You’re Not Cheating, You’re Evolving: How to Start Experimenting With AI
Eliott Ardisson
Founder & CEO - Basalt Studio
A practical guide for founder-led SMBs on how to start experimenting with AI agents — from your first single-step automation to supervised autonomy, with no hype.
Key Takeaways
- The biggest barrier to AI adoption in small businesses is psychological, not technical — the fear of “doing it wrong” keeps teams stuck in manual workflows longer than necessary.
- Starting with single-step, human-supervised automation is the lowest-risk path to building real confidence with AI tools.
- AI agents handle context and exceptions in ways that rigid rule-based automation cannot — this matters most in messy, real-world workflows.
- A phased approach (augment one step, then supervise, then extend) produces durable adoption rather than abandoned pilots.
- Not every task should be automated. Knowing where to keep humans in the loop is as important as knowing where to remove them.
The Real Barrier Isn’t the Technology
If you’ve hesitated to experiment with AI in your business, you’re probably not alone — and the reason usually isn’t technical. Most founders and operators who delay AI adoption aren’t doing so because they lack budget or engineering talent. They’re doing so because something feels off about it. Like using AI to draft a proposal, or to triage a client inbox, is somehow cutting corners.
It isn’t. Using better tools to do your work more effectively is what every professional generation has done. Spreadsheets replaced ledger books. Email replaced fax. CRMs replaced Rolodexes. None of those transitions meant you stopped thinking — they meant you redirected your thinking toward higher-value problems.
AI agents are the current version of that shift. The question isn’t whether to engage with them. The question is where to start and how to do it without breaking things.
This post is a practical walkthrough for founders and their teams who want to move from curiosity to working implementations, without overcommitting before they understand what they’re dealing with.
What Makes AI Agents Different From Standard Automation
Before getting into implementation, it’s worth being precise about terms — because “automation” and “AI agents” get used interchangeably in ways that cause real confusion.
Traditional automation works on explicit rules. If an email contains the word “invoice,” move it to the billing folder. If a form is submitted, trigger a confirmation email. These tools are reliable when the inputs are clean and predictable. They break when reality doesn’t cooperate — which, in most SMB workflows, is frequently.
AI agents use language models and contextual reasoning to interpret inputs that aren’t clean or predictable. They can read an email and infer that it’s an urgent complaint, not a sales inquiry, even if the sender didn’t use any of your trigger keywords. They can extract information from a messy PDF that would confuse a standard parser. They can evaluate a lead based on nuanced criteria rather than a checkbox list.
The practical difference matters most in workflows that involve human judgment calls — which is most of what founder-led businesses actually do. A recruitment agency doesn’t just need to know that a CV arrived. It needs to know whether the candidate looks relevant. An HVAC contractor doesn’t just need to log a service request. It needs to know if it’s urgent. AI agents can make those first-pass assessments in ways that traditional automation simply cannot.
Level One: Start With One Step, Not One System
The most common mistake teams make when experimenting with AI is trying to automate a full workflow before they’ve successfully automated a single step. The result is a complex, brittle system that fails in unpredictable ways, and a team that concludes AI isn’t ready for their business.
The better frame is this: find the single most tedious, high-volume step in an existing workflow and replace just that step with an AI-assisted version. Everything else stays the same.
Here’s what that looks like in practice:
-
A legal services firm manually reads and categorizes client intake emails before assigning them to a fee earner. An AI agent reads the incoming emails, outputs a category and a suggested priority level, and routes them to a review queue. A human confirms before anything moves. Processing time drops significantly; error rates drop further.
-
A real estate brokerage manually pulls property details from listing PDFs to populate their CRM. An AI agent extracts the structured data, formats it to spec, and queues it for a quick human review. The agent handles the extraction; the human handles the judgment call on whether the listing is worth pursuing.
-
An accounting practice gets questions from clients via email daily. An AI agent classifies each query by type (tax, payroll, advisory, administrative) and drafts a suggested response. The accountant reviews, edits if needed, and sends. The AI handles the first draft; the human handles the relationship.
In each case, the AI is doing one bounded thing. The human is still in the loop. The workflow isn’t rebuilt — it’s upgraded at one node.
This is what’s sometimes called “drop-in augmentation,” and it’s the right starting point for almost every team new to AI implementation.
Three Implementation Approaches, Ranked by Risk
Once you’ve identified a candidate step to automate, there are three ways to structure the AI’s role. Each has a different risk profile and a different time commitment.
1. Human-in-the-Loop (Lowest Risk)
The AI produces a draft output — a classification, a summary, a suggested response — and a human reviews and approves before anything is actioned. The AI never acts unilaterally.
This is the right approach for anything customer-facing, anything that touches data integrity, and anything your team doesn’t yet have confidence in the AI handling. Most first implementations should live here for at least a few weeks before progressing.
2. Supervised Autonomy
The AI acts on routine inputs without a human review step, but a monitoring layer flags anything unusual for human attention. You define what “unusual” means: low confidence scores, unexpected input formats, specific keywords, volumes outside normal range.
This is appropriate once the AI has demonstrated consistent performance on your actual data, not just test cases. A good rule of thumb: if the AI is handling a task with a high accuracy rate over a meaningful sample, supervised autonomy is likely safe. If it’s still occasionally producing outputs that need significant correction, stay in human-in-the-loop mode.
3. Full Autonomy With Monitoring
The AI operates end-to-end on a defined task class, with humans reviewing only exceptions and periodic audits. This is only appropriate for high-volume, well-understood, low-stakes task types where the cost of an occasional error is low and the pattern is genuinely stable.
Very few SMB workflows should jump here in the first six months. The value of guardrailed autonomy is real, but so is the risk of a quietly misbehaving agent that no one is watching.
Where Humans Must Stay in the Loop
AI agents are useful precisely because they can handle volume and pattern-matching at a scale humans can’t. But there are categories of work where human judgment isn’t just preferable — it’s non-negotiable.
High-stakes or irreversible decisions. Financial commitments, legal sign-offs, contract terms, hiring decisions — these require contextual judgment that AI can inform but shouldn’t make. The agent can surface the relevant information, flag the risk factors, and draft the documentation. The human makes the call.
Sensitive client interactions. When a client is frustrated, confused, or in a difficult situation, the quality of the human interaction matters enormously to the relationship. An AI can draft a response, but a human should read it before it goes out. Empathy and relational intelligence remain human responsibilities.
Brand-critical communications. External communications that represent your firm’s voice — proposals, thought leadership, key client updates — benefit from AI assistance in drafting but require human review to ensure they actually sound like you.
Novel or ambiguous situations. AI agents perform well on patterns they’ve seen before. When something genuinely unusual happens — a new type of client request, an edge case the system wasn’t designed for, a situation that doesn’t fit any existing category — a human needs to make the call. The agent should be configured to escalate rather than guess.
In our work helping founder-led firms set up intake and triage agents, the most consistent finding is that teams who define their escalation criteria clearly upfront have far fewer problems than teams who try to figure it out after deployment. Know before you build where the agent stops and the human starts.
A Realistic Implementation Timeline
You don’t need a multi-month project to get meaningful results from AI agents. A realistic phased approach looks like this:
Weeks 1–2: Identify and baseline. Pick one workflow step. Document it: what comes in, what goes out, what decisions get made, what errors currently happen, how long it takes. This baseline is what you’ll measure against.
Weeks 2–3: Build and test. Build the AI agent for that single step. Test it against a sample of real historical inputs before connecting it to any live workflow. How does it perform? Where does it fail?
Weeks 3–6: Supervised operation. Run the agent in human-in-the-loop mode on live work. Track performance, collect feedback from the team members using it, and refine the configuration based on real results.
Months 2–3: Assess and expand. If the first step is performing well, identify the next candidate. Begin reducing oversight on the first implementation if the data supports it. Build a second agent. Continue the cycle.
McKinsey and other research organizations have consistently pointed to phased, team-involved rollouts as the implementation model most likely to produce durable adoption — as opposed to top-down deployments that generate resistance and workarounds. This timeline reflects that finding.
Common Pitfalls Worth Knowing About
A few patterns show up repeatedly in failed or stalled AI implementations:
Automating a broken process. AI makes fast what you point it at. If the underlying workflow is poorly designed, you’ll get faster bad outputs. Fix the process logic before automating it.
Deploying without enough examples. AI agents configured with vague instructions and no concrete examples perform inconsistently. Before deploying, collect a set of real inputs and their ideal outputs. The more representative those examples are, the better the agent will perform from day one.
No fallback procedure. When an AI system encounters something it can’t handle — or has a technical outage — your team needs to know how to do the work manually. If they’ve never done it manually, or the process isn’t documented, you have a fragility problem. Maintain the manual fallback.
Skipping change management. If the people using the AI-assisted workflow don’t understand why it’s being introduced, don’t trust it, or feel threatened by it, they’ll route around it. Involve team members in the design process, explain what the agent is and isn’t doing, and be direct about how it changes their role.
What Good Looks Like After Three Months
By the end of a well-executed three-month implementation, a founder-led SMB should have:
- One or two AI-assisted workflow steps running in supervised autonomy with clear escalation procedures
- Measurable improvements in processing time and consistency for those specific tasks
- A team that understands how to work with the agent, what to do when it fails, and how to identify new candidates for automation
- A baseline of performance data that supports decisions about where to go next
Gartner and Forrester research both suggest that organizations achieving early wins from narrowly scoped AI implementations are significantly more likely to expand successfully than those that start with broad transformation initiatives. Scope small, succeed concretely, then build.
Moving Forward
Starting to experiment with AI doesn’t require a large technical team, a six-figure budget, or a complete workflow overhaul. It requires identifying one high-volume, repetitive step in an existing process, building a human-supervised agent for that step, and measuring what changes.
The “cheating” framing is worth releasing entirely. You’re not offloading your judgment — you’re giving yourself more time to apply it where it matters.
If you’re a founder-led business and you’d like to think through where AI agents would have the most impact in your specific workflows, you’re welcome to book a strategy call with Eliott at Basalt Studio. No sales pitch — just a practical conversation about what’s actually worth automating in your context.