THINXSTER
Blog/AI Marketing
AI Marketing8 min readAugust 20, 2026

AI Software Stack Best Practices: The 4-Layer Build

Run four layers, in order: CRM, workflow automation, one or two model APIs, and evals. Budget 12–18% of marketing spend on tooling — and skip the vector DB.

RK
Ryan Korsz
Founder & CEO, Thinxster

TL;DR

Run four layers, in order: CRM, workflow automation, one or two model APIs, and evals. Budget 12–18% of marketing spend on tooling — and skip the vector DB.

→ See how this applies to your business (free 30-min call)

Most teams should run four layers and nothing more: a system of record (CRM), a connective tissue layer (workflow automation), one or two model APIs, and an observability/eval layer. Standardize on one CRM before adding any AI. Budget 12–18% of your marketing spend on tooling, not 40%. Pick tools that your least technical employee can operate on a Tuesday afternoon without you. And do not buy a vector database until you have at least 5,000 documents that a Google Sheet plus a good prompt genuinely cannot handle — which, for a 12-truck HVAC company, is roughly never.

That is the whole answer. The rest of this page is the reasoning, the numbers, and the part where I tell you when the stack is a waste of money.

The Four Layers, In Build Order

Order matters more than tool choice. Every failed AI stack I've seen was assembled in the wrong sequence — usually starting with the shiny model integration and working backward toward the data, which is like installing a sound system before pouring the foundation.

  • Layer 1 — System of record. GoHighLevel, HubSpot, ServiceTitan, Jobber. One system. If your leads live in three places, every AI you bolt on will confidently hallucinate against stale data. Cost: $97–$497/mo for GHL, $800–$3,600/mo for HubSpot Professional at 5–10 seats, $398/mo entry for ServiceTitan-class field software.
  • Layer 2 — Automation/orchestration. Make.com ($9–$29/mo for 10k–150k operations), n8n (self-hosted ~$18/mo on a VPS, or $25/mo cloud starter), Zapier ($29.99–$103.50/mo). This layer is where 80% of your actual value gets created and it costs under $100/mo.
  • Layer 3 — Model APIs. Claude, GPT, or Gemini via API. A service business handling 2,000 inbound messages/month with 800-token prompts and 400-token responses runs roughly $8–$60/mo depending on model tier. Not a typo. Model costs are almost never the expensive part.
  • Layer 4 — Observability and evals. Langfuse (free self-hosted, $59/mo hosted), Helicone ($20/mo), or a spreadsheet of 40 golden test cases you re-run manually before every prompt change. Skipping this layer is why teams wake up to an agent that started quoting $50 drain cleanings in month four.
  • Voice deserves its own note: Bland, Vapi, and Retell run $0.07–$0.15 per minute. A 4-minute qualification call costs $0.28–$0.60. Against a $65–$200 cost-per-lead in home services, the math works — but only if the call actually books. See our Bland AI implementation work for how that gets wired to a calendar.

    Pick Boring Tools Your Staff Can Actually Run

    The single strongest predictor of whether an AI stack survives month six is not model quality. It's whether the office manager can fix a broken automation without filing a ticket.

    Use this test: if the person who answers your phones can't modify a workflow within 20 minutes of training, you bought the wrong tool. Custom Python running on a Lambda you configured at 11pm is not a stack, it's a hostage situation. When that developer leaves — and the average tenure at a small agency is 18–24 months — you inherit an unmaintainable black box.

    The best AI stack is the one that still works the week after the person who built it stops answering emails.

    Concrete version of this principle: prefer n8n over raw code, prefer GHL workflows over n8n where GHL can do it, prefer a well-written email template over an AI-generated one where the template performs within 5% of the AI version. Every layer of cleverness you add costs maintenance forever.

    The Data Layer Nobody Budgets For

    Here's the line item that kills projects. Teams budget $500/mo for tools and $0 for data cleanup, then discover their CRM has 14,000 contacts with a 31% duplicate rate, 22% missing phone numbers, and free-text job-type fields containing 47 spelling variants of "water heater."

    Realistic cleanup effort for a 10,000–25,000 record CRM: 40–80 hours, or $2,000–$6,000 if outsourced at $50–$75/hr. Do this before Layer 3, not after. An AI answering service trained on dirty data doesn't fail loudly — it fails at a 12% wrong-answer rate that nobody catches until a customer complains.

    Practical thresholds worth writing down:

  • Under 500 records: skip automation entirely, a human is cheaper and better.
  • 500–5,000 records: automation yes, retrieval/RAG no. Prompt-stuffing your 30-page service manual into a 200k-token context window works fine and costs pennies.
  • 5,000–50,000 documents: now consider retrieval. Budget 30–60 hours of engineering and expect 6–10 weeks before it beats the simple version.
  • Above 50,000: you have a data engineering problem, not a marketing problem. Hire accordingly.
  • When This Isn't Worth It — Read This Part Twice

    I'd rather lose the deal than sell you a stack that sits unused. Here is who should close this tab.

    You do more than 60% of revenue from referrals and repeat customers. Your problem is capacity, not lead flow. A $2,500/mo AI stack pointed at a pipeline you can't service produces frustrated leads and one-star reviews. Buy a truck. Hire a tech. Come back at 40% referral mix.

    Your revenue is under about $750K/year. Below that, a $1,500–$3,000/mo stack is 2.4–4.8% of revenue going to tooling, and the owner is still the one operating it at 10pm. The honest alternative: $99/mo GoHighLevel, a $300 one-time setup of three automations, and a $22/hr part-time VA. That covers roughly 70% of the value at 8% of the cost. We have said this to prospects and lost the deal. It was still the right call.

    Your close rate on qualified leads is under 20%. AI will double your lead volume and your close rate will stay broken. Doubling leads at an 18% close rate against a 35% benchmark means you're burning twice the ad spend to leave the same money on the table. Fix sales process first — it's free.

    You want it live in two weeks. Real timeline from kickoff to a stack you trust unattended: 8–14 weeks. Weeks 1–3 are data audit and CRM consolidation, which produce zero visible output and are the phase most clients try to skip. Anyone promising 14 days is either reselling a template or lying.

    Seasonal businesses with a sub-4-month season. Roofing in a hail market, holiday lighting, pool openings. You'll pay 12 months of subscriptions to use the stack for 4. Unless the platform supports genuine pause/downgrade, your effective cost per active month is 3× what the pricing page says.

    Now the failure modes, plainly:

  • Silent degradation. Model providers update models. A prompt tuned in March performs measurably worse in July with no error thrown. Without Layer 4 you find out from an angry customer. Re-run your golden test set monthly — budget 2 hours.
  • Automation loops. A Make scenario that triggers a CRM update that triggers the scenario. We have watched this burn 40,000 operations in 90 minutes and $180 in overage. Always set an execution cap.
  • Tool sprawl. The median stack we audit has 9–14 subscriptions, of which 3–5 are unused and 2 are duplicates (two schedulers, two email senders). Typical recoverable waste: $180–$400/mo. Audit quarterly.
  • The 20% tail. AI voice and chat handle roughly 75–85% of inbound cleanly. The remaining 15–25% — angry customers, complex quotes, elderly callers who hang up on robots — need a human escalation path built on day one, not month five.
  • Attribution fog. Adding 6 touchpoints across 4 tools makes it harder, not easier, to know what worked. If you can't currently answer "what did a lead cost from source X," AI makes that worse before better.
  • What Actually Belongs In Your First 90 Days

    A defensible sequence, with the numbers we'd hold ourselves to:

  • Days 1–21: CRM consolidation and dedupe. Success metric: duplicate rate under 3%, phone coverage above 95%.
  • Days 22–45: Speed-to-lead automation. Target first response under 5 minutes — the difference between a 5-minute and a 30-minute response is roughly a 5–8× drop in contact rate. This single automation typically pays for the entire stack.
  • Days 46–70: Review requests, missed-call text-back, and reactivation campaigns against your dormant list. A 12,000-contact dormant list at a 1.5% reactivation rate and $1,800 average ticket is $324,000 in pipeline from a $40 SMS spend.
  • Days 71–90: Add AI voice or chat for after-hours only. After-hours first, always — the downside of a bad AI conversation at 11pm is far smaller than at 11am.
  • Run your own numbers before signing anything with our ROI calculator, and compare against what in-house versus agency actually costs. If the payback period exceeds 6 months on paper, it will exceed 12 in reality.

    The Contrarian Take On Model Choice

    Most stack guides spend 2,000 words on which model to pick. That decision matters less than almost anything else on this page. The measured gap between frontier models on the tasks a service business actually runs — classifying an inbound lead, drafting a follow-up, summarizing a call — is small enough that prompt quality and data quality swamp model choice by a factor of roughly 5 to 1.

    What matters instead: build an abstraction so you can swap models in under 30 minutes. Route the cheap, high-volume tasks (classification, tagging, routing — probably 85% of your call volume) to a small fast model at $0.25–$1 per million input tokens, and reserve the frontier model for the 15% that touches a customer directly. Teams that do this cut model spend 60–70% with no measurable quality change.

    And write down your evals before you write your prompts. Forty real examples with known-correct answers, in a spreadsheet, is a more valuable asset than any tool on your invoice — it's the only thing that tells you whether a change helped.

    The One-Page Summary

    Standardize the CRM. Clean the data before you automate it. Spend your money on Layers 1 and 2, not on model tokens. Keep the total stack under 6 tools and under 15% of marketing spend. Build the human escalation path on day one. Test monthly with a fixed set of examples. And if your close rate is broken, your season is 4 months long, or referrals are carrying you — don't buy any of this yet.

    Frequently Asked Questions

    What are the four layers of an AI software stack?

    A system of record (your CRM), a connective tissue layer (workflow automation), one or two model APIs, and an observability/eval layer. Build them in that order. Sequence matters more than tool choice — most failed stacks start with the model API and never establish a clean system of record underneath it.

    How much should a small business budget for AI tools?

    Roughly 12–18% of marketing spend on tooling, not 40%. Above that range you are usually paying for capability nobody on staff operates. A practical test: if your least technical employee cannot run the tool on a Tuesday afternoon without you, it is not producing the value its price implies.

    Do I need a vector database for my AI stack?

    Almost certainly not. Wait until you have at least 5,000 documents that a spreadsheet plus a well-written prompt genuinely cannot handle. For a 12-truck HVAC company, that threshold is effectively never reached. Vector databases add infrastructure, cost, and failure modes before they add retrieval quality.

    Should I set up a CRM before adding AI tools?

    Yes. Standardize on one CRM before adding any AI. The CRM is your system of record, and every layer above it — automation, model calls, evals — reads and writes against that data. Adding AI on top of two competing CRMs or scattered spreadsheets multiplies the mess rather than resolving it.

    Free Weekly Briefing

    One AI Marketing Tactic.
    Every Tuesday. Free.

    What's actually working across our client accounts right now — ROAS moves, follow-up sequences, creative angles. The stuff that isn't in any blog post yet.

    No spam. Unsubscribe anytime. 1,200+ business owners already in.

    Ready to Deploy

    SEE THIS IN
    YOUR BUSINESS.

    30 minutes. We scope the exact systems that apply to your situation and give you a plan.

    ★★★★★ Trusted by 47+ local service businesses

    BOOK A STRATEGY CALL →