Bain: The Models Got Cheaper. Your AI Bill Didn't.

Via Bain & Company: How Token Economics Will Change Opex

Cheaper tokens did not make AI cheaper

Bain & Company’s 2026 analysis of token economics lands on an uncomfortable truth for every owner writing AI checks: the opex mix is shifting from headcount to tokens, but total spend is not falling the way headline model price cuts suggest. In software engineering — the function furthest along on AI enablement — token spend is still only about 1–2% of headcount cost. Across sales, support, strategy, and operations, Bain sees firms at roughly 1% today and talking about 20–30% tomorrow.

That is not a forecast you can ignore because you are “too small.” Owner-led professional services firms already pay for multiple AI subscriptions, API overages on automation tools, and Copilot-style add-ons per seat. The bill arrives as a dozen line items nobody reconciled — which is how Bain describes the general manager’s problem at enterprise scale, minus the procurement department to absorb the shock.

Bain’s bottom line: this is not a budget problem. It is a structural transformation. The economics are unsettled, nonlinear, and already punishing firms that upgraded to frontier models everywhere because the pilot felt brilliant.

The cost-per-task paradox owner-led firms already feel

Bain identifies three bottlenecks at three levels. CEOs see the destination — AI helping humans, not the reverse. General managers see budget and speed: token spend does not fit an existing line item, and scaling a pilot means navigating approvals built for annual SaaS renewals, not weekly capability shifts. Individual contributors see a chasm: early data suggests the top 5% of users in a company may consume more tokens than the other 95% combined.

Sound familiar? It is the partner who live in Claude all day, the admin team still on free tiers, and the owner discovering three overlapping writing tools on the corporate card.

Here is the paradox Bain stresses: in some workflows, agent and token costs are already more expensive than offshore human labor — sparse domains, but real. Agents burn tokens on multistep reasoning, error correction, and context loading. Speed and quality gains can be genuine while per-task economics fail, especially when every task runs on a frontier model or agents orchestrate tool calls nobody scoped.

Bain cites Wells Fargo data (October 2025) and its own analysis: average cost per token fell by roughly half from December 2024 to December 2025, while tokens consumed grew 4.5× over the same period. Net effect: the bill stays high. Three forces explain why — everyone upgrades to the new frontier instead of pocketing savings on last generation; tokens per query climb as agents take multistep work; and once a team finds one useful workflow, they find ten more.

Five moves Bain says hold up — scaled for a 25-person firm

Bain’s prescription is written for enterprises. Owner-led firms should steal the logic without the bureaucracy:

1. Create a dedicated AI compute budget. Stop hiding token and subscription spend inside “software” or “marketing tools.” Bain compares this to cloud migration lines in 2015: protected, visible, governed. If the owner has to find budget mid-quarter from margin, the pilot dies in the next review.
2. Portfolio your model spending. Frontier for high-stakes work; last generation or open-weight for volume. Bain sees three to five times cost differences between firms that model-match and those that run everything on the most powerful option. Pull last month’s invoices: how many calls actually needed frontier reasoning?
3. Instrument cost per task. What does it cost to generate a proposal, resolve a Tier-1 client question, or draft a contract section? Bain warns this is painful early and nearly impossible to retrofit. Start with your three highest-volume workflows.
4. Plan for dual-cost transition. You may pay both legacy labor and rising token bills during overlap. Build that into expectations before the CFO — or you, wearing the CFO hat — discovers it from the bank statement.
5. Evaluate on-premise or open-weight for volume. Bain flags this as the biggest unlevered option for high-volume workflows. Most enterprises have not run the math. Most SMEs have not either — but a confidentiality-driven firm might need to.

Bain highlights AT&T as an architectural example: at roughly 8 billion tokens per day, the company reorchestrated so large “super agents” route tasks to smaller worker models rather than pushing everything through frontier — reporting roughly 90% cost reduction and 3× throughput. One design decision, order-of-magnitude economics. You do not need AT&T scale to apply the principle: match the model to the job.

How AgentsROI turns Bain's framework into something you can run

AgentsROI.ai helps owner-led, information-heavy SMEs govern AI spend without hiring a FinOps team. Vendor-neutral by principle: we measure, optimize, and operate what you already use.

Workflow ROI Audit (primary). We map subscriptions, shadow tools, API overages, and the workflows driving usage — then cost per task for the jobs that matter to your P&L. You get the instrumentation Bain says most companies lack, in plain English, sized for a firm that does not have a thousand-seat procurement playbook.

Model Selection & Continuity Planning. Frontier when judgment matters; cheaper adequate models for volume; fallbacks when a vendor reprices, deprecates, or throttles access. That is how you avoid the “upgrade every six months and wonder why the bill flatlined” trap Bain describes.

Managed AI Operations. Monthly monitoring, guardrails, optimization, and ROI reporting so token economics does not decay after the first enthusiastic month. Someone owns the operating tempo besides the owner.

Instrument now — the curve will not wait for your annual budget

Bain closes with homework fit for any owner: pull your top SaaS contracts and token spend to date; instrument one workflow end to end; learn what it actually costs per outcome. That number is the one almost nobody in your organization knows yet. Once you have it, model choice, renewal negotiations, and staffing decisions get easier.

The 70/30 headcount-to-token scenario Bain discusses is a stress test, not destiny. But the direction is clear: AI opex is becoming a first-class line item whether you planned for it or not. Cheaper models will keep shipping. Usage will keep climbing. The firms that win are the ones measuring early.

If your AI bill grew while nobody can explain per-task economics — book a Workflow ROI Audit. Find out where tokens pay for themselves and where you are flying first class for errands that do not need it.

This article summarizes publicly reported information from Bain & Company’s “How Token Economics Will Change Opex” and is for general informational purposes only. It does not constitute legal, tax, financial, investment, security, or compliance advice. AgentsROI.ai is not a law firm, accounting firm, or registered investment adviser. Facts, pricing, statistics, and product capabilities cited here reflect the sources listed at the time of writing and may change. Readers should verify current information independently and consult qualified professionals regarding obligations specific to their industry, jurisdiction, and circumstances — including applicable New York State and New York City requirements. AgentsROI.ai may have commercial relationships with vendors mentioned; where material, such relationships are disclosed. Nothing in this article is an endorsement of any specific AI product, model, or provider.