How to Start With AI Agents: Pick One Boring Workflow
The best first AI agent owns one repetitive workflow with a clear owner, a source of truth and a human handoff. How to pick it, scope it and measure it.
Working through a similar problem? Tell us what you want to build.
Legacies is an independent digital product brand based in Bucharest, Romania, founded by Horia Stan and Alexandru Tălnaci, and we get one question about AI agents more than any other: where do we start? Our answer is not exciting.
Start with one boring workflow. Pick a repetitive job that already happens every week, has a clear input, a reliable source of truth and an obvious point where a person should take over. Give an agent that one job, keep a human on the consequential step, and measure the time saved before you add anything else.
Quick answer
- Pick one workflow, not an "agent strategy". The best first candidate is a task where someone moves information between tools using the same judgment every time.
- Draw the action boundary early. Reading and drafting are low risk. Sending, paying, deleting or changing customer records need a human approval step.
- Use the simplest design that works. A fixed workflow with one model call often beats an autonomous multi-agent setup.
- Measure the workflow, not the model. Minutes saved, response time and error rate matter. "It sounded convincing" does not.
Why "start small" is the official advice too
The companies building the models are more careful than the marketing around them. Anthropic's engineering guide, Building effective agents, separates workflows (models and tools orchestrated through predefined code paths) from agents (models that direct their own process and tool use). Its advice is to find the simplest solution possible and add complexity only when it clearly improves results. Sometimes that means not building an agent at all.
OpenAI says something similar in A practical guide to building agents: start with a single agent and split into multiple agents only when the logic or the number of tools really requires it. And when OpenAI launched its enterprise agent product, Presence, in July 2026, the pitch was not "replace your team". Each deployment starts with one specific job, gets only the knowledge and system access that job needs, and follows company rules about what needs approval and when a person takes over.
Analysts see the cost of skipping this step. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value and weak risk controls.
The wrong first question
"What could an AI agent do for us?" creates an endless list. Write social posts, answer emails, qualify leads, update the CRM, create invoices, summarize meetings, chase payments. Most teams then try to connect all of it at once.
The better question is: where does a person repeatedly move information from one reliable place to another, using the same judgment each time?
Good first candidates look like this:
- Turn a contact form submission into a short, structured brief before a sales reply.
- Read a support inbox, group similar requests and draft replies that a person approves.
- Pull weekly numbers from existing tools and write the first draft of the operating report.
- Prepare project handover notes from a fixed set of documents.
They are good candidates because there is an input, an expected output, a human owner and a way to notice when the system is wrong.
How to score a workflow before you build anything
We use a simple scorecard. If a workflow fails two or more of these, it is not the right first project.
| Question | Good sign | Warning sign |
|---|---|---|
| How often does it happen? | Daily or weekly | A few times a year |
| Is the input structured? | Forms, emails, a known spreadsheet | Scattered chats and memory |
| Is there a source of truth? | One CRM, one catalogue, one policy doc | Three tools that disagree |
| What happens if it is wrong? | A draft gets fixed before anyone sees it | Money moves or a customer is misinformed |
| Who owns it? | A named person who does it today | "Everyone" or nobody |
| Can you measure it? | Minutes, response time, error count | Only a general feeling |
The "source of truth" row kills the most projects. An agent that reads three systems with three different prices will confidently pick one. That is a data problem, not an AI problem.
A useful agent needs a boundary
The difference between "helpful" and "risky" is the action boundary. Researching a lead and drafting an email is low risk. Sending it, changing a customer record or issuing a refund is not. The interface should make that obvious.
We prefer to grow trust in stages:
- ObserveThe system collects and summarizes information. It cannot change anything.
- RecommendIt proposes the next action and shows the evidence it used.
- PrepareIt fills out the draft, ticket, report or task, ready for a person.
- Act with approvalA person confirms the consequential step with one click.
- Act automaticallyOnly for a narrow workflow that has earned that trust over weeks of reviewed results.
This order gives a team a safe way to learn, instead of a tool that looks impressive for a week and then becomes another tab nobody opens.
Permissions are part of the design
Give the agent the smallest set of permissions the job needs. If it only drafts replies, it does not need permission to send them. If it reads invoices, it does not need write access to the accounting system.
This matters for security. It is the same least-privilege thinking we apply to npm supply chain security. Prompt injection, where hidden instructions inside an email, web page or document steer the model, is the top risk in the OWASP Top 10 for LLM applications. Simon Willison calls the dangerous combination the lethal trifecta: access to private data, exposure to untrusted content and the ability to send data out. A support agent that reads customer emails has the second one by definition. Keep at least one leg of that triangle missing.
Why this matters for small businesses
Small teams do not need an abstract AI transformation. They need fewer handoffs and fewer forgotten follow-ups. That is why the most valuable first workflow is often administrative rather than creative.
We are a two-person studio, and the AI tools that stuck with us do one narrow job well. We wrote about the developer side of this in how AI coding agents are changing small teams.
If the workflow you picked lives in spreadsheets, inboxes and memory, the useful first step is usually a small internal tool that gives it one home. AI can be part of that system when it improves the workflow. It does not need to be the product's whole personality. That is the kind of web app we build at Legacies: a narrow tool around one real process, with clear permissions and a human handoff. Our services page lists custom web apps starting at 3,499 lei (about 665 EUR), with the final price set after a short discussion.
Measure the workflow, not the model
An agent project needs a plain success metric, written down before you build. For example:
- A coordinator spends 45 minutes less every Monday preparing a report.
- Every inbound request gets a complete first response within an hour.
- The team stops losing project context between a call and the task board.
Then track how often a person corrects the output and how long that takes. If fixing the draft takes longer than writing it, the workflow is not ready.
If the only metric is "the agent produced a convincing answer", the project is not ready either. A convincing answer can still be wrong.
Keep a log you can actually read
For every run, store the input, the output, the sources the agent used and what the human did with it. A simple table is enough for a first workflow. When something goes wrong, you want to answer "why did it say that?" in minutes.
If the agent talks to customers, also check the disclosure rules. In the EU, people have to be told when they are interacting with an AI system. We cover that in our guide to EU AI Act transparency rules.
When to do it yourself and when to bring someone in
Do it yourself if the workflow lives inside one tool that already has AI features, the output is only a draft, and a person on your team owns it. Many CRMs and help desks now ship assistants that cover exactly this. They are cheap and easy to switch off.
Bring someone in when the workflow crosses several systems, needs custom permissions, touches customer data or needs an audit trail. Also bring someone in if your data sources disagree, because no agent can fix that for you. If you are not sure where your website and data stand, our free website audit is a quick, no-signup starting point.
The tools that last will be quiet. They take one boring thing off a team's plate, keep a human in control and make Monday morning a little easier. That is enough to be worth building.
Frequently Asked Questions
What is the best first AI agent for a small business?
The best first AI agent handles one repetitive task that already happens every week, such as turning form submissions into structured briefs or drafting replies to common support emails. It should have a clear input, one reliable source of truth, a named owner and a human approval step before anything is sent or changed.
What is the difference between an AI workflow and an AI agent?
An AI workflow follows a path you define in code, with the model filling in specific steps like summarizing or classifying. An AI agent decides its own steps and which tools to use to reach a goal. Anthropic and OpenAI both recommend starting with the simplest option that works, which is often a workflow, and moving to more autonomy only when the task clearly needs it.
Are AI agents safe to connect to email and customer data?
They can be, if you limit what they are allowed to do. The main risk is prompt injection, where instructions hidden in an email or document steer the model. Give the agent only the permissions the job needs, keep a human approval step before anything is sent, paid or deleted, and log every run so you can trace what happened.
How do I measure if an AI agent is worth it?
Pick a plain metric before you build, such as minutes saved per week, time to first response or number of errors caught. Then track how often a person has to correct the output and how long that takes. If corrections take longer than doing the task manually, the workflow needs a narrower scope or better source data.
Why do so many AI agent projects fail?
Most fail because the job was never clearly defined, the data sources disagree, or the agent got too much autonomy too early. Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027 for these reasons. One narrow workflow with a clear success metric avoids most of them.
Do I need a custom-built agent or can I use existing tools?
If the workflow lives inside one tool that already offers AI features, such as your help desk or CRM, start there because it is cheaper and easy to undo. A custom tool makes sense when the workflow crosses several systems, needs specific permissions or an audit trail.