In November 2022, processing one million tokens through a model at GPT-3.5's capability level cost around $20. By late 2024, the same performance was available for $0.07. That is a 280x reduction in roughly two years — a pace of cost decline that makes Moore's Law look leisurely. Gartner projects that AI inference costs will continue dropping through 2030 as hardware efficiency improves and model architectures become leaner. The commodity tier of AI computation is approaching free.
This should mean AI is getting dramatically cheaper to run for businesses. In many cases it is. But the businesses actually capturing that benefit are in the minority. Thomson Reuters' 2026 AI in Professional Services Report found that while organization-wide AI usage nearly doubled from 22% to 40% in a single year, "adoption has hit critical mass — but now come the tough business questions." CIO.com framed 2026 as "the year AI ROI gets real" — and not as encouragement.
The paradox is real. Infrastructure costs fell. ROI expectations weren't met. If the inputs got cheaper and the outputs got better, why are so many businesses reporting that AI isn't working the way they expected? There are two answers, and they operate at different levels. One is about how businesses are buying AI. The other is about what they're doing with it.
The Buying Problem: Paying Champagne Prices for Sparkling Water Work
The AI market in 2026 has a significant tiering problem — but most of the tiering is happening on the vendor side, not the buyer side. Model providers have organized themselves into clear price-performance tiers. The cheapest models — Claude Haiku, GPT-4o mini, Gemini Flash — handle routine language tasks at fractions of a cent per execution. The premium models — Claude Opus, GPT-4o, Gemini Ultra — are priced at multiples of the commodity tier and are designed for complex reasoning, nuanced judgment, and tasks where output quality has substantial downstream consequence.
The buying behavior of most businesses does not match this architecture. Companies subscribe to premium model tiers as their default — or purchase enterprise AI agreements without specifying which tasks route to which model — and end up paying premium prices for work that a commodity model handles just as well. Formatting a client email. Classifying an inbound inquiry. Extracting structured data from a consistent document template. Generating a summary from structured meeting notes. These tasks do not require frontier reasoning capability. Running them through a premium model is like hiring a senior strategist to file your receipts. Correct outcome, wrong price point, wrong resource allocation.
This problem compounds quickly when AI gets deployed at scale without governance. Only 8% of organizations currently maintain a comprehensive AI governance framework, according to AI governance research published in 2026. That means 92% of organizations using AI have limited visibility into what their AI systems are actually doing — which models are running, on what tasks, how often, at what cost, and whether the outputs are being used. Without that visibility, there is no basis for a rational buying decision. You are paying for what someone guessed you needed rather than what the actual usage patterns show.
The Deployment Problem: Running Expensive Tools on Cheap Tasks
The model tiering failure has a structural cause. Most AI deployments are organized around tools, not tasks. A company adopts ChatGPT Teams, or buys Claude for Work, or stands up a Copilot deployment — and the deployment gives access to the premium tier to everyone for everything. The model becomes the default response to any task that touches a keyboard. This is not a strategy. It is an access policy masquerading as one.
The businesses getting better returns have organized AI deployment around task categories, not tool access. They have done the unglamorous work of cataloguing their recurring tasks, assessing what each one actually requires in terms of reasoning, judgment, and output quality, and routing each category to the model tier appropriate for it. Commodity tasks go to commodity models. Premium model time is reserved for tasks where the quality differential is measurable and consequential.
This is not a technical argument. The commodity models — GPT-4o mini, Claude Haiku, Gemini Flash — are genuinely capable for a very wide range of business tasks. They are not slower, less safe, or less reliable for routine work. They are simply less expensive because they sacrifice some ceiling performance in exchange for efficiency. For tasks that do not approach that ceiling, paying for the ceiling is waste. Most business tasks do not approach the ceiling.
The practical test for any task is straightforward: could a competent but not exceptional human complete this in under fifteen minutes using standard knowledge and a consistent process? If yes, a commodity model can almost certainly handle it. Could the same human get it meaningfully wrong in ways that would be difficult to detect without expert review? If yes, that is where premium model time earns its price difference.
The Deeper Problem: 80% of AI's Value Has Nothing to Do With the Model
Here is what makes the buying and deployment problems frustrating: fixing them is the easy part. You can fix a model tiering mistake in a week. The harder problem is the one most businesses haven't confronted yet.
NTT Data's 2026 Global AI Report made a finding that should change how founders think about AI investment: technology accounts for roughly 20% of an AI initiative's realized value. The remaining 80% comes from how organizations redesign work around it. Deloitte's 2026 State of AI report corroborates this from a different angle: only 34% of organizations are "truly reimagining their business" with AI. The remaining 66% are in what researchers describe as efficiency mode — doing existing tasks faster and cheaper.
Efficiency mode produces real gains. If you automate five hours of administrative work per week, that is five hours you did not need to pay for or personally spend. The math works. But efficiency mode captures the 20%. Work redesign captures the 80%. And the gap between them is not incremental — it is the difference between a business that is marginally cheaper to run and a business whose operating model looks fundamentally different from what it looked like two years ago.
Work redesign means asking a different question than "how can I do this existing task with AI?" It asks: "if I were starting this business today, knowing what AI can do, would I design this workflow the same way?" In most cases, the answer is no. The workflows that exist in most businesses were designed for human cognitive constraints — tasks were batched because context-switching was expensive, reports were produced weekly because gathering the inputs manually took a day, proposals took three days because the research required hours of senior attention. Those constraints no longer all apply in the same way. The workflow shape that was optimal for a world without AI is not necessarily the optimal shape for a world with it.
The businesses in Deloitte's 34% are not just adding AI to existing processes. They are rebuilding the processes themselves, using AI-native approaches — continuous instead of batched, personalized instead of templated, research-enabled instead of intuition-dependent — and finding that the workflow redesign captures value that the original process structure was never going to release, regardless of which model was running it.
A Three-Tier Model for AI Task Routing
The practical starting point is a simple tiering framework that matches task type to model tier. This is not a technical framework — it requires no engineering. It is an operational discipline that any founder can implement in a week.
Tier 1: Commodity tasks. These are high-volume, low-judgment, structurally consistent operations where correctness is binary — either the output matches what was asked for or it doesn't. Classifying inbound emails by type. Extracting specific fields from completed forms. Reformatting data between systems. Generating summary bullets from a structured meeting transcript. Checking a document against a checklist. These run on the cheapest available model — GPT-4o mini, Claude Haiku, Gemini Flash — at costs measured in fractions of a cent per task. Volume is irrelevant because unit cost is negligible. Governance here means sampling outputs periodically to confirm the error rate stays within tolerance, not reviewing each output.
Tier 2: Standard tasks. These require coherent multi-step reasoning, tone judgment, or contextual adaptation — but operate within a domain the model knows well and produce outputs a competent reviewer can evaluate quickly. Drafting a first-version client proposal using a structured brief. Generating a first-pass response to a complex inbound inquiry. Synthesizing a set of research inputs into a structured memo. Producing a first draft of a content piece from an outline and source material. These land in the mid-tier — Claude Sonnet, GPT-4o, Gemini Pro — and warrant a human review step before delivery. The model does the execution; the human validates the judgment calls and catches domain errors before they reach the client.
Tier 3: High-judgment tasks. These involve genuine strategic reasoning, nuanced evaluation against criteria that require expertise to apply, or outputs where an error has meaningful downstream consequence. Diagnosing why a business system isn't working and designing a structural solution. Evaluating whether a proposed approach has the right risk profile for a specific client context. Synthesizing competitive intelligence into a market positioning recommendation. These warrant premium model time and, always, expert human review. Premium models are not always right here — but they produce noticeably better raw material for the expert reviewer to work with, which means less correction time and lower overall cost when the judgment stakes are high.
The actual implementation is a routing rule, not an architecture: before running any recurring AI task, assign it to a tier and select the model accordingly. For teams using Make, n8n, or Zapier, this means separate workflow branches for each tier rather than a single flow that sends everything to the same model endpoint. For founders running AI manually, it means a simple decision habit: pause before prompting and ask whether this task actually needs the premium model, or whether the commodity model is the appropriate tool.
The Workflow Redesign That Unlocks the 80%
Model tiering gets you better economics on the tasks you're already running. Work redesign is what gets you to the 80%. Here is what that looks like at founder scale for three common workflows.
Client reporting, redesigned. The legacy approach: pull data from your project management tool, compile it into a document, write the narrative, review, send. An AI-assisted version of that same process saves maybe thirty minutes. The redesigned approach: your project management tool (Notion, Asana, Linear) maintains a structured data layer that an automated workflow queries every Friday, generates a client-ready status summary using a Tier 2 model and your documented reporting format, routes it to you for a ten-second sanity check, and sends. The original task took ninety minutes. The redesigned version takes ten seconds of your time. More importantly, every client gets their update on the same day at the same interval, without it depending on your workload that Friday. Consistency and quality both improve at the same time cost goes down.
Lead qualification, redesigned. The legacy approach: review each inbound lead, decide if it fits your ICP, write a first-touch response or pass. The redesigned approach: a Tier 1 model runs initial classification the moment the form submission arrives, scoring it against your explicit ICP criteria (which you have written down in a structured format). High-score leads route to a Tier 2 model that drafts a personalized first-touch message using your context library and the lead's stated need, queued for your thirty-second approval before sending. Low-score leads route to a graceful "not the right fit right now" sequence. You spend your attention on the conversations that warrant it. The commodified initial review happens in seconds, consistently, at any time of day. What previously took you two hours of intermittent attention on a Monday morning now takes you five minutes to approve a batch of pre-drafted responses.
Competitive intelligence, redesigned. The legacy approach: when you need to know what competitors are doing, you or someone on your team spends half a day pulling information and synthesizing it into a brief. The redesigned approach: a standing workflow runs weekly, pulls structured inputs (pricing pages, job postings, product updates, review sites) through a Tier 1 model for initial extraction, aggregates the signal into a structured data layer, and generates a weekly delta summary through a Tier 2 model showing what changed. You review a one-page diff instead of conducting research. What was an intermittent and expensive project becomes a continuous and nearly free background operation. The business knows more, more consistently, for a fraction of the prior cost.
In none of these cases did the redesign require significant technical investment. The tools involved — Notion, Asana, Make, n8n, Zapier, the OpenAI or Anthropic API — cost less than a part-time employee's monthly income. What the redesign required was willingness to question the workflow shape that existed before AI and rebuild it from the tasks-that-actually-need-to-happen up, rather than adding AI on top of a process designed for a different environment.
The cost collapse is real. The firms that are capturing it are the ones who noticed it came with a requirement: not just to adopt AI, but to rethink the structure of how work gets done. Most firms are still treating AI as an accelerant applied to existing workflows. The accelerant is cheap. The redesign is where the returns are.
Not sure which of your workflows warrant a redesign versus just an AI accelerant — or how to audit your current AI spend against what you're actually getting? That is the conversation we have before we build anything. Talk to us. We'd rather spend an hour mapping your actual workflow structure than watch you pay premium model prices for commodity tasks for another six months. We'd rather tell you no than waste your money.
Related: The AI Winners Aren't Cutting Costs. They're Entering New Markets. | You're Paying for AI. Do You Know If It's Working?