Satisfi Labs has launched 2,500 AI agents for 878 clients over the past decade, and at one point automated its own operations until token costs exceeded what hiring humans would have cost.
Dan Flores, Head of Tourism at Satisfi Labs, walks tour operators through what it takes to treat each AI agent like a team member instead of a feature on a website. Satisfi runs every agent against the same OKR framework it uses for its human team: define the role, write the success criteria, assign an owner, and review the agent’s performance on a cadence. Dan and Aubrie unpack what that looks like in practice for a tour operator, including how to build an AI strategic plan by department, how to segment agents by audience (group sales versus a couple booking a food tour versus a private charter inquiry), and why internal agents and external agents belong in different risk categories.
The session also gets concrete on the operating economics. Dan covers build versus buy filters, the vendor red flags to walk away from, the typical $600 to a few thousand per month price band for agent subscriptions, and the math that makes a $600 a month customer service agent compete favorably with a $20,000 to $60,000 part-time hire. He closes with a starter exercise (list your top five highest-volume workflows, pick one, assign an owner) and a take on model choice for operators chasing weekly updates across Claude, ChatGPT, Gemini, and Codex.
Resource offered to attendees: Dan extended a Tourpreneur-specific offer for anyone who books a Satisfi Labs demo. Mention Tourpreneur during outreach and start a chat at satisfilabs.com to ask company questions or schedule a demo directly.
Takeaways
- AI as workforce, not toolset, is the central mindset shift. [06:00 to 08:15] Dan frames the gap between “AI helps me write emails” and “AI is a team member that owns work for me.” Treating an agent like a tool produces vague experiments. Treating it like an employee forces you to define the role, write the job description, and measure the output the same way you would for any new hire.
- Every agent needs an owner, defined criteria, measurable outputs, and accountability. [06:50 to 08:15] Outcome-driven agents have the same four parts as a real job: a person who owns them, success criteria written down, outputs you can measure, and someone holding them accountable. Without those four parts, you’re paying for a chatbot that “answers questions,” which is not a business outcome.
- Use an OKR or KPI framework to set agent goals. [08:33 to 10:00] For each agent, write the objective (capture more leads for sales conversion, for example), the key results that prove it (conversion rate, qualified leads, response time), and the task the agent performs. The framework you already use for your human team members works on agents without modification.
- Run check-ins with your agents on the same cadence you run check-ins with people. [10:51 to 11:45] Dan talks to his agents through an internal console and asks “how are you doing this week?” The interaction surfaces drift, broken integrations, and missed KRs the same way a one-on-one surfaces issues with a team member. Same job, same review cadence.
- Start with an AI strategic plan that maps pain points department by department. [11:59 to 13:20] Even a three-person company should run the exercise: list every department (sales, marketing, ops, customer service), list pain points in each, and put them on a priority matrix of short-term wins and long-term plays. The output tells you where to deploy agents first and why.
- AI democratizes scale, which used to require capital and time. [13:30 to 15:20] Dan’s framing: a three-person company can now act like a twenty-person company without raising money or waiting years to grow headcount. For tour operators stuck doing everything themselves because hiring full time was never financially possible, the unlock is real.
- Segment your agents by audience, not by topic. [15:31 to 17:30] A group sales lead, a couple booking a food tour, a private charter inquiry, and a travel adviser checking commission terms are four different audiences with four different tones and four different success outcomes. One big agent trying to handle all of them produces mediocre answers across the board. Specialized agents win.
- Treat internal and external agents as separate risk categories. [18:00 to 19:00] Dan’s six or seven personal Claude agents fail three to four times a week and the only person affected is Dan. An external customer-facing agent that fails on a private charter inquiry costs the company a high-margin booking. Risk profile, testing rigor, and monitoring cadence should reflect that gap.
- “I can build it” does not mean “I should build it.” [19:30 to 20:45] Vibe coding, Lovable, and Claude Code make it easy to ship something on your own. The build versus buy decision still has to weigh total cost of ownership: build hours, ongoing maintenance, internal engineering time, data sensitivity, and how quickly you actually need the thing live.
- The vendor market is so noisy that filtering is now a skill. [20:58 to 22:00] Dan cites 10,000 to 12,000 new US business licenses per month for companies with “AI” in the title. Red flags to walk away from: vendor can’t show you data flow, gives vague compliance answers, has no analytics dashboard, can’t produce a ROI case study from a comparable customer.
- Just because you can automate it does not mean you should. [23:22 to 24:55] Satisfi Labs, an AI company, automated so much internally that LLM token costs exceeded what hiring humans to do the same work would have cost. Dan’s filter for what belongs in an agent: high volume, low variance, clear success criteria, speed and scale that matter. If a task happens once a month and takes ten minutes, leave it manual.
- Revisit your AI strategy on the same cadence you review the rest of the business. [26:55 to 28:30] Departments that move quickly (sales, marketing) get evaluated more often. Operations may need less. Match what you do for the rest of the business (sprints, monthly reviews, quarterly cadence) to what you do for your agents. Less frequent reviews mean bigger course corrections later.
- Good vendors hand you ROI numbers, not just usage stats. [28:00 to 31:32] Satisfi’s client console reports metrics like staff time saved, interactions per period, customer service resolution rate, and conversion rate on the ticketing agent. If a vendor only reports raw usage and not outcomes, you have no way to defend the spend internally.
- Subscription pricing for agents typically runs $600 to a few thousand per month. [42:38 to 45:30] Cost depends on which agent (customer service, ticketing, charter qualifier) and how many campaigns or skills you bolt on. For a solo operator running customer service alone, $600 a month can replace what a $20,000 to $60,000 part-time hire would cost in many markets, and that math changes how viable the agent looks.
- Set a baseline before you measure ROI, and be willing to walk away from the deal. [44:06 to 45:30] Dan asks prospects for traffic numbers, customer volume, and the shape of their current customer service operation. Sometimes the honest answer is “you don’t have enough volume to justify this; build it yourself.” That moment is the test for whether the vendor is selling or solving.
- Lay a brick every day instead of trying to build the whole wall at once. [46:09 to 47:05] Dan’s Nipsey Hussle frame: AI adoption is closer to a sustainability journey than a launch event. Operators who try to solve everything on day one stall. Operators who pick one workflow and ship something modest within a week build momentum and learn the trade-offs as they go.
- Concrete starter exercise: list your top five highest-volume workflows, then pick one. [47:20 to 48:50] Rate each workflow internal or external, assign an agent owner, and choose the narrowest entry point. For most operators that’s the overnight customer service inbox or the FAQ traffic on the website. Start narrow, then expand as the data comes in.
- Agentic means the agent acts without being prompted; that is the bar to cross. [49:22 to 50:15] Interactive AI waits for you to type. An agentic system performs a task on your behalf and hands you the output. Dan’s working definition: can this produce the outcome without you sitting in front of it? If yes, it qualifies. If no, it’s still a chat tool.
- A Slack or email summary agent is the easiest first agentic project. [50:44 to 53:00] Dan built one for himself: pulled every Slack channel he was tagged in, generated a daily visibility and action summary, and put it in one message. The agent fails a few times a week and takes a few minutes to fix. It saved him so much time it justified itself within four days.
- Pick one model, get comfortable, and ignore the weekly leapfrog. [57:32 to 59:18] Claude, ChatGPT, Gemini, and Codex are all shipping major updates monthly. Dan’s advice for operators: pick one (he says Claude is the easiest entry today), learn its modes and versions, and trust that whatever shiny feature drops elsewhere will land in your tool within a release or two. Running your business does not leave time to chase every model.