AI Offerings in Enterprise Companies: What Each Tier Costs

Enterprise AI offerings fall into four buckets, and knowing which one a vendor is selling determines what you'll actually pay. Hyperscaler platforms (AWS Bedrock, Azure AI Foundry, Google Vertex AI) sell infrastructure at $3–$15 per million tokens. Application-layer AI is bundled into software you already own — Microsoft 365 Copilot at $30/user/month, Salesforce Agentforce at $2 per conversation, ServiceNow Now Assist at roughly a 30% SKU uplift. Vertical AI products target one function: Harvey for legal, Abridge for clinical notes, Sierra for support. And services — Accenture booked $5.9 billion in generative AI bookings in FY2025; Deloitte, IBM Consulting, and boutique agencies sell the integration work around all three.

Below is what each tier actually costs, where the money disappears, and the specific conditions under which none of it is worth buying.

The Four Tiers, With Real Numbers

Tier 1 — Model and infrastructure platforms. You rent compute and tokens. Anthropic's Claude, OpenAI's GPT models, and open-weight Llama variants run on Bedrock, Azure, or Vertex. Frontier-model pricing sits around $3 per million input tokens and $15 per million output tokens; smaller models drop to $0.25/$1.25. A support-deflection bot handling 50,000 conversations a month at ~4,000 tokens each runs roughly $2,000–$4,000 in inference — trivial. The same volume with a multi-step agent that makes 12 tool calls per conversation runs 8–15x that. Token math is not linear with usage; it's linear with *reasoning steps*.

Tier 2 — Embedded copilots in software you already license. Microsoft 365 Copilot: $30/user/month, annual commit, on top of E3/E5. For 500 seats that's $180,000/year. Salesforce shifted Agentforce to consumption billing at roughly $2 per conversation after flat per-seat pricing collided with reality. Google Gemini for Workspace folded AI into Business/Enterprise tiers rather than charging separately — a competitive response, not generosity.

Tier 3 — Vertical applications. Priced per seat or per unit of work, and expensive because they replace labor rather than assist it. Legal AI in the $100–$500/user/month band. Ambient clinical documentation at $200–$400 per clinician per month, justified against 1–2 hours/day of charting time. Contact-center AI at $0.50–$1.50 per resolved ticket versus $6–$14 for a human-handled tier-1 contact.

Tier 4 — Services and integration. Where most of the enterprise budget actually goes. The rough allocation across large deployments: 15–25% software and inference, 75–85% integration, data plumbing, change management, and governance. A $200,000 platform commitment routinely carries $600,000–$900,000 of services around it.

The line item you negotiate hardest is almost never the line item that determines whether the project works.

Where the Money Actually Goes

Enterprise AI budgets get built backwards. Procurement negotiates the license; the license is the small number.

When This Isn't Worth It — And Who Should Not Buy

This section will cost us business. Read it anyway.

Don't buy enterprise AI if your process isn't written down. AI automates a defined process. If your intake, qualification, or dispatch workflow lives in three people's heads and varies by who's on shift, you will spend $150,000 discovering that you don't have a process. Fix the process first — with a spreadsheet, if that's what it takes. The AI will still be there in six months and will cost less.

Don't buy if your task volume is under roughly 500 units per month. Fixed costs of an enterprise deployment — integration, eval, monitoring, governance — run $75,000–$300,000 regardless of volume. At 400 tickets a month, a part-time human is cheaper and more accurate. The break-even for most agent deployments sits between 2,000 and 10,000 repeated tasks per month.

Don't buy per-seat copilots for populations that don't write. Microsoft 365 Copilot returns real value to people who live in Word, Outlook, Excel, and Teams 6+ hours a day. Field technicians, drivers, warehouse staff, and most frontline retail workers will not open it twice. Buying 2,000 seats to cover 400 real users means paying $576,000/year for $115,000 of value. License the 400.

Don't buy if you can't name the metric and its current value. "Improve customer experience" is not a metric. "Reduce average first-response time from 4.2 hours to under 30 minutes on 60% of inbound" is. If nobody can state the baseline number today, you cannot detect improvement later, and the vendor's dashboard will happily supply a number that has nothing to do with your P&L.

Failure modes you should expect, named plainly:

Who should genuinely wait: companies in an active ERP or CRM migration, organizations under 200 employees with fewer than 3 repeatable high-volume workflows, and any business whose current bottleneck is demand rather than delivery capacity. If you can't fill the calendar you have, an AI that books appointments faster solves nothing. Fix demand first — that's a marketing problem, and our free marketing audit will tell you which one you actually have in about 20 minutes.

The Build-vs-Buy Line Most Analyses Get Wrong

The standard framing — build for differentiation, buy for commodity — is directionally fine and operationally useless. The sharper test is rate of change in the underlying capability.

Buy anything where the capability is improving faster than you can rebuild it: transcription, OCR, translation, general-purpose reasoning, code completion. These improved 10x in 24 months. Anything you build there is depreciating on the day you ship it.

Build the layer that encodes *your* facts: your pricing rules, your eligibility logic, your escalation policy, your service-area geography, your definition of a qualified lead. That layer isn't intelligence — it's a few hundred lines of business rules plus a well-governed data source. Vendors will happily sell you a $250,000 "custom AI" that is 90% this, and it's the 90% you should own.

A practical ratio from deployments that stick: roughly 70% off-the-shelf model and platform, 30% proprietary data and rules. Programs that invert that ratio — 70% custom — take 9–18 months and are usually rebuilt within two years.

What Mid-Market Service Businesses Should Copy — and Skip

Most enterprise AI writing assumes 10,000 employees and a $4M budget. If you run a 40–400 person service business, three enterprise practices transfer directly and the rest don't.

Copy the evaluation discipline. Before deploying an AI receptionist or intake agent, write 50 real historical calls or inquiries into a test set with correct outcomes. Run every prompt change against it. This costs a day and prevents the most expensive failure mode there is.

Copy the ownership model. One named person owns the metric. Not a committee, not "the vendor."

Copy the staged rollout. 10% of traffic, then 30%, then 60%, with a human-review sample at each stage and a documented rollback.

Skip the platform commitment. You don't need a $200,000 annual Bedrock commit to run 40,000 conversations a month. Consumption pricing is your friend at your volume; enterprise commits are discounts on volume you don't have.

Skip the multi-agent architecture. Two-agent systems fail in ways that are hard to debug and harder to price. A single well-instructed agent with 4–6 tools handles most service-business workflows — booking, qualification, follow-up, review requests — at a fraction of the operational complexity.

Practical entry points at this scale: an AI voice agent answering after-hours calls at roughly $0.09–$0.25 per minute versus $1,200–$2,800/month for an answering service; automated speed-to-lead follow-up, where responding in under 5 minutes rather than 30 raises contact rates several-fold; and review-request automation, which moves a 4.2-star profile to 4.7 in about 90 days at typical volumes. Our AI marketing agency page covers how these get sequenced, and the ROI calculator will tell you whether your current lead volume clears the break-even threshold before you talk to anyone.

How to Evaluate Any Enterprise AI Offering in One Meeting

Five questions. If a vendor can't answer them with numbers, the offering isn't ready.

Then set a kill date before you start. A pilot with no expiration becomes a permanent line item that nobody has authority to cancel. Ninety days, one metric, a named owner, and a documented decision at the end — kill it, scale it, or explicitly extend it once with a reason in writing.

The organizations getting real returns aren't the ones that bought the most AI. They're the ones that picked two workflows with high volume and clear metrics, instrumented them properly, and refused to expand until the numbers held for a full quarter.