How to Build AI Agents for a Small Business: Which to Build First

Picture a window and door manufacturer with €1.2 million in revenue and fourteen employees. The owner sits through a demo and comes away sold. The first agent they commission answers quote requests from the website on its own. In week three, it promises a customer an installation date the fitting crew can’t hit. The project gets switched off, and so does the owner’s trust in anything with the word “agent” attached.

They built things in the wrong order. Spent on an agent that drafts internal documents, the same budget would have bought back hours, produced numbers worth deciding on, and trained the team to fix what the machine hands them. To work out which agents to build, in what order, and when to move on to the next, you need five scored criteria, an autonomy ladder and a 90-day plan.

Building any single agent takes the same four steps. You map the process, test it in chat, configure it and estimate what it will cost. It’s the same method you’d use to build an AI agent that automates your small business’s marketing. This article answers a different question: which agent you build first, which comes next, and which you never build at all.

What do we mean by an AI agent in a small business?

An AI agent is a program built on a language model that handles one specific task with some degree of autonomy. It reads information, uses whatever tools you connect it to (email, spreadsheets, your CRM) and hands back a result. Either a person signs off on that result, or the agent puts it into effect on its own. Every step in the sequence below hinges on that distinction.

A review published in Information Fusion in 2025 separates two families that everyday language lumps under one name. AI agents proper are modular systems built on language models and dedicated to automating one specific task. They get more capable when you connect them to tools, give them good instructions and improve their reasoning. The second family, so-called agentic AI, covers systems where several agents work together, break tasks down on the fly, keep a persistent memory and coordinate themselves. For the first family, the authors list customer support, appointment scheduling and data summarisation as typical uses. They assign research automation, robot coordination and medical decision support to the second ¹.

For a business of 10 to 20 people, the practical takeaway is simple. In year one you build agents that each do one job. Self-organising multi-agent systems can wait until you have three single agents running properly.

The autonomy ladder, from safest to most sensitive

Every agent sits on one of these five rungs. The higher it sits, the more damage a mistake can do.

1. Reads and summarises. It takes emails, reports or transcripts and gives you the gist. If it gets something wrong, you’ll notice as you read.
2. Writes a draft that a person approves. Posts, replies, product descriptions. Nothing goes out without a human sign-off.
3. Sorts and routes. It tags incoming requests and assigns them to the right person. A mistake puts an email in the wrong inbox.
4. Acts on internal systems. It updates the CRM, creates tasks, edits a shared spreadsheet. The mistake stays in-house, but it pollutes your data.
5. Acts on customers or budget. It sends emails, publishes, changes offers, moves money between campaigns. The mistake leaves the building.

The same review lists the problems that come with these systems: hallucinations, brittleness when inputs are unexpected, emergent behaviour, and coordination failures once there’s more than one agent. As countermeasures it points to reasoning-and-acting loops (ReAct) and retrieving information from verified documents (RAG) ¹.

Take a wrong price. At level 2, you fix it in thirty seconds before you hit “publish”. At level 5, a customer reads it and brings it up on the phone later. That’s why the right sequence climbs the ladder one rung at a time, and why every agent starts at least one level below what you think it could handle.

Score the candidates before you choose

List every repetitive task in marketing and sales, then score each one from 1 to 5 on five criteria.

• Frequency. How often it comes round. Daily scores 5, once a month scores 1.
• Written procedure. Is there an SOP or a checklist? Written down and up to date scores 5; living only in someone’s head scores 1.
• Cost of a mistake. This score runs backwards. A mistake that costs nothing scores 5; one that could lose you a customer scores 1.
• Reversibility. Also reversed. If you can undo the mistake in one click it scores 5; if it can’t be undone it scores 1.
• Hours saved per month. More than 15 hours scores 5, under 2 hours scores 1.

The table shows an example for a service business with a small marketing team. Your numbers will be different.

Candidate	Frequency	Procedure	Cost of mistake	Reversibility	Hours	Total
Repurposing articles into social posts (draft)	4	4	5	5	3	21
Weekly marketing and sales report	3	3	4	5	3	18
Routing incoming requests	5	2	3	4	3	17
Internal procedures assistant	3	2	4	5	2	16
Draft replies to reviews	2	3	4	5	1	15
Automatic replies to quote requests	4	2	1	1	4	12

The agent the owner in our opening story wanted first lands at the bottom. It scores well on frequency and hours at stake, but the procedure is weak and a mistake can’t be taken back.

You can also do the scoring with an assistant like Claude or ChatGPT. Paste in your task list and use a prompt along these lines.

You are a process consultant for a [industry] business with [number] employees.
Below is the list of repetitive tasks in our marketing and sales,
with a one-line description of each.

Score each task from 1 to 5 on these criteria:
frequency, whether a written procedure exists, cost of a mistake (5 = no cost),
reversibility of a mistake (5 = undone in one click), hours saved per month.
Also state the minimum autonomy level required (1 reads, 2 approved draft,
3 routes, 4 acts on internal systems, 5 acts on customers or budget).

Return a table sorted by total score. For every score below 3, write
one line on what we should fix in the process before building the agent.
If you're missing information needed to give a score, ask me for it instead of guessing.

[task list]

That last line does a lot of work. Leave it out and the model makes up how often your tasks happen, and the table looks precise even when it isn’t.

The first three agents to build, in this order

1. Content repurposing, in draft

The first agent takes an article you’ve already published, a newsletter or a webinar transcript and turns it into three or four draft posts for LinkedIn and Instagram, in your brand’s voice. A person reads them, edits them and schedules them. That’s level 2 on the ladder: a mistake costs you one edit, and almost every team already carries the procedure in their heads.

The most solid research on how these tools affect skilled work comes from software development, and it’s useful here for the logic it describes. Three randomised experiments at Microsoft, Accenture and an anonymous Fortune 100 company, published in Management Science in 2025, gave an AI assistant to a randomly selected group of developers during their normal work. Across 4,867 people, those with the assistant completed 26.08% more tasks, with a standard error of 10.3%. Less experienced developers adopted it more and gained the most ². An earlier analysis of 1,974 developers at the first two companies had found weekly pull requests rising by between 12.92% and 21.83% at Microsoft and between 7.51% and 8.69% at Accenture, estimates the authors themselves describe as imprecise ³.

A small business can take two lessons from these numbers. First, the gain is real but varies a lot from one setting to another, so you have to measure the hours saved yourself, on your own team. Second, less experienced people benefit most. That makes the junior who runs your social channels the right person to hand the agent to first, with the marketing lead signing off.

The heart of the agent is an instruction like this.

You will receive an article from our blog. The audience is small business owners.
Write 3 draft LinkedIn posts (1,200 characters max) and 2 Instagram captions
(400 characters max).

Rules:
- every post opens with a figure or an example from the article, never a generic question
- use only information contained in the article; if a number isn't there, don't write it
- direct tone, short sentences, no emoji, no more than two hashtags per post
- at the end of each draft, name the paragraph of the article you took it from

Mark every draft as DRAFT. Do not suggest posting times.

Naming the source paragraph is for the approver, who can check it in ten seconds without rereading the whole article.

2. Weekly marketing and sales report

The second agent reads the week’s data (visits, new leads, quotes sent, deals closed) and writes a one-page summary for Monday’s meeting. It sits at level 1 and falls under data summarisation, one of the typical uses for single-task agents.

There are two situations where you should start here and push content repurposing back. The first is when you publish fewer than two pieces a month, so there’s little to repurpose and few hours at stake. The second is when the owner wants to see the numbers before putting money into marketing. A report that lands on time every Monday is more persuasive than ten well-written posts.

3. Routing incoming requests

The third agent reads emails and website form submissions, tags them (quote request, support, supplier, job application, spam), assigns them to the right person and drafts a first reply. That’s level 3. A mistake sends a request to the wrong inbox, so bring this agent in only once the team has learned to correct the first two. Before you build it, write down the routing rules that whoever opens the mail currently applies from memory.

The 90-day plan

Weeks	What happens	Who owns it
1-2	Content repurposing SOP, chat tests on 10 real articles	Marketing lead
3-6	Agent 1 live in draft mode, every correction logged	Social media assistant, with sign-off from the marketing lead
7-8	Review of agent 1’s numbers, then report SOP and tests	Marketing lead and sales
9-11	Agent 2 running every Monday, checked against the manual report	Sales
12-13	Routing rules written down and agent 3 tested in chat	Whoever currently handles incoming mail

One rule governs the handover between agents. You start building the next one only when the current one has passed the check described in the next section. Every agent has a named owner who spends half an hour a week reading the correction log and updating the instructions.

When are you ready for the next agent?

You move on when the current agent has held up on its own for four weeks straight. In practice, that means fewer than one draft in five needs substantial edits, the measured hours saved are close to the estimate in your scoring table, and the owner keeps using it without anyone having to remind them.

The correction log tells you all of this. A spreadsheet with four columns is enough: date, draft, what you changed and why. After a month the reasons start repeating, and every recurring reason becomes a new line in the agent’s instructions.

If the correction rate is still high after six weeks, the problem is usually the procedure. Go back to the SOP before you switch models or tools.

How many agents make sense? For a business of 10 to 20 people, three to five agents in the first year is a number the team can keep on top of. Every extra agent is another set of instructions to maintain and another log to read. If nobody has half an hour a week to look after it, that agent quietly degrades.

Agents to leave until last, or never build

Some tasks show up in every demo and are still poor candidates for a business just getting started.

• Hiring. Decisions about people carry the most legal exposure and the highest risk of discrimination. At most, an agent can summarise CVs for whoever makes the call, and rejecting a candidate stays a human decision.
• Pricing and discounts. A price you’ve communicated is a commercial commitment. It belongs with whoever answers for the margin.
• Running the ad budget on its own. Moving money between campaigns is a level 5 action with an immediate cost. The agent can prepare the proposal; a person approves it.
• Replies published without review. Reviews, comments, messages to customers. Drafts are fine, with someone reading them before they go out.
• Automatic quotes. That’s the window manufacturer’s mistake from the top of this article.

The prerequisite nobody checks is the written procedure

Every guide to AI in business tells you that you need quality data. For a small business, the real prerequisite comes even earlier: a written procedure for the task. An agent follows instructions. If the way your team repurposes an article depends on who’s doing it that day, the agent has nothing to follow and you have no yardstick for judging its drafts.

You can write the SOP in an hour with an assistant’s help. Record a five-minute voice note explaining to a colleague how you do the task, transcribe it and use this prompt.

Below is a transcript of someone explaining how they carry out a task
on our marketing team. Turn it into an operating procedure with:
- the goal of the task in one sentence
- the inputs needed and where to find them
- numbered steps, one per line, each with the criterion for a step done well
- common mistakes to avoid
- who approves the final result

If a step is ambiguous or missing, list it at the end as a question to ask the person.
Do not add steps that aren't in the transcript.

[transcript]

The questions at the end are worth as much as the procedure. They show you where the process currently relies on one person’s memory, which is exactly where the agent would trip up. It’s the work we do with marketing and sales teams in our SOP training, and it pays to do it before you touch any tool.

Tools, level by level

You pick the tool after you’ve chosen the agent and written the procedure.

Stage	Tools	What they’re for
Chat testing (levels 1-2)	Claude Projects, ChatGPT custom GPTs	Testing instructions and SOPs on real cases before automating
Draft agent with connected data	Flowise	Building the agent with documents and tools connected
Connecting email, forms, CRM (levels 3-4)	Make, Zapier, n8n	Feeding inputs to the agent and getting its drafts to where the team works

On Monday morning, step one is a five-minute voice note on how your team repurposes an article. The procedure comes out of that, the first agent comes out of the procedure, and its correction log tells you what the second one should be.

Sources

Footnotes

1. AI Agents vs. Agentic AI: A Conceptual taxonomy, applications and challenges. doi.org ↩ ↩²

2. The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers. doi.org ↩

3. The Productivity Effects of Generative AI: Evidence from a Field Experiment with GitHub Copilot. mit-genai.pubpub.org ↩