Published: October 25, 2026 | Category: AI | By Mahesh

AI native finance has two jobs that most coverage mixes together. The first is running the finance function with AI agents that reconcile accounts, flag anomalies and draft forecasts. The second is accounting for a company whose product runs on inference, where every customer request creates a real cost of revenue. Both belong in this cluster article under Depth Grid's guide to the AI native company, because a company is only what an AI native company is if its books can show what the AI costs and what it earns.

The numbers explain the urgency. ICONIQ Capital data, as reported, puts average AI product gross margin near 52% in 2026 against 75% to 85% for traditional software^[1]. Meanwhile Gartner found that finance AI adoption had stalled at 59% of functions in 2025, after jumping from 37% to 58% the year before^[2]. Finance teams are adopting AI slowly while the finance problem AI creates is arriving quickly.

What Does AI Native Finance Mean?

Direct answer: AI native finance runs the finance function with governed AI agents and treats inference as a variable cost of revenue. It changes two things at once: who does the routine accounting work and how gross margin, pricing and forecasts are built for products that cost money every time they are used.

Traditional software finance rests on one comfortable fact. After the build, the cost of serving one more customer is close to zero, so gross margin sits at 75% or higher and a CFO can forecast with simple growth curves. AI breaks that fact. Each query, each agent step and each retry draws on compute that someone pays for. The finance team inherits a cost line that behaves like a utility bill and not like a salary.

The same technology also arrives inside the finance team. Agents can match invoices to purchase orders, chase exceptions and prepare variance commentary. The CFO therefore plays two roles at once: buyer of AI for the back office and steward of the cost structure of an AI product. Treating these as separate projects is a common mistake, since the same data quality and control weaknesses undermine both.

How Far Along Are Finance Teams?

Direct answer: About 59% of finance functions used AI in 2025 according to Gartner, almost unchanged from 58% in 2024. Agentic AI is earlier still. A May 2025 vendor survey found 6% of finance leaders using it and 38% planning to within a year.

The Gartner 2025 AI in Finance Survey covered 183 chief financial officers and senior finance leaders and ran from May to June 2025^[2]. It found 59% reporting AI use in the finance function, against 58% in 2024 and 37% in 2023^[2]. Gartner's own analyst described the momentum as slower after a sharp rise^[2]. Among those already using AI, 67% were more optimistic than a year earlier^[2]. The 2024 survey had named inadequate data quality and availability as finance leaders' top challenge, followed by low data literacy and technical skills^[3].

Agentic AI shows a steeper curve from a lower base. A CCH Tagetik survey of 392 finance leaders, taken on May 21, 2025, found 6% already using agentic AI and a further 38% intending to adopt it within twelve months, which would put 44% in use by 2026^[4]. That is a vendor survey at a vendor event, so treat the sample as friendly. PYMNTS Intelligence reported that by the start of the third quarter of 2025 more than one in ten companies had begun using or testing agentic AI, while in May 85% of CFOs had said they had no plans to adopt it^[5]. Sentiment moved in about 90 days.

Gartner's guidance on agentic AI in finance carries a plain caution. It says agents can hallucinate, misread data or fail on open-ended tasks, and that human feedback and explainability stay essential^[6]. For an auditor that is the whole story. An agent may prepare the reconciliation, but a person with authority must own the sign-off.

What Does Inference Do to Gross Margin?

Direct answer: Inference averages about 23% of revenue at scaling-stage AI B2B companies, according to ICONIQ. Average AI product gross margin sits near 52% in 2026, compared with 75% to 85% for traditional SaaS. The gap is mostly the cost of running models for customers.

ICONIQ Capital's State of AI work reports inference at an average of 23% of total revenue for scaling-stage AI B2B companies^[7]. It puts AI product gross margin at about 52% in 2026, up from 41% in 2024^[1]. Other coverage of the same research quotes 45% for 2025 and about 53% for 2026 with 59% projected for 2027^[8][9], so the exact figure varies by report edition and readers should cite the range. Bessemer Venture Partners found that fast-scaling AI startups it labeled Supernovas averaged about 25% gross margin early on while steadier Shooting Stars averaged around 60%^[9].

The pattern repeats across sources. One summary says 84% of scaling AI B2B companies see gross margin erosion of six points or more from AI infrastructure costs^[10]. A cost-management study reported by the same research hub found that 94% of companies say they track AI costs but only 34% have mature cost management^[11]. Counting the cost is common. Controlling it is not.

Not every finding points the same way. Benchmarkit reports that median software gross margin stayed above 80% and stable across four years, which suggests AI costs have not yet compressed the median across all software^[8]. The squeeze is concentrated in AI native cohorts. That distinction matters for an investor deciding which benchmark a given company should be judged against.

A worked example shows what the averages mean in dollars. Take $1 million of AI product revenue. At 23% inference, the model bill is $230,000. A 52% gross margin means total cost of revenue is $480,000, so other costs of delivery (hosting, support, data licensing and similar items) must be about $250,000, or 25 points. Gross profit is $520,000. A traditional software firm at an 80% margin would keep $800,000 from the same revenue. On Depth Grid's arithmetic the gap is $280,000 per $1 million, and inference explains $230,000 of it.

What Is the Cost per Successful Task?

Average margin hides the number that decides whether a product survives, which is the cost of one successful outcome. A model call can fail, retry or loop, and the company still pays for every attempt. We call this the cost per successful task: inference cost per attempt divided by the success rate. This is a Depth Grid working metric built on a simple pricing example from SoftwareSeni's analysis of AI pricing^[12], in which a task priced at $0.50 costs $0.10 to run when simple and $0.45 when complex.

Source: Depth Grid calculation. Cost per success = cost per attempt divided by success rate. The $0.50 price and the $0.10 and $0.45 attempt costs come from a published pricing example^[12]; success rates are illustrative.

The last row is the one a CFO should remember. A complex task that earns a 10% margin when everything works loses money once success drops to 70%, because retries push cost per success above the price. Under outcome-based pricing, where the customer pays only for a completed result, failed attempts are a direct loss^[12]. That is why ICONIQ found that 37% of AI companies planned to change their pricing model within twelve months^[12].

The cost per attempt is also a choice. A cheaper model, a smaller routed model or a locally run open model can cut it. Recursive released the open-weight Flint coding model in October 2026 under the Apache 2.0 license, and its figures on speed and benchmarks are self-reported, as we noted in our Recursive Flint coverage. Running such a model on owned hardware turns a per-token bill into a fixed hardware cost, which changes the accounting from variable to fixed. Our Claude Haiku 5.5 launch coverage shows the other route, a low-priced hosted model. Which is cheaper depends on volume, and finance should model both rather than accept either by default.

What Controls Do AI Agents Need in Finance?

An agent that touches the ledger needs the same discipline as a junior accountant plus a few extra rules. Four controls cover most of the risk.

Spend and action limits

Set a hard ceiling on what an agent can approve, pay or post without a person. Put the limit in the system, not in the prompt, so no instruction can talk the agent past it. Gartner's security research makes the parallel point that deterministic controls beat asking AI to police itself^[13].

Segregation of duties

The agent that prepares a payment should not be the one that approves it. Many teams give one agent broad access because it is easier to build. That recreates the single point of failure that finance controls were invented to remove.

Reconciliation with human sign-off

Let agents do the matching and exception triage, and require a named person to approve the result before it reaches the close. The aim is to move human time from keying to judgment, not to remove the person.

An audit trail tied to each agent

Log what the agent read, what it decided and what it changed, under its own identity. Without that, an auditor cannot tell an agent error from a human one, and neither can you.

How Does AI Native Finance Differ From Traditional Finance?

Source: Depth Grid analysis using ICONIQ, Bessemer and Gartner figures cited above.

How Do You Forecast When Costs and Retention Move?

Two inputs make AI native forecasts unstable. The first is cost. Inference prices have changed fast, and a plan built on this quarter's model prices can be wrong by next quarter in either direction. The second is retention. CRV reported a median gross revenue retention of 40% for AI native companies and a median net revenue retention of 48%, against roughly 101% net retention for the broader SaaS market^[14]. The source does not define its cohort, so treat those numbers as a warning about volatility and not as a benchmark to plan on.

The practical answer is to forecast in scenarios. Build a base case, a case where inference cost per success falls 30% and a case where it rises 30%, then run each against a retention range wide enough to cover both the SaaS norm and the AI cohort figures. On a $10 million revenue plan with the 23% inference share, a 30% swing in unit cost moves cost of revenue by $690,000 either way, on Depth Grid's calculation. That alone is large enough to change a hiring plan.

Funding conditions add a third variable. Crunchbase data covered in our Q3 2026 global venture funding report shows AI companies took 64% of global venture dollars in the quarter, which means investors are comparing AI native margins with each other and not with old SaaS. Our comparison of startup funding in India, the USA and Europe shows how those expectations differ by region. A company that cannot explain its inference line will face harder questions in either market.

What Will AI Native Finance Look Like by 2031?

The table below is Depth Grid analysis. It extends the dated figures above and is a reasoned projection, not a measured result.

The most uncertain row is 2030. Companies classify inference differently today, with some placing it in cost of revenue and others in research and development^[9]. Until accounting guidance settles the question, margin comparisons across firms need a footnote that says how each company classifies the cost.

How Does This Play Out in India, the US and Europe?

In the United States, public and late-stage investors judge AI companies against 80% software margins and expect an explanation for the gap, which makes the cost-per-success metric a communication tool as well as a control. In India, many large finance operations sit inside shared service centers that run accounting for global parents, so agent adoption there tends to start with reconciliation and payables and moves quickly once a client approves the control design. Our reading of workforce tools such as the one in SAP Buys TechWolf: What the AI Work Deal Means suggests that the same centers will also face questions about how roles change. In Europe, auditors and works councils are likely to focus on explainability and on who is accountable for an agent's output, so documentation should be ready before the first review.

What Should a CFO Do in the Next 90 Days?

The plan below is Depth Grid's suggestion and assumes a company of any size with at least one AI product or one finance agent in use.

1. Days 1 to 30: measure. Split cost of revenue so inference, hosting and support each have their own line. Compute inference as a share of AI revenue and compare it with the 23% reference.
2. Days 31 to 60: price against cost per success. Measure success rates by task type, compute cost per success and flag every task where it exceeds 60% of price.
3. Days 61 to 90: control the agents. List every agent that can touch the ledger, set action limits in the system, separate preparation from approval and turn on per-agent logging.
4. Then decide the sourcing question: model the hosted option and the owned-hardware option side by side at your real volume and revisit the choice each quarter.

Most companies find that two or three task types create most of the margin loss, so the first pricing change usually needs only a small slice of the product.

What Is Still Uncertain?

Several caveats apply. ICONIQ margin figures reach this article through secondary coverage that quotes slightly different values by edition. The CCH Tagetik agentic survey was taken at a vendor conference. The CRV retention figures come without a cohort definition. The cost-per-success table uses illustrative success rates and a single published price example, so it demonstrates the mechanism and does not predict any company's margin. Inference prices can change quickly, which can make any cost figure dated within months.

The direction is still clear. Finance leaders who measure inference as a variable cost and put agents under ordinary controls will have better answers for boards and auditors, whatever the exact benchmarks turn out to be.

Common Questions

Sources

1. The SaaS CFO, "How to Calculate the Inference Efficiency Ratio," citing ICONIQ Capital's 2026 State of AI snapshot. thesaascfo.com
2. Gartner, "Gartner Survey Shows Finance AI Adoption Remains Steady in 2025," November 18, 2025. gartner.com
3. CFO Dive, coverage of Gartner's 2024 finance AI survey of 121 finance leaders. cfodive.com
4. CPA Practice Advisor, "Finance Leaders Plan to Increase Use of Agentic AI in Next Year," May 28, 2025, reporting a CCH Tagetik survey. cpapracticeadvisor.com
5. PYMNTS Intelligence via CFO.com, "CFO interest in agentic AI surged in Q3 but skepticism lingers," 2025. cfo.com
6. Eastgate Software, "Agentic AI in Finance: What CFOs Need to Know Now," summarizing Gartner guidance. eastgate-software.com
7. SoftwareSeni, "Why AI gross margins are so much lower than SaaS," March 18, 2026, citing ICONIQ's January 2026 State of AI report. softwareseni.com
8. CloudZero, "AI gross margin" guide, citing ICONIQ and Benchmarkit 2026 reports. cloudzero.com
9. Rework, "Why AI Gross Margins Differ From SaaS," citing ICONIQ and Bessemer. rework.com
10. Medium, "AI for Startups in 2026: What Actually Matters Now," citing ICONIQ's January 2026 report. medium.com
11. SoftwareSeni, "What the AI inference cost crisis means for growing software companies," citing Mavvrik and Benchmarkit. softwareseni.com
12. SoftwareSeni, "How to Design AI Product Pricing That Survives Variable Inference Costs," March 18, 2026. softwareseni.com
13. Quisitive, summary of Gartner's January 2026 research "How to Secure Enterprise Agentic AI Ambition." quisitive.com
14. CRV, "AI SaaS vs Traditional SaaS: Why AI SaaS Economics Work Differently," June 8, 2026. crv.com

Related Analysis

• The AI Native Company: The Complete Guide to Building AI-First Businesses in 2026
• What Is an AI Native Company? Complete Definition, Examples and Guide
• Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks
• Global VC Funding Q3 2026: Why Money Is Piling Up
• Fintech in 2026: How Technology Is Permanently Rewiring Banking and Payments

Article by Mahesh | Depth Grid