In the title of a November 3 report by The Register, there are two figures: 90 percent and one-twentieth. In “Open-Weight Models Will Run 90% of Enterprise Agentic Workflow Steps,” two Gartner analysts conclude that nine out of ten steps in agentic workflows are routine work, and that open-weight models, at comparable performance, cost about one-quarter as much as the cheapest closed-source model; if every step had originally invoked a frontier model, the cost can be as low as one-twentieth.
What really made me stop wasn't the amount of savings, but another reminder: an agent task is never a single clever answer, but hundreds of classifications, tool calls, parameter passing, validations, and reconciliations. A company's AI bill is mostly spent on these trivial steps that no one writes into a reporting PPT.
Core argument of this article: What enterprise AI should recalculate is not “which model to buy,” but “how much each step should cost, and whether data crosses the boundary.” Gartner’s strategic planning assumption is that by 2030, 90% of enterprise operational agent workflow steps will run on open-weight models; but saving money is only a side benefit—the main reason is control. The deciding factor is shifting from model procurement to two things—whether the harness can route by step, and whether the contract allows you to shift routine steps away.
1. Ninety percent of the steps are "manual labor"; align the four sets of definitions before discussing the budget
The report was written by Gartner Senior Director Mark Petty and Group Vice President Raymond Paquet, released around October 26, and reported by The Register on November 3. What it provides is not a single number but a set of measures that are easily conflated by the media. Before budgeting, first distinguish clearly:
| caliber | The figures given by Gartner | Nature |
|---|---|---|
| Routine step percentage | 9 of 10 steps are classification, tool invocation, parameter passing, validation, and reconciliation | Analyst observations of typical operational workflows |
| Compared with the cheapest closed-source model (same score) | Open weights at about 1/4 the price | Analyst estimates |
| Compare using frontier models at every step | As low as 1/20 | Best-case scenario (best-case) |
| Actual savings per task (for frontier models) | About 5–6 times, not 20 times | Analyst self-report correction: open models are more verbose and retry more |
| By 2030, 90% of steps run open weights | Strategic Planning Assumptions | Predicted, not measured values |
Don’t put “one-twentieth” into your business case.Gartner itself acknowledges in the report that current open models “may be more verbose and retry more frequently,” so on a per-task basis, the real savings relative to frontier models are close to 5–6x. 5x is still substantial—an agent with a monthly bill of 10,000 yuan can theoretically be brought down to the 2,000-yuan range—but using 1/20 for annual budgeting will definitely leave you slapped in the face by the bill at year-end. This also reminds us that when looking at any AI cost report, first find that “best-case” correction clause.
Gartner strategic planning assumption: by 2030, 90% of enterprise operational agent workflow steps will run on open-weight models (The Register, 2026-11-03; this is a prediction, not a measured value)
Another judgment in the report is more worth noting: models are becoming commodities; what is truly valuable is the harness—the software layer that routes each step to the appropriate model, applies guardrails, and manages a fleet of agents. A good harness saves money and blocks risk without the user noticing; a poor harness leaves the work to people. According to a paraphrase of a Gartner summary, routing through a gateway first can reduce inference costs by about two-thirds; this is paraphrased wording, and implementation should be based on actual measurements (Debugged Pro, 2026-11-03).
II. Saving Money Is Merely Incidental: What Is Really Being Moved Away Is Workflow History
“Years of workflow history are a large part of what acquirers pay for in a SaaS company; every routine step sent to a frontier API moves a small piece of it outside your boundary. What open weights preserve is your control; the money saved is merely a bonus.” — Mark Petty, Raymond Paquet, Gartner, quoted in The Register (2026-11-03)
This passage pulls the discussion back from a financial accounting perspective to an asset perspective. Every validation and exception handling by the agent across ERP, CRM, and MES deposits a record of “how this company actually does things”; sending the entire batch to an external API amounts to permanently relocating process assets outside. For organizations with data residency requirements, this is no longer a cost issue, but an admission issue of whether AI can enter core business.
Accordingly, Gartner assigns the two types of models to two reports: frontier models continue to serve as R&D engines and are more like capital expenditure; open-weight models take on massive stable execution and become operating expenditure. The report recommends that “a small number of highly stable steps that run millions of times per month” prioritize domain-specific open-weight models—exactly the profile of ERP reconciliation, document validation, and material code matching.
Same-week China signal: weight opening and Harness hosting happen simultaneously
In the early hours of November 5, Moonshot AI officially released the complete model weights of Kimi K3.1 Pro: a MoE architecture with 2.8 trillion total parameters, 104 billion activated per inference, a 1 million token context, and native multimodal visual understanding; Ascend announced same-day 0Day inference adaptation, relying on vLLM Ascend and SGLang to support Atlas deployment validation (The Paper, 2026-11-05). On the same day, Kimi also launched Hosted Agents, turning the Harness framework into a hosted service, with the company stating that building a reliable Harness in-house "may take months".
Looking at the two things together, the signal is clear: weights have become a purchasable commodity, and the orchestration layer has become a new product form. This is on the same line as the October 29 “inflection point for enterprises building their own software” and the October 21 “wave of sovereign open-weight models”—this round has advanced from “whether the model can be obtained” to “whether the model can be scheduled step by step”.
Three qualifications that must be clearly stated. First, open weights ≠ open source as defined by OSI: K3.1 Pro has not disclosed its training code and original training data, and commercial use must comply with a custom license (The Paper, 2026-11-05). Second, “data does not leave the domain” has hardware thresholds: according to CSDN DevPress’s compilation based on the Hugging Face model page and vLLM Recipes (2026-11-05), its MXFP4 weights are about 1.56 TB, and deployment starts with 8 Blackwell cards; for small and medium-sized teams, continuing to use the API is often more cost-effective. Third, the Revolut case from the same week (about 2,000 engineers switching to self-hosted GLM, 56 B200s, about $0.30 per million tokens, lead time -70%) comes from a relayed account of a conference presentation, is a single company’s self-report and unaudited, and should not be used as a benchmark.
III. The Real Risk Lies in the Contract: Step-Based Billing Sells You 90% of Routine Work at the Ceiling Price
ERP Leaders' November 3 interpretation for the CFO was sharp: the budget question has shifted from "which model to buy" to "which steps need to be expensive"; and a per-step pricing agreement that also fails to state "the right to move routine work elsewhere" amounts to paying the highest unit price for 90% of routine steps.
This is not hypothetical. In 2026, the billing pendulum has swung from seat-based to hybrid, usage-based, and outcome-based models (following the 08-25 billing restructuring): Salesforce Agentforce charges about $0.10–$0.15 per autonomous action, Intercom Fin charges $0.99 per resolution, and Fini starts at $0.49 per resolved ticket (public pricing pages from vendors in 2026). The model itself is not the problem; the problem is, when an agent runs hundreds of steps a day, who holds the authority to define the billing unit.
| Contract Terms | Content that must be clearly stated | The consequences of not writing clearly |
|---|---|---|
| Billing unit | What counts as one step/one resolution; whether retries, failures, and manual handoffs are billable | Failures and retries are also billed, and the bill scales with the retry rate |
| Measurement and Verifiability | Which system performs the counting, and can customers view and verify it in real time? | In the event of a dispute, only the manufacturer's reports may be accepted. |
| Limits and Alerts | Spending limit, alerts before reaching the limit, behavior after reaching the limit | Sudden surge in usage leads to unexpected quarter-end bill |
| Price adjustment and routing rights | Price adjustment notice period; can routine steps be moved to self-hosted open-weight models? | locked into the unit price of a single frontier model |
| Exit and Remedies | Exit clauses, logs and data export, refunds or exit rights for non-compliance | When model prices rise, you can neither walk away nor have any basis for negotiation. |
The governance side must be accounted for as well. The IDC November 5 roundtable cited the 'Worldwide Agentic AI Adoption Guide' (2026-06): 95.3% of enterprises are already running agents in production environments, about 70% of organizations have at least half of their projects spanning multiple functions, and active agents are expected to grow from 31 million in 2025 to 2.5 billion in 2030; but only 35% operate under the 'agent-assisted decision-making + human approval' model, and 47% list 'lack of formal governance protocols' as their top security concern.
95.3%
Enterprises are already running agents in production (IDC, 2026-11-05, citing the 2026-06 guidelines)
22%
Share of Forrester respondents who saw AI improve EBITDA (2026-10-25)
Data from the same period can cross-calibrate one another: Gartner (11-02) found that 47% of supply chains made AI-related layoffs, averaging 10.5%; Forrester (10-25) forecasts that AI will account for 20% of technology spending in 2027, with only 22% of respondents seeing EBITDA improvement; Gartner (10-22) found that 44% of CEOs find it difficult to prove AI returns (all cited from ERP Leaders, 2026-11-03). Scale is growing, but return evidence has not kept up—'calculating costs clearly step by step' is more urgent than 'deploying another agent.'
IV. Four Practical Recommendations for CIO / CFO
1｜First do a “step inventory,” then discuss model procurement. Choose a high-frequency process and break it into steps, annotating item by item which are classification, validation, reconciliation (open-weight models can handle), and which truly require frontier reasoning. Without this checklist, any proposal to “cut costs by 90%” cannot be verified; start pilots with low-risk, rollback-capable processes, and use your own data to actually measure retry rates and per-step costs.
2｜Write routing capability into the acceptance criteria. Require the harness / gateway to support per-step routing and to be able to observe which model was used at each step, how much was spent, and where the data went; write “routine steps can be switched to self-hosted open-weight models” as a contractual right, rather than a vendor’s future feature.
3｜Three questions for re-examining the existing contracts. Following the checklist that ERP Leaders (2026-11-03) gave to CFOs: What exit clauses do the largest software and services contracts each have, and when can they be used? How much does it cost to leave, and how many months does it take? In a dispute, who reads the contract, and have they read it before? If the answers are vague, fill in the terms first, then scale up the agents.
4｜Governance gate and routing go live simultaneously. Permissions, approvals, auditing, and recall are all indispensable, echoing the three hard constraints of October’s mandatory national standard for agent security (plan number 20263116-Q-252). IDC’s 47% “lack formal governance protocols” are often the first bill to blow up after costs have been saved.
The essence of this change is not that open-source models have finally become cheap enough to use, but that for the first time enterprises are in a position to split the AI bill down to “every step.” Only by breaking it down by step can they know which 90% should stay in their own data center and which 10% is worth paying a premium for, and only then can they know what the “billing by usage” clause in the contract is actually measuring. Getting this account straight is closer to the original intent of digital transformation than signing another more expensive model.
This article is compiled and analyzed based on public information: Gartner, "Open-Weight Models Will Run 90% of Enterprise Agentic Workflow Steps" (Mark Petty, Raymond Paquet; reported by The Register on 2026-11-03); ERP Leaders, "Gartner says cheaper open AI models can do nine in ten AI agent steps" (2026-11-03); cross-check of the report by ExplainX and CyberMaterial (2026-11-03); The Paper, "Moonshot AI's Kimi K3.1 Pro Opens Its Weights, Simultaneously Launches Hosted Agent Service" (2026-11-05); CSDN DevPress deployment compilation based on the Hugging Face model page and vLLM Recipes (2026-11-05); IDC × Google Cloud × Workday roundtable citing "Worldwide Agentic AI Adoption Guide" (2026-11-05; guide, 2026-06 edition); Debugged Pro citing the Gartner summary (2026-11-03); vendors' public pricing pages (Salesforce Agentforce / Intercom Fin / Fini, 2026). Cost ratios and forecasts are cited according to the original wording of the sources; best-case and revised values are marked separately; Revolut data are self-reported by a single company and unaudited. Data as of 2026-11-06.
Zhongdi Information · Specializing in Odoo and digital solutions for enterprise management
