📘 Definitive Guide
Custom AI vs SaaS AI for Service Businesses (2026 Guide) →
Every dollar your phone leaks. Every fix that works. 5,000-word reference built from 60+ deployments.

SaaS AI demos are designed to sell you. The sales rep is good. The product looks polished. The pricing slide has a single number on it that feels reasonable. By the end of the call, you have a proposal in your inbox and a soft deadline to sign before "the Q3 promotional pricing expires." What you don't have is any of the information you'd need to actually evaluate the deal.

This post is the buyer's playbook. Five questions every operator should ask before signing with ServiceTitan AI, Avoca, Goodcall, or any other home-services AI voice platform. They're the questions vendors won't volunteer because the honest answer makes the deal harder to close. Each one comes with a sample script you can read off the page, and a flag for what answer should make you walk.

01.What's Your Per-Call or Per-Minute Pricing Escalator?

Every SaaS AI vendor sells a base plan with included minutes or calls. Almost none of them lead with what happens when you exceed the cap. The reason is that overages are where their highest-margin revenue comes from. A storm-season month or a heat-wave week can blow past the cap inside a single afternoon, and the overage rate is typically 3-5x the marginal cost of the underlying voice infrastructure.

What you need to know specifically:

"Walk me through exactly what happens if we go over our included minutes in a given month. What's the per-minute overage rate? Is that rate the same at 10% over and 200% over? If we have a storm event that 5x's our normal volume in 72 hours, what's the realistic invoice impact?"

Red flag: the rep can't answer without "checking with the team," or quotes a range like "it depends on your usage." Both mean the overage math will not be in your favor. Get a specific number in writing before signing.

02.What Can I Export If I Leave?

This is the single most important question and the one vendors least want to answer. The honest answer for most SaaS AI platforms is "very little." Recordings stay with the vendor. Transcripts stay with the vendor. Trained voice models stay with the vendor. Custom FAQ entries you painstakingly tuned over 14 months? You can copy them out of the dashboard, one at a time, if you're patient.

What you actually need on exit:

"If I decide to leave you in 18 months, what specifically do I get back? I'd like that in writing as a one-page exhibit attached to the contract. Specifically: recordings, transcripts, customer profiles, custom prompts, trained voice profiles. Yes or no on each, and in what format."

Red flag: the rep says "we don't typically have customers leave" or pivots to talking about their renewal rate. That's a non-answer. If they can't put a specific export policy in writing, assume you get nothing on exit. The Texas roofer we wrote about in The True Cost of SaaS AI paid an extra $1,500 for a "data services fee" that produced a partial CSV with no recordings. The vendor wasn't lying — they were just doing exactly what their export policy permitted.

If the vendor cannot put their data export policy on a single page and attach it to the contract, the honest reading is that you don't own your data. The vendor does.

03.Can I See and Edit the System Prompt?

The system prompt is the instruction set the LLM reads on every call. It determines how the AI greets callers, what questions it asks, when it escalates, and what tone it uses. On most SaaS AI platforms, the system prompt is shared infrastructure — one prompt across all customers, configured at the vendor level. The customer cannot see it, cannot edit it, and cannot version it.

This matters because the prompt is the AI. If you can't see or change it, you're not running a customized solution — you're running the vendor's solution with your business name pasted into a few fields. When the vendor updates the prompt to fix a bug for someone else, your AI behavior changes too.

"Can I see the actual system prompt the AI uses on every call? Can I edit specific sections — the qualification questions, the escalation triggers, the tone? If I don't like a change you make at the vendor level, can I roll back to the prior version for my account?"

Red flag: the rep says "we handle that for you" or "the prompt is proprietary." Both translate to "no, you can't see it." That's the boundary between configured-SaaS and actually-custom. It's a fair tradeoff at the SaaS price tier — but you should know which side of that line you're buying on.

The five things a real custom prompt should encode (and that SaaS prompts almost never do well) are covered in the full Custom vs SaaS guide.

04.How Do You Handle Multi-Location, Multi-Trade, or Specialty Dispatch?

This is the question that breaks most SaaS AI sales pitches. The standard pitch assumes a single-location, single-trade operator. The moment your business has more than one branch, more than one service line, or any dispatch logic that depends on tech skill, certifications, or territory, you've stepped outside the SaaS happy path.

What to probe specifically:

"I have [N] locations across [N] states. We run [trade A] and [trade B]. We have a different on-call rotation per branch. Walk me through exactly what happens on a 2am emergency call from zip 75201 — which calendar does the AI book? Who gets paged? What if the caller is in a zip we don't serve?"

Red flag: the rep waves their hand and says "we handle multi-location" without walking through the specific scenario. The Wegner Roofing case (9 locations across 5 states) we cover in Why Multi-Location Service Businesses Need Custom AI is the canonical example of why SaaS multi-location handling almost always falls short of what the demo suggested.

05.What's Your Integration Depth With My CRM?

This is where most operators get burned the slowest and most painfully. The vendor's marketing says "integrates with ServiceTitan/AccuLynx/JobNimbus/FieldEdge." What that usually means in practice is a one-way webhook that pushes a new lead into the CRM after the call. What it almost never means is that the AI can read customer history, recognize repeat callers, look up existing jobs, check warranty status, see open invoices, or update dispatch in real time.

The depth matters. A one-way push means every caller is treated as a brand-new lead, including the customers you've worked with for ten years. A two-way integration means the AI greets the repeat customer by name, knows their last service was a roof tune-up in 2024, and can offer them the right next-step without going through full intake again.

"Walk me through what data flows in and out of my CRM specifically. Can the AI read prior job history before greeting a caller? Can it update an existing job's status? Can it create a follow-up task assigned to a specific technician? Can it see whether the caller has an open invoice and decline to schedule until it's resolved? Show me — don't tell me."

Red flag: the rep talks about "Zapier" or "make.com" or any third-party connector. Those are fine as glue, but they're not real integrations. If the vendor's actual integration story is "you connect it through Zapier," you're going to be doing a lot of manual data engineering. Real integrations are direct, two-way, and maintained by the vendor.

Need a second opinion on a vendor proposal?

Free 30-minute audit. Bring us the proposal you're evaluating. We'll pressure-test it against your specific call mix and tell you the honest answer — whether to sign, negotiate, or walk.

Book Free Audit →

06.The Bonus Question: How Do You Get Paid?

This is the question nobody asks and the one that explains everything else. SaaS AI vendors are incentivized to maximize seat retention and minute consumption — that's where their revenue grows. They are not incentivized to optimize your booking rate or your recovered revenue, because those are your metrics, not theirs.

Custom AI agencies (SimpliScale included) are incentivized differently. We get paid a flat monthly retainer that doesn't scale with your minute usage. Our renewals depend on whether your booking rate actually moved. That alignment isn't perfect, but it's structurally different from SaaS — and it changes how aggressively we tune the system for your specific outcomes.

Knowing how a vendor gets paid tells you what they're going to push you toward. SaaS will push you toward more minutes, more locations, more add-ons. Custom will push you toward better booking rate, deeper integration, and ownership. Neither is inherently better — they're just different alignments. Pick the one that matches what you're actually optimizing for.

07.The Walk-Away Test

If you've asked all five questions and the rep can't put specific answers in writing, the right move is to walk. Not because the product is bad — most SaaS AI products are competent — but because you're being asked to sign a 12-month contract on assumptions you can't verify. Any vendor who is uncomfortable with five specific buyer questions is signaling something about their product or their pricing.

The operators we work with who came from SaaS-AI experiences almost universally describe the same regret: "I wish I'd asked these questions before signing." This post exists so you don't have to learn it that way.