Blog · AI
Every vendor is selling AI. Most of what gets sold to small businesses as AI is solving a problem that traditional automation already solved cheaper. Here's the framework we use to decide which workflows deserve AI and which ones don't.
AI pays back when the workflow's input is unstructured (free-text, images, PDFs with weird layouts, voice) and the cost of being wrong is low. AI is the wrong tool — and almost always more expensive — when the input is structured (forms, database records, API payloads) and the rules are stable. About 70% of the 'we need AI' requests we get from small businesses can be solved cheaper and more reliably with traditional automation. The 30% that genuinely benefit from AI usually pay back in 3-9 months. The 70% that don't would have paid back in 2-4 months with non-AI automation and would have cost 60-80% less.
Right now there's a tax on every automation conversation we have with small business owners. Someone — a vendor, a consultant, a LinkedIn thought leader — told them they need AI. They walk in asking for AI before they've described the problem. The conversation now has to walk back through "what are you actually trying to do?" before we can give a real answer.
That's the AI hype tax: the wasted scoping cycles, the inflated budgets, and the disappointment that follows when a $25,000 AI agent does worse than a $5,000 Zapier workflow would have done. We're not anti-AI — we ship AI features. We're anti-spending on AI when traditional automation solves the same problem cheaper, faster, and with better uptime.
Below is the framework we use internally to decide which workflows deserve AI. It takes about 10 minutes and saves you from buying a Ferrari to deliver pizza.
AI earns its premium in problems that traditional code can't solve cleanly. The common thread: the input is too messy or too varied to write rules for, but humans can do it instinctively. That's the AI sweet spot.
Inbound emails or chat messages need to be triaged: is this a billing question, a bug report, a feature request, a sales lead? Writing rules for this is a losing battle — customers don't use your category names, they describe the symptom. An LLM can classify these with 90%+ accuracy in under a second.
Suppliers send invoices in 47 different PDF layouts. Customer agreements come back with redlines as PDFs. Receipts get emailed in. Traditional OCR works for clean invoices but breaks on edge cases. Vision-capable LLMs handle the messy 20% that used to require a human eyeball.
"Take this CRM data + this prior conversation and write a first-draft follow-up email in the voice we use." This is the highest-leverage AI use we see at SMBs. The output is always reviewed by a human, so the cost of being wrong is "you have to edit a paragraph" — a 30-second cost.
"Here's a 4-paragraph free-form lead form. Who should it go to, with what urgency, and what's the suggested first response?" Routing rules can be written for structured inputs (industry dropdown, urgency dropdown). They fall apart when the input is "tell us what you need". AI bridges that gap.
These are the conversations where "we should add AI" is almost always the wrong answer. Traditional automation is cheaper, more reliable, and faster to ship.
Form submission → CRM row → notification email. Zero ambiguity, structured input, clear rules. Adding an LLM here is using a $5/token tool to do what a $0/run if-else statement does. We've seen agencies pitch "AI-powered lead intake" for $15K when the real work is a $4K Zapier or custom integration.
"Calculate the commission based on these tiers." "Apply the discount if order ≥ $500 AND customer is VIP." "Reconcile the payment against the invoice." All of these have crisp rules. Code does this perfectly, deterministically, and traceably. LLMs do it non-deterministically with hallucination risk. The audit trail of "the LLM decided to apply a 12% discount because…" is a compliance nightmare.
"When order ships, send the customer a tracking email." "When a deal moves to closed-won, notify Slack and create the invoice." These workflows are predictable, structured, and high-volume. Every dollar spent on AI here is wasted. The cheapest version is a template + a workflow tool ($0-$50/month all-in).
Financial reporting amounts. Medical record fields. Legal document fields. Anything where being slightly wrong has regulatory or financial penalty exposure. LLMs are probabilistic; compliance requires determinism. If your auditor asks "why did this number change?", "the AI decided" is a bad answer. Use AI for the messy upstream work (extracting fields from PDFs), then use deterministic code for anything that ends up in a regulated system.
From what an automation project actually costs in 2026, here's the AI-specific delta over a base automation project. These are the ranges we quote at Kivolaro.
| AI tier | What it looks like | Build adder | Monthly tokens |
|---|---|---|---|
| None | Pure automation — rules, integrations, no LLM. | $0 | $0 |
| Light | Classification, drafting, parsing — single-step LLM calls with structured outputs. | +$1,500 | +$80 |
| Heavy | Multi-step reasoning, agents that take actions, RAG over your docs, vision on PDFs. | +$5,000 | +$350 |
These are SMB-scale numbers (1–50 employees, 200–5,000 LLM calls/month). At enterprise scale the math changes — token costs grow linearly, and the engineering investment in guardrails grows non-linearly.
The monthly token number assumes you're using cost-effective models (Claude Haiku, GPT-4o-mini, Gemini Flash) for the bulk of calls and the heavier frontier models (Claude Opus / Sonnet, GPT-4o) only for the calls that need them. Agencies that pitch "we'll use the best model for everything" are signing you up for a 5-10× token bill.
Run any "we should add AI" idea through these five questions. If 4+ answers point toward AI, it's the right call. If 2 or fewer do, traditional automation is almost certainly cheaper and better.
For a guided 9-question assessment that maps your workflow's AI readiness, the AI Readiness Quiz on /resources walks through the same logic plus data, process, team, tooling, and risk dimensions.
✏️ Note: illustrative compositions, not specific clients. Same underlying workflow (incoming lead intake), three different scopes, three different right answers.
Same problem shape, three different answers. The variable that drives the answer is volume × variability, not the buzzword density of the conversation.
Sometimes, but rarely as a top priority. At under-10 employee scale, the highest-ROI moves are usually basic automation (form-to-CRM, spreadsheet replacement, payment reconciliation) — not AI. AI becomes worthwhile when you have a workflow handling unstructured input at volume (200+ items/month) and saving 4+ hours/week. Below that, the build cost and monthly token spend rarely pay back.
At SMB scale (1–50 employees): light AI integration adds ~$1,500 to build cost and ~$80/month in tokens; heavy AI integration adds ~$5,000 to build cost and ~$350/month in tokens. So a typical 'AI-powered' workflow lands at $7K–$22K all-in for build, $200–$500/month ongoing. Anything quoted at $50K+ for SMB scale is either including services you don't need or assuming enterprise volumes.
A chatbot is a UI; AI automation is a backend pattern. Chatbots talk to users; AI automation processes data behind the scenes. They overlap when you build a customer-facing chatbot that uses LLMs to answer support questions — but the higher-ROI use of AI in SMB is almost always behind-the-scenes (classification, routing, drafting), not user-facing chatbots.
No, it augments them. Zapier (and Make, n8n) are great at structured-input → structured-output workflows; AI adds the ability to handle unstructured inputs and ambiguous routing. The combination is more powerful than either alone — and we usually build production AI workflows on top of one of these platforms when the SMB-scale fit is right.
Three quick tests. (1) Can you describe the workflow's exceptions in plain language? AI projects fail when nobody can articulate the edge cases. (2) Is your data accessible? If the input lives in 3 disconnected systems with no API, the data plumbing dominates the project, not the AI. (3) Do you have someone who can review outputs? AI projects need a human-in-the-loop for the first 30-90 days to tune the prompts and catch failure modes. The AI Readiness Quiz on /resources scores you on these five dimensions in 9 questions.
Buying 'AI' before defining the workflow. The conversation usually goes: 'we need an AI chatbot' → vendor builds an AI chatbot → 6 months later nobody uses it because it solved a problem that wasn't a real bottleneck. The fix: spend 1–2 weeks mapping the actual workflows and their volumes BEFORE deciding what's AI vs traditional. The vendors that lead with the workflow conversation are the ones to work with; the ones that lead with the technology are usually selling hype.
Blog
Send us the workflow — what triggers it, what it does, how often it runs, and what the messy part is. We'll come back with a verdict: AI, traditional automation, or a hybrid. No hard sell either way.
Tell us the problem →
