Two years past the AI hype peak, here's an honest take on what AI in a CRM actually does well — proposal drafting, call summaries, email replies — and what's still mostly vendor theater (lead scoring, predictive close, agentic deal-closers).
Short version: we're two years past the AI-in-CRM hype peak. Most of the "AI features" your CRM vendor is marketing in 2026 are still theater — predictive lead scoring is approximately astrology, agentic deal-closers don't close deals, and "AI insights" is a search bar with extra steps. But three things actually work and save real hours per week: AI proposal drafting, call summarization, and reply drafting. Here's where each one earns its keep and how to set them up without paying vendor markup.
Every CRM in 2026 lists 15-25 "AI features" on their marketing site. They divide cleanly into three tiers based on whether they actually do useful work.
This is the part most vendors won't tell you. Predictive ML models — lead scoring, churn prediction, close probability — need a meaningful volume of historical labeled data to train. The minimum useful threshold is roughly 1,000-2,000 positive examples for a binary classification (will close / won't close).
A service business doing 200 deals a year accumulates 200 closed-won labels per year. After 3 years you have 600. The model trained on 600 examples is going to give you predictions, but the predictions are barely better than rule-based heuristics like "deals at Negotiation stage with a proposal sent close at 60%." You don't need ML for that. You need stage discipline (see how to build a pipeline that doesn't lie).
The vendors selling predictive AI to sub-$10M ARR businesses are either (1) selling fancy weighted-average math dressed up as ML or (2) using a generic pre-trained model that wasn't trained on your data, which means the predictions have no grounding in your specific business.
You give the AI: client name, vertical, problem statement, scope, price. The AI gives you back a 2-page proposal in your brand voice. You edit for 5 minutes. Ship.
This is the single highest-ROI AI feature in a CRM today. A senior practitioner takes 45 minutes to write a good proposal from scratch. With AI drafting, it's 8 minutes. Across an agency doing 8 proposals a week, that's 5 hours/week saved.
Setup: prompt the assistant with 3-5 of your best proposals as voice samples. From then on, every draft matches your voice. Without the voice samples, the output is generic AI prose that needs heavy editing.
Recording in, summary out. Action items extracted with owner and due date. Decisions captured. Next-meeting agenda drafted.
This is the feature that justifies the AI add-on by itself if you do 5+ external calls a week. The honest gotcha: the summary is only as good as the audio. Bad mic, multiple speakers talking over each other, jargon-heavy conversations — accuracy drops. Use a real meeting recorder (Otter, Fathom, Read) and pipe the transcript into the CRM's summarizer.
Incoming customer email, AI drafts a 3-paragraph response in your voice using your CRM's context about the customer. You edit for 90 seconds.
The trick is the customer context. If the AI knows the customer's company, deal stage, last 5 conversations, and current SLA, the reply is grounded. If it doesn't have context, you get generic AI sludge. Context-grounding is what separates a useful reply drafter from a wasted feature.
Customer sends a long email with project requirements scattered across 4 paragraphs. AI extracts: budget range, timeline, scope items, decision-maker, urgency. Populates the deal record.
Underappreciated feature. Saves the "now I have to manually go through this email and update the CRM" tax.
Most CRM vendors built their AI features by paying OpenAI / Claude / Gemini for tokens behind the scenes, then marking up the tokens 2-5x and selling you "AI credits."
Realistic markup math: an AI feature that costs the vendor $0.04 of OpenAI tokens to generate a proposal draft gets sold to you as "1 AI credit" inside a "200 credits/month for $99" plan. Effective cost to you: $0.50 per draft. The vendor margin: $0.46 per draft, scaling.
BYO API keys flips this. You connect your own OpenAI / Claude / Gemini key in the CRM settings. The CRM uses your key for all AI operations. OpenAI bills you directly. No markup.
Real numbers for a mid-volume agency:
That's why most vendors don't offer BYO. The credit system is the margin business. Zay CRM uses BYO on Growth ($400/mo) and Pro ($950/mo) tiers. You pay us for the CRM. You pay OpenAI for the tokens. No middleman markup. See the full pricing comparison for the math on this across vendors.
BYO means you pick the model. Practical guidance as of mid-2026:
Most agencies running Zay CRM use GPT for proposals + Claude for long summaries + Gemini for bulk operations. You can route different tasks to different models in Settings → AI Routing.
Until you have 1,000 closed-won + 1,000 closed-lost deals of historical data with consistent stage labels, AI lead scoring is fitting noise. You're better off with a simple rule-based score: "Industry = target vertical (+10), Title = decision maker (+15), Engaged in last 7 days (+10), Replied to outbound (+25)." Rule-based scores are explainable, debuggable, and roughly as accurate as ML scoring on small datasets.
The pitch is "AI will respond to inbound leads in 60 seconds, qualify them, and book meetings." The reality is the AI messages get flagged as spam, real prospects hate the conversation flow, and the meetings booked have a 70%+ no-show rate because the prospect never actually committed.
What works instead: AI drafts the response, a human reviews and sends. The 60-second responsiveness is real. The "AI does it autonomously" part is theater that costs you brand trust.
AI assistant lives across the entire product on Growth ($400/mo) and Pro ($950/mo) tiers. Starter tier ($100/mo) does not include AI — keeps the entry price low for solo operators who don't need it.
To enable: Settings → Integrations → AI → connect your OpenAI / Claude / Gemini key. From that point on, the AI buttons appear inline across deals, contacts, emails, contracts. Usage is billed directly by the AI provider to your account. We don't see the bill, we don't mark it up, we don't take a cut.
Voice training: upload 3-5 sample documents in Settings → AI → Voice Training. The assistant references these for all proposal drafts and email replies.
Start the 7-day Zay CRM trial — Growth tier covers the full AI stack on BYO keys. Or browse the live demo workspace where the AI assistant is running on sample data with a connected key.
AI in a CRM in 2026 is a small number of features that genuinely save hours per week — proposal drafting, call summaries, reply drafting, data extraction — surrounded by a much larger marketing surface of features that don't work yet. Buy the working ones. Skip the smoke. Bring your own API keys so you stop paying token markup. The AI conversation is about practical hours saved, not vendor talking points.
Try it — Zay CRM 7-day trial · live demo · pipeline discipline post.
30 minutes. We look at your business and tell you exactly what we'd do. client or not.
Book Your Free Audit