How much does it cost to use an AI assistant?

The Real Price Tag: Why AI Assistant Costs Are So Confusing

Ask ten companies what they pay for an AI assistant and you'll get ten answers that don't remotely resemble each other — because the honest answer ranges from zero dollars to over a million.

At the consumer level, the ladder runs from genuinely free tiers up to $300/month for legacy premium plans, though most standard paid subscriptions cluster tightly around $20 source. Business subscriptions span $39 to $999+ per month, while custom enterprise builds start around $75,000 and can exceed $1 million source. That spread alone would make budgeting hard enough. The real problem is deeper.

The pricing model — not the plan name — determines the bill. Industry analysis identifies six dominant models: per-seat, flat subscription, per-resolution, per-conversation, usage credits, and custom contracts source. The same workload can swing the bill dramatically depending on which one you're locked into. One documented example — 2,000 conversations per month handled by a small team — cost $79/month on a flat plan versus $1,077/month under per-outcome pricing source.

Why the spread? AI has reintroduced marginal costs that traditional SaaS never carried. Every query burns real compute, which is why AI companies typically run 50–60% gross margins versus 80–90% for conventional software vendors source. Vendors are passing that cost structure straight through to you, in increasingly creative ways.

And here's the trap: the plan fee is the only number printed in large type on a pricing page. The full cost of ownership includes several lines that rarely make the headline:

• Usage overages and token consumption, which analysts note often run 5–10x initial estimates due to context bloat source
• Add-ons like AI copilots billed at $50 per human agent per month source
• Human seats for the people still supervising the assistant
• Setup and maintenance, which for custom builds includes retraining at $2K–$10K per quarter and integrations at $5K–$25K per API source

These hidden lines routinely exceed the plan fee itself. That's why comparing two vendors by their headline price is almost meaningless — you're often comparing two different pricing models wearing similar price tags. If you're extending an assistant with agent skills to squeeze more value from what you already pay for, SkillzBench's free directory and MCP connector can help you test which skills actually deliver lift before they add to your bill.

The Six Pricing Models That Decide Your Bill

The pricing model—not the plan name—is what actually decides your bill. The same AI assistant workload can cost eight times more under one pricing structure than another, which is why understanding the six dominant models matters more than comparing brand logos.

1. Per-seat pricing charges for human users, like traditional SaaS. Zendesk seats run $19–$115/agent/month, while Intercom tiers span $29 to $132 per seat. This works when humans stay central to the workflow, but it ignores how much the AI actually does.

2. Flat subscription offers predictable monthly pricing. Most small businesses land in the $29–$99/month range, with paid plans commonly running $39 to $999/month. You pay the same whether the bot handles 100 conversations or 10,000.

3. Per-resolution (outcome) pricing bills only when the AI actually solves something. Intercom's Fin set the template at $0.99 per resolution, with Zendesk following at $1.50–$2.00, according to pricing research. It aligns incentives—but carries a catch we'll get to shortly.

4. Per-conversation pricing charges per interaction regardless of outcome. Botpress runs $150/month for 250 conversations up to $750/month for 1,500. As one analysis puts it, a conversation ending in customer frustration costs the same as one that succeeds.

5. Usage/credits pricing sells prepaid capacity. Chatbase spans $40/month (700 credits) to $500/month (15,000 credits), with overages at $40 per 1,000 credits.

6. Custom contracts dominate enterprise, where list prices stop being the real price. A median Ada contract runs $72,000/year, with buyers negotiating roughly 17% off on average.

The headline finding from SiteGPT's comparison: an identical 2,000-conversation workload with three human agents cost $79/month on a flat plan versus $1,077/month on Fin's outcome-based pricing—an 8x gap worth $11,376/year.

The counterintuitive risk with outcome-based billing: your bill grows as the AI gets better. When Fin's resolution rate climbs from 50% to 70%, the monthly line jumps from $990 to $1,386. You're celebrating success while your budget absorbs it.

To compare offers honestly, model your own numbers across all six structures:

• Estimate monthly conversation volume and realistic resolution rates
• Add human seat costs, add-ons, and overage scenarios
• Model both low and high resolution-rate outcomes before committing
• Treat enterprise list prices as a ceiling, never a floor

Once you've picked a pricing model that fits your workload shape, extending your assistant's capabilities with vetted agent skills can stretch that budget further—SkillzBench is building a free directory and MCP connector to help you find and test them.

What You'll Actually Pay: Consumer Plans, SMBs, and Enterprises

The gap between what individuals pay and what businesses pay for the same AI capability is wider than most buyers expect — and the pricing model you choose matters more than the brand on the box.

Consumer plans cluster tightly

Individual users face a remarkably consistent market. Current subscription pricing shows ChatGPT Plus, Claude Pro, Google AI Pro, and Perplexity Pro all landing within a penny of $20/month for their standard paid tiers. xAI's Premium is the outlier at $30.

The full ladder runs from $8/month (ChatGPT Go, xAI Plus) to $300/month for legacy SuperGrok Heavy — a $292 spread. Premium tiers like ChatGPT Pro and Claude Max 20x sit at $200/month, while Google AI Ultra bundles extra storage and YouTube Premium at $99.99.

Small and midsize businesses: $29–$99/month is the norm

Most SMBs land on subscription plans in the $29–$99/month range, with paid tiers spanning $9–$999/month depending on volume and features, according to pricing analysis. Usage-based alternatives charge roughly $0.50–$6 per resolved conversation.

The catch: enterprise pricing pages often mislead small buyers into thinking the whole category is out of reach. A 12-person firm and a 1,200-person firm are effectively buying two different products that share a name.

Enterprise: list prices stop being real prices

At scale, published rates become a negotiation ceiling. Data from 117 Ada purchases shows a median contract of $72,000/year (range $36,158–$295,400), with buyers negotiating 17% off on average, per enterprise cost research. Treat any published enterprise number as a starting point, not a final figure.

Buy vs. build: the math most teams underestimate

Custom builds sound appealing until the invoices arrive. Here's what the real cost stack looks like:

• Custom AI bot development: $75,000–$500,000+ depending on complexity and domain specificity
• In-house team salaries: $400,000–$700,000/year before benefits or infrastructure
• Model retraining: $2,000–$10,000 per quarter to keep accuracy current
• Custom integrations: $5,000–$25,000 per API connection
• Ongoing maintenance: 15–30% of initial build cost annually

For most businesses, buying beats building — building only wins at very large scale or when requirements are genuinely unique. A 60–90 day rental period before commissioning custom work lets real usage write the spec.

Once you've picked a tool and a budget tier, agent skills can extend its workflows without new engineering spend. SkillzBench is building a free directory of agent skills with security scanning and performance benchmarking, plus a free MCP connector — everything on the platform stays free.

How to Choose the Right Pricing Model and Cut Costs

The same workload can cost eight times more under one pricing model than another, so the model — not the vendor — decides what you pay.

Start by matching the pricing model to your workload shape. A flat subscription rewards predictable, high-volume traffic where you can spread a fixed fee across many interactions. Per-resolution pricing aligns with value when most conversations actually get solved, but it punishes low resolution rates and quietly grows the bill as the AI improves. Per-conversation models charge even for abandoned chats, so a frustrating dead end costs the same as a solved problem. Usage-based credit systems suit spiky or experimental workloads, while custom contracts only pencil out at large scale.

Model both scenarios before signing anything. A 2,000-conversations-per-month workload with three human agents cost $79/month on one flat plan and $1,077/month on per-outcome pricing — a $948/month gap for identical traffic. Push your resolution rate from 50% to 70% under per-resolution billing and watch the line climb from $990 to $1,386/month, because the bill grows with the AI's success.

Budget five cost lines, not the headline fee:

• Plan fee (what's printed in large type)
• Overages and usage beyond included quotas
• Add-ons like copilots, premium models, or analytics modules
• Human seats for supervisors and escalations
• Setup, integration, and ongoing maintenance time

For custom builds, hidden items stack up fast: maintenance runs $1,000–$20,000+ per year, model retraining $2,000–$10,000 per quarter, and each custom integration adds $5,000–$25,000. Enterprise deployments alone can take three to six months.

Watch the hidden tax on API bills. Token waste from context bloat — system prompts, tool call outputs, accumulated conversation history — routinely inflates estimates by 5–10x. A single agent run with 30,000 input tokens costs roughly $0.075 at GPT-4o pricing; at 10,000 runs per day, input tokens alone run $22,500 per month. The highest-leverage lever is model routing: send 60–70% of routine traffic to cheap Tier 1 models at $0.15–$0.30 per million tokens, and reserve premium models for the queries that actually need them. Prompt caching can cut cached input costs by about 90%, and semantic compression trims 70–90% off tool output tokens.

Before any custom build, rent for 60–90 days. The rental writes the spec for the build — real usage reveals which workflow your team actually runs, not the one described in the kickoff meeting.

Once you've picked your assistant, agent skills can automate the rest of the workflow. SkillzBench is building a free directory, security scanning, and benchmarking of agent skills for Claude Code, Codex, Cursor, Gemini CLI, and chat-app agents, plus a free MCP connector — everything free, with the waitlist live now.