The AI Seat You Bought Costs More Than You Think
• 8 min read
Part 2 of 2: Vendor selection, plan mix, usage monitoring, and the KPIs that survive a board meeting.
List price is not your cost. Seat count is not your ROI. And the two most expensive mistakes are buying premium tiers for people who won’t use them, and buying nothing while your team quietly uses free accounts.
Part 1 covered the decisions that come before the purchase — scope, license structure, and where the data line sits. This one covers the money.
Vendor selection involves two key questions
Most owners collapse these into a single comparison and pick randomly – that’s bad practice.
Question 1: Where does the work actually happen? And Who’s doing the work?
This decides more than model quality does.
If your company lives in Microsoft 365 — Outlook, Teams, Word, Excel, SharePoint — Copilot has home-field advantage. It runs inside your tenant, inherits your existing security and data rules, and meets people where they already work. The catch is that it requires a qualifying M365 license underneath, so the all-in cost is higher than the add-on price suggests.
If your company lives in the browser — Google Workspace, Slack, Notion, a CRM, custom tools — a standalone platform fits better. Gemini is the natural analogue for Workspace shops. Claude and ChatGPT run in their own apps and connect out to your stack, which suits teams doing dense document work, analysis, or anything technical.
Practical adoption note from deployment data: teams that anchor pilots on email drafting and meeting summaries tend to convert to full rollout. Pilots anchored on spreadsheet-first use cases tend to stall until a manual review habit forms around the output. Choose your pilot use case accordingly.
Question 2: Will the contract survive a customer’s security review?
This is the gate that quietly kills deals, and the one owners skip. Before signing, get written answers on:
- A data processing agreement that names sub-processors individually
- A stated default retention period — in writing. A vendor that can’t state one hasn’t thought it through
- Training exclusion by default, and whether zero-data-retention is available on your tier
- SOC 2, and data residency if you have regulated or non-US customers
- Admin console access with per-user usage logs
- Exit terms: what happens to your data and your conversation history when you leave
If you have EU exposure, note that buying a tool makes you a deployer under the EU AI Act, not a provider — and deployers carry their own obligations for high-risk uses, including human oversight, monitoring, and log retention of at least six months. Those duties don’t transfer to the vendor because someone else built the model.
The plan mix > plan choice
Current list anchors (verify at signature; these move quarterly):
| Tier | Approx. price |
|---|---|
| Individual pro tiers (ChatGPT Plus, Claude Pro, Gemini) | $20 / user / month |
| Copilot Business | ~$21 / user / month |
| Claude Team | $25 / seat monthly, ~$20 annual, 5-seat minimum |
| ChatGPT Team / Business | ~$25 / seat, ~$20 annual |
| Microsoft 365 Copilot | ~$30 / user / month annual, on top of a qualifying M365 plan |
| ChatGPT Enterprise | ~$45–75 / user / month, 150-seat annual minimum |
Two things to notice.
The headline prices are nearly identical. At the business tier, the spread between the major vendors is small enough that price should not be your deciding variable. Fit and contract terms should be.
Enterprise tiers are usually the wrong buy for an SMB. The seat minimums alone disqualify most companies under a few hundred people, and you’re paying for administrative controls you don’t yet have the maturity to use.
The structure most mid-market companies converge on is layered, not single-vendor: a broad, lower-cost tier for everyone as the sanctioned default, plus a small pool of premium seats for identified power users. The combined cost per power user typically still lands under a single enterprise seat of either competitor, and it gives you leverage at renewal because you’re not fully migrated onto one platform.
The counter-argument is real: multi-vendor creates fragmentation, and fragmentation is one of the top complaints in SMB AI surveys. Manage it by making one tool the default and the second one deliberate and named, not by letting both spread organically.
For budget scale: published SMB averages put total annual AI tool and subscription spend around $18,000, and per-employee AI spend at larger companies around $1,240 a year.
The long-tail hidden cost: utilization abandonment
This is the number that matters most and gets modeled least.
In most enterprise deployments, active daily usage concentrates in 20–30% of licensed seats. The other 70–80% use the tool occasionally or not at all — but the invoice doesn’t distinguish. Zylo’s 2026 SaaS Management Index puts unused licenses across all SaaS at 36%.
Do the arithmetic on your own deployment. Twenty seats at $25 is $500 a month. If eight people use it weekly, your real cost per active user is $62.50, not $25. That’s the number to compare against the value you’re claiming, and the number to bring to a renewal negotiation.
Two structural controls:
- Negotiate true-down rights at 6 and 12 months. Vendors are competing hard for AI-era positioning right now and are unusually flexible. That window closes as the market stabilizes.
- Set an automatic inactivity reclaim. Ninety days is the standard default for AI tools. Reclaimed seats go back to the pool, not to the invoice.
Also watch the consumption layer. Seat pricing is no longer the whole bill — agentic features increasingly bill per credit or per token on top of the seat price, which is a variable line item that didn’t exist a year ago. Ask what the metered components are and set a spend alert before rollout, not after the first surprise invoice.
Monitoring usage without building surveillance
The goal is aggregate visibility, not reading people’s prompts. Reading prompts destroys trust and kills adoption faster than any policy.
Monitor at three levels:
- Seat level, monthly: active vs. licensed seats, last-active date per user, seats crossing the inactivity threshold. Every business tier admin console gives you this natively.
- Spend level, quarterly: total AI spend, cost per active user, and overlapping tools. If three departments are separately paying for three assistants, that surfaces here.
- Workflow level, per deployment: which specific tasks the tool is being used for. This comes from asking people, not from telemetry — a ten-minute monthly check-in with each pilot group is more informative than any dashboard.
Tell people what you’re measuring, in the policy, before you start. Aggregate usage metrics land fine when disclosed and land badly when discovered.
Lightweight productivity KPIs
The measurement bar has moved. Among 830 IT decision-makers surveyed in 2026, productivity gains fell from 23.8% to 18.0% as the primary ROI metric, while direct financial impact — revenue and profitability — nearly doubled to 21.7%. “Saves four hours a week” is no longer a credible justification on its own.
That said, an SMB doesn’t need an ROI model. It needs three or four numbers it can actually capture. The non-negotiable part is capturing them before rollout — if your controller can’t say how long reconciliation takes today, you can’t prove AI shortened it.
Pick one from each category and baseline it for two weeks pre-launch:
Volume — output produced per person per week in the target workflow. Proposals sent, tickets resolved, posts published, invoices processed. Easiest to capture, easiest to argue with.
Cycle time — time from request to delivered output. Quote turnaround, support first-response time, content brief to publish. This is usually the most persuasive single number because customers feel it.
Quality proxy — rework rate, revision rounds, error rate, or escalations. This is the one people skip, and skipping it is how you end up with a “win” that just moved work to the reviewer.
Cost line — contractor or overtime spend in the affected function, or headcount deferred. This is the one a board or an investor actually wants, and it’s the only one that survives real scrutiny.
Review at 30, 60, and 90 days, and force a decision each time: expand, revise, hold, or shut down. A pilot that saves drafting time but creates more review work should not be expanded.
One realistic expectation to set with whoever approves the budget: PwC’s 2026 study of over 1,200 executives found that 74% of AI’s economic value is captured by roughly 20% of organizations, and estimated that the technology itself delivers about 20% of an initiative’s value. The other 80% comes from workflow redesign, governance, reskilling, and measurement. If you buy seats and change nothing about how work flows, you have bought the 20%.
Missed Part 1? It covers the decisions that come before the purchase: company-wide vs. department-first rollout, why company-administered accounts beat expensed personal subscriptions, and how to write a data boundary policy people will actually follow.
A note on the data: much of the published research on AI adoption comes from vendor-adjacent sources with an interest in the adoption story. Treat directional trends as reliable and specific percentages as soft. Pricing was current as of July 2026 and changes frequently — verify before you sign.