We've built AI agents into about a dozen small-to-mid business operations in the last year. Some stuck. Some got quietly turned off after 30 days. The pattern is clear enough that I can say what works and what doesn't.
This is the honest field report. No hype, no demos, no "AI-native business" framing.
What works
1. Review response agents (hospitality, local service, marketplaces)
If you run multiple locations or a high-volume review inbox, this is the highest-ROI AI project we ship. The agent drafts responses in your voice, routes to managers for approval, and escalates anything sensitive. Most businesses move ~80% of their routine review work to the agent within a month.
Why it works: The job is narrow, the stakes per reply are low, the voice is learnable, and the alternative (a human typing identical replies all day) is miserable.
2. First-draft intake agents (advisory, healthcare, legal)
Clients filling out forms is friction. Clients having a short conversation with a chat agent that asks good clarifying questions and produces a clean intake summary is not. We build these on Claude with strict clinical/legal review layers.
Why it works: The conversation feels human-paced, the summary lands on the advisor's desk structured, and the client feels heard.
3. Content production agents (content brands, B2B marketing)
Not "write my blog posts" (that's a recipe for slop). Specifically: agents that draft social variants from a single long piece, generate metadata, write alt text, or produce SEO-targeted subpages from a structured brief.
Why it works: Clear input → clear output. The human still writes the core thing; the agent multiplies it.
4. Sales research agents (B2B with complex buyers)
Agents that, given a company name and a domain, produce a research brief: recent news, probable pain points, relevant case studies to reference, suggested opener. Not autonomous outreach — research, for a human sender.
Why it works: The research was always going to happen. The agent just compresses 40 minutes of tab-flipping into 2 minutes of review.
What doesn't
1. "AI concierge" chatbots on homepages
Almost never earns its keep on a small business site. Customers either know what they want (they click through) or don't (they leave or email). The chatbot just adds a decision.
2. Fully autonomous sales outreach
Every client who tried it reported the same thing: response rates dropped after an initial novelty bump, and the brand took a reputational hit. LLM outreach reads like LLM outreach within the first two sentences. Use AI for the research stage, not the sending stage.
3. AI "assistants" that do everything
Generic assistants trying to do ten things fail at all of them. The agents that work are scoped to one job, with one success metric, and one failure fallback.
4. Internal knowledge base chat
Works great at Fortune 500 scale. Rarely worth it below ~50 employees. Slack search + a well-maintained Notion usually wins on ROI at small scale.
The rule I'd give myself at $500K–$10M revenue
Pick one agent that automates one clearly-painful, high-frequency task. Measure the baseline honestly. Ship a narrow version. Iterate for two weeks. Either it earns its place or you turn it off — both are fine outcomes.
The one thing you should not do is launch 3–5 agents simultaneously. That's the most common failure we see. The team can't evaluate any of them properly, and all of them slowly rot.
One agent. One job. Measurable outcome. That's the shape that works.