Agentic AI in Non-Engineering Business Functions

Executive summary

Adoption of AI in non-engineering business functions is broad, but autonomous production is narrow. As of mid-2026, the U.S. Census Bureau's Business Trends and Outlook Survey puts firm-level AI usage between 17% and 20% [1], and the Federal Reserve estimates roughly 18% of firms had adopted AI by year-end 2025 [2]. McKinsey's 2025 survey finds that in any given business function, no more than 10% of organizations are scaling AI agents [3]. The distance between "we use AI" and "an agent autonomously runs this workflow" is the central fact of the topic.

Independent measurement of return is weak and consistent across three large studies. MIT's Project NANDA reported that about 95% of enterprise generative-AI pilots produced no measurable P&L return [4]; McKinsey found more than 80% of organizations see no tangible enterprise-level EBIT impact [3]; BCG found only 5% of companies are "future-built" value generators while 60% are laggards [5]. These are independent of vendor case studies and corroborate one another.

The flagship rollback is recurring. Klarna publicly reversed its 2024 claim that an AI assistant did the work of 700 agents [6], with its CEO conceding that cost-driven AI cut quality and that human support is "the way of the future" [7][8]. The same human-in-the-loop reassertion appears across functions. Gartner predicts 50% of organizations expecting to cut service staff via AI will abandon those plans by 2027 [9], and that over 40% of agentic-AI projects will be canceled by end of 2027 [10].

Where agents work, they are scoped, supervised, and hybrid. Real production scale exists — Sierra reached roughly $100M ARR on outcome-based pricing [11], and Salesforce reports (vendor-sourced) nearly $1.4B in combined Agentforce and Data 360 ARR [12] — but Salesforce's own support runs about 50/50 agent and human [13], and rigorous field experiments show gains concentrate in routine work and novice users: +14% support productivity overall, +34% for novices [14]. Headcount effects are claimed loudly and measured weakly: CEOs attribute cuts to AI (Klarna ~40% [15], Salesforce 9,000 to 5,000 support staff [13]), while independent labor economists find "largely stability, not major disruption" [16], little AI-specific decline in job postings [17], and credible "AI-washing" where AI is cover for over-hiring corrections [18][19][20].

This snapshot draws on 67 distinct sources — 54 independent and 13 vendor or self-reported — anchored by government statistics, peer-reviewed and working-paper RCTs, and analyst forecasts, with vendor metrics held separate and downgraded unless independently corroborated.

Introduction

This snapshot answers one question: what is the actual operational state of agentic AI in non-engineering business functions — sales, marketing, operations, finance, HR, and customer support — separating what is genuinely deployed in production, what is working, and what is being quietly rolled back, from vendor case studies. The scope excludes engineering-team AI tooling, pure consumer products, and speculative future capability.

"Agentic" here means non-trivial autonomous decision-making: systems that take multi-step actions and make decisions, not single-shot LLM prompts or static chatbots. The bar matters because much of what is marketed as "agents" does not clear it, and because the loudest evidence in the market — vendor case studies and CEO statements — is the evidence most subject to selection and incentive bias. Two recurring framings organize the body: the deploy–reassert cycle, in which an ambitious autonomous deployment fails on the complex tail and walks back to human supervision, and the inverse relationship between an agent's autonomy and the cost and regulation of its errors.

The report covers the weak independent return signal first, then walks each function in descending order of deployment maturity — customer support, sales and marketing, finance and HR, operations — before turning to the reliability and security ceiling that explains the rollbacks, the bounded productivity gains that complicate the failure narrative, the contested headcount effects, and the patterns that recur across all six functions.

The independent return signal

Genuine value capture is concentrated in a small minority of firms despite near-universal experimentation, and this is the most robust finding in the topic because it triangulates across three large independent studies plus government data.

MIT's Project NANDA "State of AI in Business 2025" report concluded that about 5% of AI pilot programs achieve rapid revenue acceleration while the vast majority stall with little to no measurable P&L impact — the widely cited ~95% framing [4]. The study drew on 150 leadership interviews, 350 employee surveys, and analysis of 300 public deployments, and attributed failure not to model quality but to a "learning gap": most systems do not retain feedback, adapt to context, or improve over time [21]. NANDA also found that buying tools from specialized vendors succeeded about 67% of the time, roughly triple the rate of internal builds [4] — a structural argument for procurement over in-house construction.

McKinsey's 2025 global survey corroborates the gap from a different sample: 88% of organizations report regular AI use in at least one function and 72% use generative AI, up from 33% in 2024, yet more than 80% say their organizations aren't seeing a tangible impact on enterprise-level EBIT [3]. For the agentic layer specifically, McKinsey found only 23% of organizations are scaling any agentic system, and in any given business function no more than 10% [3]. BCG's study of 1,250 executives adds a third confirmation: only 5% of companies globally are "future-built" value generators, 60% are laggards reporting minimal gains [5], and agent deployment is concentrated among leaders — 33% of future-built firms use agents versus 12% of "scalers" and almost none of laggards [22].

Government statistics ground the base rates. The Census Bureau's Business Trends and Outlook Survey put firm-level AI usage between 17% and 20% in late 2025 and early 2026 [1], and the Federal Reserve estimated about 18% of firms had adopted AI by year-end 2025, even though roughly 78% of the labor force works at firms that have adopted something [2]. Adoption is real but shallow and employer-size-skewed, and the autonomous-agent layer sits well below it. The counterweight is that rigorous experiments do find real gains, but bounded ones, which makes the "no measurable P&L" result a story about organizational integration rather than model incapability.

Customer support deployments and rollbacks

Customer support has the deepest agentic deployment and, at the same time, the most documented walk-backs, which makes it the bellwether function.

The canonical case is Klarna. In February 2024 the company announced that its OpenAI-powered assistant handled two-thirds of customer service chats in its first month, doing the equivalent work of 700 full-time agents [6] — a vendor claim that became the industry's headline proof point. By 2025 the story inverted: CEO Sebastian Siemiatkowski said cost "seems to have been a too predominant evaluation factor," producing lower quality, and that "investing in the quality of the human support is the way of the future for us" [7], with Klarna resuming human hiring. A Forrester analyst diagnosed the failure as underestimating the complexity of customer service operations, noting that AI struggles with "novel scenarios, nuanced exceptions, high emotionality — all areas where humans excel" [8].

Production scale nonetheless exists. Sierra reached roughly $100M ARR in under two years using an outcome-based pricing model that charges for completed work rather than seats [11] — a credible signal because customers pay per resolution. Salesforce's Agentforce reports (vendor-sourced) nearly $1.4B in combined Agentforce and Data 360 ARR with 9,500+ paid deals [12], though independent Salesforce practitioners describe early versions as "very buggy ... it would lose context," improving only through late 2025 [23]. Even at Salesforce, support runs roughly 50% with agents and 50% with humans [13] — full replacement has not occurred at the vendor most incentivized to demonstrate it.

The metrics deserve scrutiny. Intercom reports its Fin agent resolves an average of 67% of customer inquiries (vendor-reported) [24], but independent analysis citing Gartner finds AI deflects more than 45% of customer queries while only around 14% reach full self-service resolution [25] — headline deflection numbers can overstate genuine resolution by a wide margin, and when deflection is optimized as a KPI, bots can route issues away from resolution rather than solving them. Two failures bound the autonomy risk. A Canadian tribunal held Air Canada liable for its chatbot's hallucinated bereavement-fare policy, rejecting the argument that the bot was a separate legal entity and ordering $812.02 in damages [26][27]; and in 2025 Anysphere's support bot for Cursor invented a non-existent login policy, reportedly triggering cancellations [28]. Support is where agents are most deployable, but "resolution" has to be measured at the customer-outcome level, and the complex and emotional tail stays human.

Sales and marketing agents

Sales and marketing produced both the most aggressive autonomy claims and the most spectacular failure, while the durable deployments are concentrated inside incumbent platforms.

The defining cautionary tale is 11x.ai. A March 2025 TechCrunch investigation found the a16z- and Benchmark-backed AI-SDR vendor listed customers it did not have: ZoomInfo said "we are not a customer ... 11x's product performed significantly worse than our SDR employees," and Airtable said its trial was never used in production [29]. Reported ARR was inflated by counting three-month trials as annual revenue — about $14 million claimed against roughly $3 million that survived the trial period — and the company was losing 70–80% of customers that came through the door [29]. Sifted's reporting added that employees found the product "barely functional" [30]. This is direct evidence that vendor traction claims in the category cannot be taken at face value.

The category-level problem is deliverability and quality. Independent analysis of 100,000 paired cold emails found AI senders flagged as spam at roughly 8% versus 3% for humans, with lower reply rates (4.1% versus 5.2%) and fewer meetings booked (0.7% versus 1.1%) [31]. Even vendors marketed as fully autonomous, like Artisan's Ava, required 5–10 hours of weekly human oversight in independent testing [32]. Gartner's forward call is that AI agents will outnumber sellers tenfold by 2028, yet fewer than 40% of sellers will report that agents improved their productivity, citing a value ceiling beyond which more AI does not mean more productivity [33].

Labor displacement is occurring at the role level even where tool efficacy is contested: Emergence Capital's survey of 560+ B2B firms found 36% cut SDR/BDR headcount in 2025, the steepest reduction of any sales role [34]. The durable deployment, meanwhile, is platform-native. Meta shipped an Ads CLI in 2026 explicitly so that AI agents can list campaigns, check spend against a budget target, and pause underperforming ad sets without a human UI [35] — autonomous ad operations inside a walled garden, a more defensible surface than standalone outbound SDR bots.

Finance and HR autonomy under controls and discrimination law

Finance and HR show strong vendor activity and executive intent but the least genuine autonomy, because regulatory and legal liability structurally penalize unsupervised agent decisions.

In finance the demand signal is loud: Deloitte's Q4 2025 CFO Signals survey found integrating finance AI agents a top transformation priority [36]. But Deloitte's own polling found trust is the leading barrier to tool use in finance and accounting [37], and the production reality is autonomy-with-a-gate. Ramp shipped controller agents that autonomously approve low-risk expenses or provide a recommendation with a rationale to the approver [38]; BlackLine's Verity markets end-to-end accruals that close three days faster but with human final review (vendor-reported) [39]. The structural brake is auditability: COSO's February 2026 guidance requires a complete, non-editable audit trail with preserved records of inputs, prompts, logic, outputs, model version, and human review [40] — a requirement fundamentally at odds with lights-out autonomous posting under SOX.

In HR, deployment is broader at the high-volume frontline but legally exposed. LinkedIn debuted a Hiring Assistant agent that sources candidates and asks screening questions [41], and Workday acquired Paradox, maker of the Olivia recruiting agent, for about $1B, completed October 2025 [42] — concentrating production autonomy in screening and scheduling. The legal risk is acute: a federal court conditionally certified Mobley v. Workday as a nationwide ADEA collective action on May 16, 2025 [43], and the ACLU filed an ADA and Title VII complaint over HireVue/Intuit AI video interviews, which the vendor calls "entirely without merit" [44]. Much of the major-vendor agent roadmap remains pre-production; Workday's Illuminate agents are largely slated for general availability in 2026 [45]. In both functions, expect agents to remain assistive and recommendation-mode with mandatory human sign-off for the foreseeable term — the binding constraints are SOX/COSO controls and anti-discrimination law, not capability.

Operations and the RPA-versus-agent contrast

Operations is where the "agent washing" phenomenon is most visible and where the contrast between marketing and production reality is sharpest.

The RPA incumbents are repositioning: UiPath's FY2025 annual report frames its strategy as "the path to agentic automation" [46]. But a 2025 academic comparison found LLM computer-use agents, while faster to build, are not yet production-ready, and that classic RPA outperforms agentic computer-use in execution speed and reliability, particularly in repetitive, stable environments [47] — for the high-volume, stable back-office processes that dominate operations, deterministic automation still beats agents on the metrics that matter. Practitioners describe a persistent "last-mile problem" where agents stall in legacy data silos and enter "prompt doom loops" on complex processes [48].

Where agents do land in production, they are supervised and require substantial preparation. ServiceNow's Rolls-Royce deployment reported 54% help-desk deflection, but only after the customer had to almost rewrite its knowledge articles to make them AI-ready, and a Forrester analyst observed that people want deterministic and auditable workflows, which are hard enough to govern on their own [49] — an implicit ceiling on autonomous operations. Adoption breadth is also thinner than the narrative suggests: Microsoft Copilot reached about 15M paid seats, roughly 3.3% of its 450M-subscriber base, after two years, with Forrester estimating most enterprises are 12–18 months from scaled deployment [50]. For stable, repetitive back-office work, deterministic RPA may remain the correct tool; agents fit variable, judgment-light tasks, and data and knowledge-base readiness is a heavy line item.

The reliability and security ceiling

The recurring rollbacks are predicted by measured agent reliability, which is the mechanism connecting every function's failures.

On Carnegie Mellon's TheAgentCompany benchmark of consequential real-world office tasks, the most competitive agent at the benchmark's late-2024 release completed 30% of tasks autonomously, with long-horizon tasks beyond the reach of the systems tested [51]. The problem is sharpest under repetition. The tau-bench pass^k metric showed 2024-era agents that often succeed once dropping to only about 25% success when repeating the same task eight times (vendor-published, methodologically transparent) [53]. The mechanism is compounding error — a 20-step process with 95% per-step reliability succeeds only 36% of the time [54] — so any workflow long enough to be worth automating is also long enough to fail unpredictably without a human checkpoint.

Security compounds the reliability problem. OWASP ranks prompt injection the #1 LLM application risk, and the EchoLeak vulnerability (CVE-2025-32711, CVSS 9.3) demonstrated a zero-click prompt-injection exploit that could exfiltrate enterprise data through Microsoft 365 Copilot via a single crafted email [55]. An agent that takes actions is an agent that can be hijacked into taking adversarial actions — a category of risk that static chatbots did not carry. Deployments that decompose work into short, checkpointed, verifiable steps and assume an adversarial threat model for any agent with action authority are responding to this ceiling.

Measured productivity gains

Well-designed studies find genuine gains; the nuance is where they land, which shapes both deployment strategy and headcount expectations.

The strongest single study is Brynjolfsson, Li, and Raymond's field RCT of 5,179 customer-support agents: AI access raised productivity, measured as issues resolved per hour, by 14% on average, including a 34% improvement for novice and low-skilled workers but minimal impact on experienced and highly skilled workers [14]. The Harvard/BCG experiment with 758 consultants found large gains — speed +25%, quality +40%, completion +12% — but only for tasks inside the AI "frontier," with degraded performance on tasks outside it, the "jagged frontier" [56]. Field experiments at an online retailer found generative-AI sales effects ranging from 0% to 16.3% across workflows, depending on the AI's marginal contribution relative to existing firm practice [57] — gains shrink where the firm was already good.

The consistent pattern is that AI compresses the gap between novice and expert and helps most on routine, well-scoped tasks, which explains both why support sees the clearest gains and why the entry-level workforce is the most exposed. It also reframes the "95% no P&L" result: gains are real at the task level but fail to aggregate to the P&L because of integration, learning, and verification overhead.

Headcount effects and AI-washing

The headcount question is where vendor and executive incentive bias is strongest, and where independent evidence most sharply contradicts the public narrative.

The claims are prominent. Klarna's CEO said AI helped shrink the workforce about 40%, via hiring freeze and attrition rather than direct swaps [15]; Salesforce's Benioff said he cut support from "9,000 heads to about 5,000, because I need less heads" [13]; IBM's CEO said AskHR automation replaced "a couple hundred" HR workers while total IBM headcount grew [58]; Shopify required teams to demonstrate why they could not get work done using AI before hiring [59]. The strongest pro-displacement signal is Stanford's "Canaries in the Coal Mine," which found early-career workers aged 22–25 in AI-exposed occupations saw a roughly 13–16% relative employment decline since 2022, with employment affected more than wages [60].

Independent labor economists are skeptical that AI is the cause of broad labor weakness. The Yale Budget Lab found the labor market reflects largely stability, not major disruption, with the occupational mix changing no faster than in prior technological shifts [16]. A NY Fed analysis of job postings found AI is not the main driver of the slowdown in hiring [17], and a separate NY Fed study attributed much of rising young-graduate unemployment to remote work impeding on-the-job training rather than AI [61]. The AI-washing critique is explicit: Yale's Martha Gimbel argued CEOs prefer blaming AI over admitting mismanagement or over-hiring [18]; Block's Jack Dorsey conceded "yes we over-hired during COVID" after publicly linking layoffs to AI [19]; Amazon's CEO said its roughly 14,000 corporate cuts were "not even really AI-driven ... It's culture" [20]; and UPS's roughly 48,000 cuts tracked declining package volume [62]. Challenger's tally — 54,836 U.S. layoffs attributed to AI in 2025 out of roughly 1.2M total cuts — frames AI as a minor named cause amid an over-hiring correction [63]. The labor signal is real at the entry level but confounded by macro correction, remote-work effects, and incentive-driven attribution.

Patterns recurring across functions

Three patterns hold across the six functions. The first is the deploy–reassert cycle: an ambitious autonomous deployment, a quality or reliability failure on the complex tail, and a walk-back to human-in-the-loop supervision. Klarna, the AI-SDR category, and Duolingo's reversal of its "AI-first" memo [64] all follow this shape, and Gartner has effectively forecast it as a market-wide event with its 50%-abandonment [9] and 40%-cancellation [10] predictions.

The second is that autonomy is inversely proportional to consequence and regulation. The functions with the most genuine autonomy — support deflection, recruiting screening, ad operations — are those where errors are cheap and reversible; the functions with the least — finance posting, hiring decisions — are those where errors are costly and legally actionable. COSO/SOX controls [40] and ADEA/ADA litigation [43][44] are the rate-limiters, not incidental friction. The EU AI Act's Article 14 human-oversight requirement [66] sits in the same role.

The third is that vendor metrics systematically overstate. Deflection is not resolution [25], trial revenue is not ARR [29], "fully autonomous" is not unsupervised [32], and "agentic" is not always agentic — Gartner counts only about 130 of thousands of vendors as offering real agentic AI [10]. Every function's headline metric required downward adjustment where independent data existed.

Two structural observations follow. The "95% failure" finding and the real RCT gains are the same story: task-level gains are real [14][56][57] but fail to aggregate to enterprise P&L [4][3] because of integration cost, the learning gap [21], and a verification tax — the human effort to check agent output, which the reliability ceiling makes unavoidable. And the most defensible deployments are platform-native: Meta's Ads CLI [35], LinkedIn's Hiring Assistant [41], and in-platform ServiceNow and Salesforce agents [12][49] outperform standalone third-party agents because they own the data, the action surface, and the guardrails — the exact things the last-mile problem [48] breaks on. Outcome-based pricing functions as an honesty filter for the same reason: Sierra's ~$100M ARR on pay-per-resolution [11] is more credible than any seat-based figure because the vendor earns only when the agent completes work. Governance and data-readiness surveys point the same way, with controls and stack readiness now treated as prerequisites for production agents [65][67].