025) that the Gemini app can run scheduled and recurring actions for Google AI Pro and Ultra subscribers and qualifying Workspace accounts. Vendor announcement, cited for the existence and plan conditions of the feature, not for how well it works.
R11 · 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).