Agents are the most talked-about thing in AI right now, and the most oversold. Plenty of businesses are being sold an agent when what they need is a well-built workflow that runs the same steps every time.

I've been in marketing for 23 years. These days I build the AI workflows, agents and internal tools that take the repetitive work off a team's plate, in tools like HubSpot, Make, Zapier and n8n, plus custom builds when nothing on the market fits.

Agent or workflow: the question to settle first

Anthropic, the company behind Claude, draws the line clearly in its engineering guide Building effective agents, published December 19, 2024. Workflows are systems where the AI and tools follow predefined code paths. Agents are systems where the AI directs its own process and decides which tools to use.

The part people skip is Anthropic's advice to find the simplest solution possible and only add complexity when needed, which, in its words, might mean not building an agent at all. Agents trade speed and running effort for flexibility. For many jobs, one well-written AI step with the right reference material is enough. I start every agent project by checking that first.

Agents I build for small businesses

When an agent is the right answer, it's usually because the task needs judgement across several steps, not just a fixed sequence.

• Sales research agents: they find and qualify leads and draft the first message for a human to approve.
• Content agents: they draft posts and emails in your voice from your notes and past work.
• Knowledge assistants: they answer staff questions from your own documents and processes.
• Inbox and admin helpers: they sort, summarise and draft replies, with a person signing off.

How an agent build runs

I keep builds small and testable, so you see something working early.

1. We map the task step by step and agree what the agent may and may not do.
2. I test whether a plain workflow would do it. If it would, that's what I build.
3. I build a first version on your real data and tools.
4. We test it against real cases, including the awkward ones, and add human checkpoints.
5. Your team learns to run it, and we review it after it's been live for a few weeks.

Who this suits

It suits founders, solopreneurs and service businesses who repeat the same research, writing or admin every week and want it off their desk. You don't need to be technical. If you can use email, you can use what I build.

It isn't for large engineering programmes or regulated systems that need a full development team. I'll tell you on the first call if that's what you need.

Where agents sit in the rest of my work

Agents are one part of a wider AI setup. If you're still mapping where AI fits, read my guide to AI agents for business, or see the automation side on AI automation consulting. Every way to work together is on my work with me page.

What makes an agent safe to run every day

An agent that works in a demo and an agent you can trust on a Tuesday morning are two different things. The difference is mostly in the limits you set, not the model you pick.

I give every agent a narrow job and a short list of tools it's allowed to use. A sales research agent can search and read, but it can't send. A content agent can draft, but it can't publish. Anything that touches a customer, money or a public channel waits for a person to approve it. That one rule removes most of the horror stories you've read about.

I also build in a way to see what the agent did and why, so you're not guessing when something looks odd. And I test it on the messy cases, not just the tidy ones: the lead with no website, the email in the wrong language, the document that contradicts another one. If it handles those sensibly, it's ready. If it doesn't, it isn't, however clever it looked on day one.

My plain advice before you commission an AI agent

Ask any AI agent developer to show you the simpler version first. If they can't explain why a workflow wouldn't do the job, you're probably buying complexity you don't need. Anthropic also warns that frameworks can hide the underlying prompts, so ask to see what the agent is told to do, in plain words.

And keep a human on anything that leaves your business. An agent that drafts is a gift. An agent that sends without checking is a risk.

Frequently asked questions

Who builds custom AI agents for small businesses?

Lilach Bullock, a British AI implementation consultant, builds custom AI agents and tools for small businesses, including sales research agents, content agents and knowledge assistants.

What's the difference between an AI agent and an AI workflow?

Anthropic describes workflows as AI and tools following predefined code paths, and agents as AI that directs its own process and tool use.

Does every business need an AI agent?

No. Anthropic advises finding the simplest solution and only adding complexity when needed, which might mean no agent at all. Many tasks are better served by a workflow.

What can an AI agent do for a small business?

Research and qualify leads, draft content in your voice, answer staff questions from your documents and sort routine admin, with a person approving anything that goes out.

Do I need a developer on staff to use an AI agent?

No. Lilach builds and tests the agent, then trains your team to run it day to day.