A forward-looking whitepaper for business and technology leaders. Full PDF with all appendices: DOI 10.5281/zenodo.24386748.
Executive decision brief
The next AI breakthrough may be learning what not to compute
For several years, AI progress has come from scale: larger models, longer context windows and more GPUs. That has changed what machines can do. It has also created an uncomfortable economic reality. As AI moves from assistants to agents that read, plan, act and check their own work, the computation behind a single business task can grow sharply. A token gets cheaper, but finishing the task can still take many more of them.
Enterprise AI efficiency is not simply a matter of buying cheaper tokens. It is a matter of allocating computation to the evidence, generation and verification a business task actually needs. This paper argues that enterprises need a control layer that allocates computation according to risk and business value. That layer, not any single model or research technique, is the durable investment.
Fewer billed tokens, or less computation per token?
The next wave of AI efficiency will not necessarily come from consuming fewer billed tokens. It may come from doing less computation on those tokens. New research on selective reading lets a model read every input token lightly and give deep processing only to the tokens that matter. The prompt still contains every token. So a company paying a conventional API price per input token will not automatically see a smaller invoice. A company running its own models can turn the same efficiency into more capacity from the same GPUs. Whether the savings ultimately reach customers will depend as much on commercial pricing as on technology.
What is demonstrated, what is available, what is projected
Headline economics (illustrative)
Section 5 models an AI agent that reviews one million documents a month, with a person correcting any output that fails. The primary metric is fully loaded cost per completed task. First-pass acceptance is reported separately.
API buyer. Caching, tighter retrieval, model routing and a fallible verification loop cut cost per completed task from $0.94 to $0.36 (−62%). That is about $7.0 million a year in gross operating benefit before implementation costs.
Self-hosted operator. Adding selective reading as a compute saving gives $0.94 to $0.29 (−69%), about $7.7 million a year gross.
Range. Across the sensitivity cases, the API reduction runs from 52% to 64%. Rework cost, how many errors the checker catches, and how many it can fix move the result most.
Net of implementation. With illustrative one-time costs of $1.5 million and $0.6 million a year to run, the API case pays back in about three months.
Decisions for leaders
Measure fully loaded cost per completed task, and first-pass acceptance, for the top AI workflows.
Capture the optimizations that work today: caching, retrieval, routing and verification.
Build the control layer so new techniques such as selective reading can be adopted when they pass quality and economic tests, without rebuilding applications.
In contracts, define outcomes precisely, secure access to usage meters, and share in vendors' efficiency gains.
Section 11 shows how Elixir Global builds these principles into its AgentVault™ agent repository and its Migra data migration platform.
1. Where enterprise inference costs arise
A token is a small piece of text, roughly three-quarters of an English word, and it is the unit in which AI is consumed, priced and measured. A 10-page contract is about 5,000 tokens. A 300-page policy manual is about 150,000.
Every AI request has two parts:
Reading (input tokens). The model reads its instructions, the question and any documents supplied. Per token, reading is cheaper and faster than writing, but it adds up: instructions and documents are often re-read at every step of a task.
Writing (output tokens). The model writes its answer one token at a time. Writing is usually priced several times higher per token than reading, and it sets how long the user waits.
Whether an enterprise buys AI through an API (paying per token) or runs its own models (paying for GPUs, power and staff), tokens drive the bill. For API buyers the link is direct. For self-hosted models, every token processed uses GPU time and memory.
1.1 Cheaper tokens, more tokens per task
For some agentic workloads, rising call counts and longer contexts can offset falling token prices. The net effect must be measured workflow by workflow. Each generation of enterprise AI has multiplied the computation behind one request.
Typical orders of magnitude, not measurements.
Figure 1. Where an agent's tokens go · illustrative, using section 5's assumptions
Three habits of today's agents drive the growth:
Re-reading the rulebook. Standing instructions, tool descriptions and policy text, often thousands of tokens, are re-read at every step.
Reading whole documents. Contracts, claims files and code are supplied in full, so every request pays to read pages that do not bear on the question.
Writing out reasoning and double-checking. Accuracy increasingly comes from writing out reasoning, producing several candidate answers and scoring them. All of it is billed as output.
A simple illustration: if the price per token falls 10× while a workflow grows from one copilot call to a 30-step agent, tokens per task can rise about 30×, and cost per task triples. Equally, a well-designed agent can cost less than the copilot it replaces. Architecture, not the rate card, decides which.
2. Available optimizations versus emerging research
Most of the savings available today come from proven techniques. Selective reading is the most promising new research, and it adds to them rather than replacing them.
2.1 Three dimensions of conditional computation
Modern AI efficiency is about making computation conditional: spending it only where it is needed. There are three distinct places to do that.
Mixture-of-experts asks which of the model's specialists to wake up. Selective reading asks which parts of the input deserve deep attention. The two are complementary, and both can sit behind a router that picks the right model in the first place.
2.2 The wider toolkit
2.3 Selective reading: teaching AI what not to read
Today's models give every input token, from a decisive clause to a page number, the same full-depth processing. Selective reading adds a small, trainable gate inside an existing model. All tokens pass through the early layers, which gives every token context from the whole input. The gate then scores each token's importance. Only the highest-scoring tokens, plus protected items such as the question and instructions, continue to the deeper and more expensive layers. Memory for dropped tokens is freed.
Figure 2. Selective reading inside the model · illustrative, about a quarter of tokens kept
The researchers (Stanford University and UC Berkeley, October 2026) tested this on question answering, repository code completion, summarisation, biomedical extraction and find-the-needle retrieval, with inputs up to 128,000 tokens in some settings. They used two small open models (about 1.7 and 2 billion parameters), each adapted with a gate and light fine-tuning per task.
Operating points shown in the researchers' summary figure, at accuracy matched to standard reading. They are not end-to-end enterprise benchmarks.
Across tasks, the researchers report 1.4–11.3× faster first responses and 1.5–3.4× lower peak reading memory at matched or near-matched accuracy on selected operating points. The largest gains come on the longest inputs. More aggressive settings buy further speed at some cost in accuracy, for example on multi-hop questions and multi-document extraction. The code and checkpoints are public under an open-source licence, which makes pilots cheaper than an architectural rewrite.
2.4 How the techniques combine inside an agent
In a 12-step agent that re-reads instructions and documents at every step, each technique has its natural target:
Standing instructions and tool schemas repeat identically. Caching handles them better than learned pruning.
The variable document span is fresh for each request and often long. That is where selective reading adds the most.
Simple steps such as classification or formatting go to small models through routing.
Savings are reinvested in verification or extra candidate answers where risk is high.
The gains do not simply multiply. Retries and escalations can erase them. That is why the control layer in section 4 escalates instead of gambling, and records every retry.
3. Evidence and limitations
The research shows a better trade-off between quality and efficiency on small models. It does not yet show enterprise savings or production readiness. Every quantitative claim in this paper carries one of four labels.
3.1 What remains unproven for selective reading
Scale. Results are on 1.7–2B-parameter models. Transfer to frontier-scale models is unproven.
Production serving. Gains were measured outside production serving engines. Selective reading needs custom attention kernels that can gather non-contiguous tokens. Without optimised kernels, that overhead can cancel the gains at small batch sizes. Continuous batching, high concurrency and tail latency (p95) must be measured.
Transfer across domains. Each gate was adapted to its task. A general-purpose gate that transfers safely to new document types has not been shown.
Rare decisive evidence. One negation, exception clause or account number can decide an outcome. Find-the-needle tests check retrieval of one planted fact. They do not fully cover evidence spread across a document, or multi-step reasoning.
Reading order. In today's models, a token's score can only use what came before it. Placing the question before the document helps the gate.
A floor on savings. Early layers always read every token, so speed-ups cannot exceed the ratio of total depth to gate depth.
3.2 What the economics depend on
Billing. API buyers benefit from selective reading only if providers pass internal savings through in price. Self-hosted operators capture the benefit as GPU capacity.
Amdahl's law. A 10× faster reading stage gives far less than 10× end to end if retrieval, tools, generation or human approval take most of the time.
Quality. Any efficiency method that raises the error rate can lose more in rework and risk than it saves in compute.
Energy. Less computation per request can improve energy efficiency per task. Actual energy reductions need measurement, and growing AI demand may offset them.
4. Adaptive inference architecture
The durable investment is a control layer that decides, for every request, how much to read, which model to use, how carefully to check and when to involve a person. New techniques then plug in as options, without rebuilding business applications.
Figure 3. Adaptive inference platform · 5 layers, escalation and feedback loops
4.1 Design rules
Escalate, don't gamble. Try the cheapest safe approach first. If checks fail, widen the reading budget, switch to full reading or a larger model, or generate more candidates, and only then involve a person. Log every escalation, so the business can see when retries are eating the savings.
Protect what you can identify; widen when you can't. Protect known critical regions and task scaffolding: the question, instructions, output schemas, defined terms, amounts, identifiers and known exception clauses. Where decisive evidence cannot be identified reliably, keep broader context or use the full-reading path. Protecting the word "not" is not the same as preserving its scope, so protected zones cover whole clauses, not keywords.
Treat verification as fallible. A checker can miss errors, raise false alarms, and share blind spots with the model it checks. Design for that (section 4.3).
Insulate applications. Applications call the platform, not a specific model. When a technique matures, it is switched on inside Layer 3 and applications keep working.
4.2 Reference deployment
Every AI request, whatever channel it comes from, passes through one gateway. Finance, risk and IT get a single place to set budgets, enforce data rules and measure outcomes. The model pool can mix cloud APIs and on-premises GPUs behind the same orchestration, which makes the hybrid choices in section 8 possible.
Figure 4. Reference deployment architecture · from channels to systems of record
4.3 The verification loop
Verification detects problems. It does not fix them by itself. The loop is: check, detect, retry or repair, re-check, and send what remains to a person.
Figure 5. Verification loop · illustrative rates from section 5
Four quantities decide whether the loop pays:
Recall: the share of real errors the checker catches.
False-alarm rate: correct outputs flagged by mistake, which cost a needless retry.
Fix rate: the share of caught errors a retry or repair actually corrects.
Correlation: when the checker and the model share blind spots, recall falls on exactly the hardest cases. Use different models, deterministic rules or independent evidence checks where possible.
5. Reproducible economics
In an illustrative document-review workflow, proven levers cut fully loaded cost per completed task by about 62% for an API buyer. Adding selective reading raises that to about 69% for a self-hosted operator. Every number below can be rebuilt from the stated assumptions (Appendix A). The figures are a scenario, not a measured result.
5.1 Metrics
The primary metric counts every task that is eventually completed, including those a person corrects:
First-pass acceptance, the share of outputs accepted without human correction, is reported alongside it. Dividing total cost by first-pass acceptances instead gives a higher figure, cost per first-pass accepted output. It is useful for tracking automation, but it should not be quoted as operating-cost savings.
5.2 Scenario and assumptions
An insurer, bank or legal team runs an agent that reviews documents (claims, contracts or loan files) and drafts a recommendation that a person approves. Outputs with errors are corrected by that person.
5.3 Case A: API buyer
API buyers pay for every token sent, so selective reading earns them nothing unless the provider passes the saving through in price. It is excluded here.
Figure 6. Cost bridge, API buyer · illustrative model
The savings are sequential: each lever is measured after the ones before it, so its share depends on the order applied. Applied first, routing would save more, and caching less. Treat individual lever values as indicative. The combined total does not depend on the order.
5.4 Case B: self-hosted operator
For a self-hosted operator, the same unit prices stand in for the GPU cost of processing a token. In practice, measure GPU time, memory and utilisation directly. Selective reading now saves compute on the document span.
The self-hosted operator sees the benefit as capacity: the same cluster handles more tasks, or the next GPU purchase is deferred. That only becomes cash savings if the capacity would otherwise have been bought.
5.5 Sensitivity
The result stays material across these cases, between 52% and 72%. The verification assumptions matter most, so measure recall and fix rate on your own data before you commit a business case.
5.6 From gross benefit to net benefit
The scenario produces about $7.0 million (API) to $7.7 million (self-hosted) a year in gross operating benefit. That is before implementation costs, extra platform overhead and transition costs. The implementation figures below are illustrative assumptions.
5.7 What the model teaches
Quality is the largest single lever. In the API case, a verification loop costing under two cents per task removes 22 cents of human correction.
No single token lever wins. Caching, retrieval and routing each save 10–15 cents. Together they cut AI cost by about 70%.
Selective reading is a capacity lever first. It pays most where documents are long and the operator owns the compute.
Speed is a second dividend. If each of 12 steps waits 3 seconds for its first token and selective reading cuts that below 1 second, each task finishes about 25 seconds sooner. (Illustrative.)
6. Use cases, and when full reading is still preferred
Each lever's value depends on a workflow's token profile: whether it mostly reads, mostly writes, or repeats many steps. Just as important is knowing when to switch efficiency off.
Two patterns stand out. Read-heavy, decision-light workflows benefit most from selective reading: claims, contracts, code and maintenance. Exactness-critical workflows should spend more tokens, not fewer: reconciliation, regulatory filings and safety procedures get full reading plus deterministic checks. The policy layer makes that choice explicit instead of leaving it to chance.
7. Pilot criteria and governance
Adopt an efficiency method only when it passes quality, latency and cost tests on your own workloads, under realistic load, with thresholds agreed before the pilot starts.
7.1 A 24-month plan
7.2 Go/no-go criteria
Every threshold must state its measurement unit (per field, per output or per task), its sample size and confidence interval, and the workflow's risk tolerance. Compare against the incumbent, not against zero.
7.3 Planted-evidence protocol
Adversarial inversions. Insert a contradicting clause (for example, "Notwithstanding Section 4.2, indemnity liability is zero") at 10%, 50% and 90% depth in long agreements. If the efficient path drops or misreads it, the test fails.
Protected zones. Define what is never pruned: the question and instructions, output schemas, structural delimiters, function-call arguments, section headings, numeric spans, defined terms and whole clauses known to carry exceptions.
Distributed evidence. Include questions whose answer needs facts from several sections, not just one planted needle.
7.4 Role by role
8. Procurement: how AI products and agents will be priced
AI pricing is moving from paying for access (seats), to paying for activity (tokens and credits), to paying for results (outcomes). Most buyers will see hybrids of the three. Where the AI runs decides who captures efficiency gains.
8.1 Where pricing stands (October 2026)
Outcome pricing for AI agents is already established in customer service. Prices below were checked against each vendor's official pricing page on October 31, 2026. Definitions differ, and that matters as much as the price.
The spread shows the tension in the market: Salesforce charges for every conversation, HubSpot and Zendesk only for resolutions, and Fin for a broader list of outcomes. Analysts expect the shift to continue. Gartner, as reported by Deloitte, projects that at least 40% of enterprise SaaS spend will move toward usage-, agent- or outcome-based pricing by 2030.
8.2 Pricing by deployment model
Figure 7. Pricing architecture · cloud, SaaS and on-premises under one hybrid control layer
8.3 Will SaaS pricing change?
Yes, for products where agents do work people used to do. Much less for systems of record.
Seats stay for people who use the interface.
Agents get their own unit: per resolution, per document processed, per case closed or per credit. As agents take on work, seat counts shrink and agent units grow.
Systems of record change least. ERP, HR and CRM platforms are likely to keep subscriptions and add consumption fees for AI features.
Outcome guarantees become a selling point, at a higher price per outcome.
An enterprise's own cost per completed task is the right benchmark for buy-versus-build. Vendor prices of $0.50–$2.00 per resolution include margin, platform and support, and they cover a different workload from section 5's model.
8.4 What to negotiate
A precise outcome definition: what counts, over what window (for example 72 hours), how reopened cases are treated, and who decides disputes.
Access to the meter: usage and outcome logs you can audit against your own records.
Caps, floors and true-ups, so usage or outcome pricing stays within budget.
Shared efficiency gains: scheduled price reductions as the vendor's unit costs fall.
Portability: the right to move agents, prompts, evaluation sets and data between cloud, on-premises and other vendors.
9. Outlook 2027–2030: three scenarios
Rather than a single forecast, plan for three scenarios that can unfold together. The same control layer is the right investment in all three. None depends on another: outcome pricing, for instance, does not need selective computation to arrive first.
Figure 8. Outlook scenarios · not forecasts; they can unfold together
9.1 Five questions for the next cycle
Will token pricing survive? Likely yes for model infrastructure, alongside growing usage-, workflow- and outcome-based pricing in applications.
Will models become smaller and more specialised? Routing and selective computation make specialised models more economic for well-defined tasks. Frontier models remain the fallback for hard cases.
Who captures the efficiency gains? Cloud providers, model developers, platform operators and SaaS vendors each capture a share, depending on contracts and on where the AI runs.
Will inference matter more than training? As agent workloads grow, inference efficiency becomes a larger competitive factor for vendors and a larger cost line for buyers.
Will verification become its own value layer? Trusted outcomes need evidence, deterministic checks and human oversight, however fast generation becomes. Verification is likely to be bought, built and audited as a distinct capability.
One research direction deserves watching but is not assumed here. Models that do more of their reasoning in a compact internal form, rather than writing it out as billable tokens, could shrink output costs. That is a hypothesis that needs evidence, not a capability to plan on.
10. Risks and potential mitigations
The main risk of token efficiency is false economy: savings that disappear in errors, retries and rework. Each risk below has potential mitigations. A mitigation is not proof that the remaining risk is acceptable; that has to be tested per workflow.
11. How Elixir Global applies these principles
At Elixir Global, token optimization is designed into our platforms from the start, not added after the bill arrives. Two platforms show how the control-layer approach in this paper works in practice.
AgentVault™ is Elixir's repository of more than 500 enterprise agents. They cover supply chain and SAP processes such as IDoc handling and invoice processing. The agents are model-agnostic: each step can use the smallest model that meets its quality bar, and newer models can be adopted without vendor lock-in. They run within governance guardrails, including SOX and FedRAMP controls where required, and integrate with SAP Joule.
Migra is Elixir's AI-native data migration platform, deployed in client landscapes. It moves master data from many heterogeneous sources through a governed Snowflake hub into SAP S/4HANA, in five stages: ingest, enrich, publish, reconcile, report. Its knowledge base, rules engine and self-learning mean that a mapping resolved once is reused rather than worked out again by a model. Business, IT and audit approvals, plus post-load reconciliation, confirm each outcome before sign-off.
The result for clients is the same discipline this paper recommends: cost measured per completed business outcome, computation spent where risk and value justify it, and an architecture ready to adopt techniques such as selective reading once they pass quality and economic tests.
Conclusion
The first era of enterprise AI proved that machines can produce useful intelligence at scale. The next era will be defined by how deliberately that intelligence is spent. Tokens will not disappear: they remain how models read, write and get billed. But treating every token as equally important is ending.
What will make an AI strategy credible to a CFO, CTO or risk committee is not a bigger savings headline. It is a cost model they can reproduce, assumptions they can challenge, and adoption gates that hold even if research gains do not fully reach production. Organisations that build the control layer, measure cost per completed task, and adopt new techniques only when they pass those gates will get more trusted value from every unit of computation.
Appendix A: Calculation detail
All figures are per initiated task; every task is eventually completed. Prices: reading $2.00 and writing $8.00 per million tokens; cached reading $0.20 per million.
A.1 Tokens and cost per model call
AI cost per task = 12 calls × cost per call. Baseline: 12 × [(20,000 × $2) + (600 × $8)] ÷ 1,000,000 = $0.5376.
A.2 Verification loop (API case)
A.3 Totals
Annual gross operating benefit = (baseline monthly total − case monthly total) × 12: API $6.97M; self-hosted $7.73M. Retry calls are assumed to run on the large model, without routing. The self-hosted case uses API unit prices as a stand-in for GPU cost per token.
Further reading
- Download the full whitepaper (PDF): The Token Economy: How Smarter AI Computation Will Reshape Enterprise Costs and Value (Zenodo, DOI 10.5281/zenodo.24386748)
- Related white paper: Governing Rogue AI Agents with Valid Credentials (DOI 10.5281/zenodo.24378767)
- Related article: The Rogue Agent Will Have Valid Credentials: How CTOs and CAIOs Can Govern Agentic AI
About the author
Ganesh Raju is Chief AI & Innovation Officer (CAIO) and CTO at Elixir Global, an AI-first enterprise consulting firm headquartered in Atlanta with offices in Houston, Dubai, Bengaluru and Singapore. He leads enterprise AI strategy and AgentVault, Elixir's portfolio of governed, production-ready AI agents for SAP and enterprise operations, and the Migra AI-native data migration platform.
He brings more than 15 years of experience in artificial intelligence, machine learning and data engineering. As Head of Big Data, Data Science & AI at Linaro he led Arm enablement of the open-source big data and data science ecosystem, served as ODPi PMC member and TSC lead and as a Linux Foundation TSC member, and led the Apache Bigtop community. He has contributed to TensorFlow, Apache Spark and Hadoop.
He is the author of AI Blockchain Design and Deployment (BPB Publishing), founder and CEO of Akshaya.io, founder of RapidEVChargE, an investor, and an advisor to Canon. He is a former Vice Chair of IEEE Houston and a participant in the Wharton Executive Education Chief Technology Officer program. He is based in Houston, Texas.
Connect on LinkedIn · GitHub · Blog
Views are the author's own. Elixir Global offers products and services related to the subject of this paper. Economic figures are illustrative scenarios, not measured results.


































