Where is the money in the agentic AI market?

Where would a company pay for an AI agent instead of a junior, offshore worker, or BPO?

Companies pay for AI agents when work is high-volume, repeatable, software-mediated, and measurable. Basically, when it’s replacing a queue, not only a job.

The best tasks arrive as tickets, chats, leads, invoices, forms, emails, or CRM objects. The action is bounded by rules, policies, or past examples. The agent can act directly: refund, route, update a record, qualify a lead, draft a reply, or open a case. Success is measurable: resolved ticket, qualified lead, booked meeting, accepted document, completed workflow, or lower handling time.

That is why agents land first in support, SDR, data enrichment, back-office ops, content production, and QA-like review. They do not need career growth or months of context. They need a narrow production lane with enough volume to justify setup and monitoring.

When are people willing to pay a lot for an AI agent?

Buyers will pay a lot when an agent touches revenue, removes a visible cost center, or clears a bottleneck.

In short, willingness to pay is highest when the buyer can point to an existing budget line.

The strongest zones are customer support deflection, sales pipeline creation, and operational backlogs.

Buyers pay less when the agent is just “productivity software.” A writing assistant, research helper, or generic analyst may save time, but saved time is hard to price.

If you want more recent data on this point, please see our latest agentic AI market report.

When is an AI agent too expensive compared with a human?

An AI agent is too expensive when volume is low, failure is costly, or the workflow needs too much supervision.

The evidence points in that direction. A 2025 study by Cornell University, named “How Do AI Agents Do Human Work?”, comparing human and agent workflows found agents could be much faster and cheaper on programmable tasks, but also found weaker quality and risks such as fabrication and tool misuse. That means raw “agent cost per task” can look great while the true cost moves into review, correction, and governance.

Software is the warning case. Early Copilot experiments showed developers completing a constrained coding task 55.8% faster with Copilot, but later studies on mature repositories were much less clean: one 2025 randomized trial found experienced open-source developers were 19% slower with AI tools, and another study found AI-assisted code increased rework burden on core maintainers.

Do we have real-economy examples where an agent replaced an identifiable cost unit?

Yes, but today the strongest examples are still concentrated in customer support and outsourced digital labor.

We do not yet have broad proof that AI agents are replacing full occupations across the economy, but we do have clear examples where they replaced identifiable units: tickets, conversations, support agents, contractors, and online freelance spend.

Klarna is the cleanest public example. In February 2024, Klarna said its OpenAI-powered assistant handled 2.3 million conversations in its first month, covered two-thirds of customer service chats, did the equivalent work of 700 full-time agents, reduced repeat inquiries by 25%, cut resolution time from 11 minutes to under 2 minutes, operated in 23 markets and 35+ languages, and was expected to improve profit by $40 million in 2024. That is not a vague productivity claim but a direct mapping from agent output to a support-labor cost unit.

Salesforce is another direct example, though the source is executive disclosure rather than a formal audited cost study. Marc Benioff said Salesforce reduced its support staff from about 9,000 to 5,000 after deploying AI agents, with AI handling roughly half of customer interactions. That is one of the clearest “agent replaced support headcount” claims from a major software company.

Duolingo is a different version: less “agent replaces employee,” more “AI production system replaces contractor output.” The company cut about 10% of its contractor workforce in late 2023 as it used AI to streamline content production, and later told investors that AI tools changed course creation speed: 20,500 course units published in Q1 2026, versus 7,100 per quarter in 2025 and 1,800 per quarter in 2024. That is a real cost-unit shift from human content production toward AI-assisted production, even if it is not a pure autonomous-agent replacement.

The best aggregate evidence comes from firm spending data. A 2026 paper using U.S. expense-management data tracked company spending on online labor marketplaces and AI model providers from Q3 2021 to Q3 2025. Firms most exposed to online labor adopted AI earlier and reduced contracted labor spend; the paper estimates that a $1 decline in online labor spending was associated with about $0.03 of additional AI spending among the most exposed firms. That is exactly the pattern we would expect if companies are replacing outsourced digital tasks with cheap AI services, not buying AI as another seat on top.

So, yes, replacement is real, but it is not yet evenly distributed.

Are big “AI layoff” companies actually replacing people with agents?

Mostly no, except in a few narrow cases where support or contractor work is explicitly automated. Many “AI layoffs” are better read as cost-cutting plus AI reallocation, not proven one-for-one agent replacement.

Company	What happened	What it proves	What it does not prove
IBM	In 2023, IBM said it would pause or slow hiring in back-office roles, including HR. It also said roughly 30% of non-customer-facing roles could be replaced by AI and automation over five years.	This is a strong automation signal. IBM clearly expected AI to reduce demand for some back-office labor.	It does not prove that deployed agents had already replaced those workers at the time.
Duolingo	Duolingo cut 10% of contractors and connected the move to AI-enabled content production. Later investor materials showed a step-change in content output per quarter.	This supports a real automation story in contractor content workflows.	It does not prove broad full-time employee replacement across the company.
Salesforce	Marc Benioff described a reduction of roughly 4,000 support roles and tied it directly to Agentforce handling a large share of support interactions.	This is the clearest “yes”: a real agent-replacement claim in a specific function.	It is still specific to support, not proof that AI agents are replacing jobs broadly across all functions.

For broader layoff lists, the evidence is weaker. Business Insider’s mid-2026 roundup says many companies cite AI in job cuts, but also notes the risk of “AI washing” and that some companies simultaneously redeploy or hire in AI-related areas.

So, we don’t really have proof of deployed agents replacing those jobs one-for-one.

If you want more recent data on this point, please see our latest agentic AI market report.

Are any industries being revolutionized by AI agents with reliable proof?

As of today, customer support is the only industry where the agent revolution is already commercially visible and measurable. Coding and sales development are changing fast, but support has the strongest proof because the unit of work is clean: conversations, resolutions, escalations, CSAT, handling time.

Klarna’s numbers show why support moved first: millions of conversations, two-thirds of chats automated, 700 full-time-agent equivalent, lower repeat inquiries, faster resolution, and a $40 million profit-improvement estimate. Those are the exact metrics a support organization already uses to manage BPOs and internal teams.

Intercom, Zendesk, Sierra, Decagon, and Salesforce are all converging on the same category: autonomous customer-service agents. Intercom prices Fin around successful outcomes; Zendesk says its outcome-based pricing focuses on automated resolutions; Sierra publicly argues for pricing only when an agent achieves a valuable customer outcome; Salesforce prices Agentforce through conversations, actions, or Flex Credits. This is not one vendor narrative. It is a market structure forming around support work as measurable digital labor.

Software development is being transformed, but not yet cleanly “revolutionized” in the economic sense. A 2026 study of 129,134 GitHub projects found coding-agent adoption between 15.85% and 22.60% in the first half of 2025, very high for a category only a few months old. But productivity evidence is mixed: some tasks accelerate, while mature-codebase work can slow down due to review, waiting, and correction.

Sales development is the next serious candidate. Vendors such as Qualified and 11x sell AI SDRs that prospect, engage, qualify, and route buyers. The proof is more commercial than academic: the category exists because the SDR function already has priced units, such as qualified meetings and pipeline. But compared with support, public third-party performance data is thinner.

Which agentic AI companies say they are selling strongly into specific agent categories?

Today, the agentic AI market is clustering around a few sellable agent jobs: support agent, sales/SDR agent, coding agent, and workflow/back-office agent. The strongest signal is when vendors build the company around one labor category.

Intercom is explicitly selling Fin as an AI customer-service agent. Its pricing and help materials define chargeable “outcomes” around Fin successfully delivering value, and its sales expansion now includes Fin for Sales priced at $10 per qualified lead, where the customer defines what “qualified” means. That shows Intercom is moving from generic AI support into distinct agent jobs: support resolution and lead qualification.

Sierra is also focused on enterprise customer-service agents. Its public pricing thesis says AI agents executing processes autonomously make outcome-based pricing possible, where customers pay only when software achieves specific valuable outcomes. That tells us Sierra believes the money is in resolved customer workflows, not generic seats.

Qualified is focused on the AI SDR. It describes Piper as an AI SDR agent and PiperX as a full-funnel AI SDR “superagent” that can engage, nurture, and convert buyers through multimodal interactions. That is a very specific labor category: sales development, not broad automation.

11x sells “digital workers” for sales, RevOps, and GTM teams, with Alice and Julian positioned around revenue workflows. Its site frames the product as AI workers that execute, learn, and optimize across customer touchpoints. The interesting signal is the buyer: GTM leaders, not CIOs buying generic automation.

Decagon is another customer-experience agent company. Its case-study page highlights Hertz using proactive outbound agents to resolve issues before they arise and free human agents for higher-value interactions. That is important because it moves support agents from reactive ticket deflection to proactive issue resolution.

It looks like agentic AI is not becoming one horizontal software market but fragmenting into labor markets.

Are agentic AI companies moving from seat pricing to result pricing?

Yes, and the clearest migration is in customer service, where vendors are moving from seats or conversations toward resolutions and actions. This is one of the strongest commercial signals in the market.

Intercom is the canonical example. Fin’s model charges per resolution, and Stripe’s customer story says Intercom adopted outcome-based pricing so users pay 99¢ per resolution, charged when the customer confirms the AI answer resolved the issue or does not ask for more help after the final AI answer. That is not seat pricing. It is a priced unit of completed labor.

Zendesk publicly says it implemented outcome-based pricing for AI agents, focused on automated resolutions that deliver value to customers. IDC’s summary of Zendesk’s move describes it as charging customers only when AI agents fully resolve inquiries autonomously.

Salesforce is not pure outcome pricing, but it is moving away from simple seats. Agentforce launched at $2 per conversation, and Salesforce later introduced Flex Credits so customers pay for actions the agent performs, such as updating records, automating workflows, or resolving cases, at $0.10 per action. That is consumption/action pricing, not “buy one seat per employee.”

Sierra is the clearest philosophical version. It says autonomous agents enable a model where customers pay only when software achieves specific, valuable outcomes. Public details on exact rates are not disclosed, but the model direction is explicit.

If you want more recent data on this point, please see our latest agentic AI market report.

Which vendors accept being paid only when the agent succeeds?

Intercom, Zendesk, and Sierra are the clearest public examples of vendors accepting success-based AI-agent economics. Salesforce is adjacent, but its model is action-based rather than strictly success-only.

The clearest pricing shift is from seats or usage toward measurable agent work. But not every model is equally outcome-based.

Company	Pricing model	What it signals
Intercom	Fin is priced at 99¢ per resolution, charged only when the AI answer resolves the issue or the customer stops asking for help.	The money is in resolved support tickets, where vendors can prove the agent completed useful work.
Zendesk	Zendesk prices around autonomous resolutions, charging when AI agents fully resolve inquiries without human help.	The money is in autonomous support resolution, especially work that would otherwise require human agents.
Sierra	Sierra advocates outcome-based pricing, where customers pay only when AI achieves valuable outcomes. Pricing is custom and not public.	The money is in high-value enterprise workflows, not generic seats or raw usage.
Salesforce	Flex Credits charge for agent actions, such as updating records, automating workflows, or resolving cases.	The money is in agent-executed business tasks inside enterprise systems, even when they are not final outcomes.

Who is actually paying more for AI agents?

Insurance might be the strongest “more budget” signal.

So, we made some research and, unfortunately, no named enterprise publicly discloses agent spend in a way that lets us rank industries by actual AI-agent budget.

But we do have some signals.

The best hard signal is insurance. KPMG’s 2025 Insurance CEO Outlook says 73% of insurance CEOs prioritize AI investment for underwriting, claims, and customer experience, and a related industry summary says 67% plan to allocate 10–20% of budgets toward AI.

That is AI broadly, not only agents, so we should not overclaim. But they are still saying AI is a top budget priority, not just a productivity experiment.

Who has AI agents handling real customers?

DNB and Substack have some of the cleanest examples of AI agents handling real customers.

Indeed, DNB, Scandinavia’s largest bank by market value, worked with boost.ai to manage high chat volume and automated more than half of online chat interactions in six months. This is real customer interaction, not internal productivity and not lead qualification.

Substack is another clean customer-facing case. Decagon says Substack launched an AI agent that handles and resolves more than 90% of user inquiries. The proof level is vendor case study, not independent audit, but the claim is specific: real users, real support inquiries, real resolution rate.

It looks like the strongest category is still customer support, because the work is already digital, repetitive, and measurable.

Who has AI agents touching real money?

R1 is the cleanest example we found of an AI agent touching payment workflows.

Sierra says R1 deployed an AI agent for high-volume patient inquiries including balance checks, payment processing, payment-plan setup, and account questions.

That means the agent is not just answering FAQs but also operating inside a revenue-cycle workflow where money is owed, paid, or structured into a plan.

We should be precise: this does not mean the agent is autonomously making financial risk decisions. The proof says it handles payment processing and payment-plan setup in patient revenue operations.

That is enough to say it touches real money, but not enough to say it controls financial policy.

Who has AI agents making real decisions?

Today, we do not have a clean public example of a named company letting generative AI agents make top-level business decisions autonomously.

We found examples of agents routing, resolving, processing, qualifying, and executing workflows. We did not find a strict example where a company says an AI agent independently makes major pricing, hiring, lending, underwriting, investment, or strategy decisions without human governance.

The broader research supports that caution. A 2026 study of agentic AI adoption across 12 industrial companies, named “Agentic AI in Industry: Adoption Level and Deployment Barriers”, found most organizations still at assistant or compensator levels, with only one at multi-agent orchestration.

The main blocker was verification: companies had higher-level experimental capabilities, but could not integrate them into production because human-in-the-loop verification remained the trusted control mechanism.

If you want more recent data on this point, please see our latest agentic AI market report.

Who showed a real ROI from agents?

As of today, we found two clean examples of companies showing quantified ROI or direct savings from AI agents: Salesforce and Engine.

Salesforce says Wiley uses Agentforce and Service Cloud with Einstein AI for customer service operations, and that AI-powered agents help customers resolve common issues.

The disclosed result: 50% faster onboarding for seasonal agents, 213% return on investment, and $230,000 in savings. This is not perfect proof of autonomous agents replacing a whole function. But it is a clear named-company ROI claim tied to AI-powered customer-service agents and quantified business impact.

A second strong example is Engine, the business-travel platform. In April 2026, they deployed an Agentforce AI agent called Eva which now manages more than 50% of customer cases end-to-end, including rescheduling reservations and recommending accommodations.

The disclosed business result is that this cut handle times and saves “millions annually.” We do not get the exact dollar number, so Wiley is cleaner. But Engine is still a strong real-economy example because the source ties the agent to end-to-end customer-case handling and annual savings, not just internal productivity.

Who says agents made them more money?

Xanterra is the clearest revenue-growth example we found when it comes to the AI agentic market.

Cresta says Xanterra Travel Collection used Cresta AI Agent across National Parks guest operations and realized a $3.3 million revenue increase, while also avoiding hundreds of thousands in potential guest recovery costs.

A related Cresta post from October 2025 says the AI Agent strategy delivered 74% containment and unlocked $3.3 million in new revenue.

That’s interesting because the agent did not just cut costs, it also improved guest-service execution and reinforced sales behaviors enough to create incremental revenue.

A second example exists, but it is weaker because it is Salesforce talking about Salesforce.

Salesforce says Agentforce on its own website powered more than 30,000 new leads in 2025. It also said in February 2025 that visitors who engaged with Agentforce were 1.8x more likely to convert to a lead than the website average, creating an incremental $600,000 increase in marketing-driven pipeline per 1,000 site visitors from Agentforce.

This is not booked revenue, so we should not present it like Xanterra. But it is still a real monetization signal: the agent changed visitor conversion and created measurable pipeline.

Who really cut costs with AI agents, with numbers?

Yes, we have real cost-reduction examples, but the clean ones are still mostly in customer service.

We ranked them from strongest to weakest proof: named company, clear agent use, quantified cost/headcount/time impact, and direct link between the agent and the cost reduction.

Rank	Company	Agent type	Context	Cost-reduction numbers
1	Klarna	Customer-service AI assistant	Klarna’s OpenAI-powered assistant handled real customer-service chats across refunds, returns, payments, cancellations, disputes, and invoice issues.	2.3 million conversations in month one, two-thirds of customer-service chats, equivalent work of 700 full-time agents, 25% drop in repeat inquiries, resolution time down from 11 minutes to under 2 minutes, estimated $40 million profit improvement in 2024.
2	Salesforce	Internal customer-service AI agents / Agentforce	Marc Benioff said Salesforce reduced customer-support staffing while AI handled a large share of customer interactions.	Support staff reduced from about 9,000 to 5,000; AI handled around 1.5 million customer interactions, roughly half of engagements; Salesforce said AI customer service saves about $100 million annually.
3	Engine	Agentforce customer-support agent, “Eva”	Engine uses Eva to manage customer cases end-to-end, including reservation changes and accommodation recommendations.	Eva manages over 50% of customer cases end-to-end; Salesforce says it cuts handle times and saves millions annually. Diginomica gives the sharper number: 15% lower average handle time and an estimated $2 million annual cost saving.
4	Wiley	Agentforce + Service Cloud AI agents	Wiley uses AI-powered service agents during seasonal customer-service peaks.	50% faster onboarding for seasonal agents, 213% ROI, and $230,000 in savings. This is clean ROI proof, but the dollar impact is smaller than Klarna, Salesforce, or Engine.
5	Reddit	Agentforce advertiser-support agent	Reddit uses Agentforce to answer repetitive advertiser-support questions.	Case deflection increased 33%; average resolution time fell from 8.9 minutes to 1.4 minutes, an 84% improvement; Salesforce says this saves 760 live-rep hours per year.
6	1-800Accountant	Agentforce tax-support agent	1-800Accountant used Agentforce to absorb tax-season support demand.	Agentforce resolved over 1,000 client engagements in the first 24 hours; later Salesforce materials say it reached 90% case deflection during tax week.

Who says AI agents reduced headcount?

Salesforce is the clearest headcount example in the AI agentic market.

Marc Benioff said Salesforce cut its customer-support workforce from about 9,000 to 5,000 and replaced part of that work with AI agents.

The reported operating signal is also specific: AI handled around 1.5 million customer interactions and roughly half of engagements.

We would not call every AI layoff a true agent replacement. This one is stronger because the function is named, the before/after headcount is given, and the agent workload is quantified.

Who says AI agents reduced outsourcing?

Klarna is the cleanest outsourcing example in the AI agentic market.

The Guardian reported that Klarna’s internal AI program steadily reduced its use of outsourced workers, including in customer service, with technology carrying out the work of 853 full-time staff, up from 700 earlier.

It also said Klarna’s headcount fell from 5,527 to 2,907 since 2022, mostly through attrition, with departing staff replaced by technology rather than new hires.

Strictly, this is not “we cut BPO invoice by X dollars.” But it is the best public proof we found that agents reduced outsourced labor, not just internal tickets.

Who says AI agents reduced ticket volume?

Reddit is the cleanest named-company ticket-deflection example in the AI agentic market.

Salesforce says Reddit’s Agentforce deployment increased case deflection by 33%, reduced average resolution time from 8.9 minutes to 1.4 minutes, improved advertiser satisfaction by 20%, and saved 760 live-rep hours per year.

A second example exists, but it is weaker because the customer is unnamed: Braincuber says an AI support agent cut weekly ticket volume for a U.S. e-commerce brand from 1,900 to 763 after 90 days, a 60% reduction, while reducing needed agent headcount from 11 to 6 and cost per resolved ticket from $8.40 to $1.90.

So the strict named-company answer is Reddit. The sharper numeric ticket-volume reduction exists, but the company is not named.

Who says AI agents reduced backlog?

SilencerCo is the clearest named backlog example we found, but the proof comes through a vendor article.

CoSupport says SilencerCo managed around 3,500 tickets per month in Freshdesk and saw a 60% reduction in ticket backlog after using CoSupport AI. The agent was trained on ticket history and help-center content so routine inquiries could be resolved automatically.

We would use this, but with a proof caveat: it is a named customer and a precise backlog number, but it is not an independent audit or investor disclosure.

Who tried AI agents and lost money?

Air Canada is the cleanest company-level example where an AI customer-service bot directly cost money.

In the Moffatt v. Air Canada case, the airline’s chatbot gave a customer wrong information about bereavement fares. The British Columbia Civil Resolution Tribunal rejected Air Canada’s attempt to distance itself from the chatbot and ordered the airline to pay damages, interest, and fees. Reports put the amount around CAD $812 total, including CAD $650.88 in damages.

That is a small dollar loss, not a strategic failure. But it is strong because it is legal, named, factual, and directly tied to chatbot output.

Who killed an AI agent project?

DPD is the cleanest named example of a company disabling an AI customer-service feature after failure.

In January 2024, DPD disabled part of its AI chatbot after a customer made it swear, criticize the company, and generate inappropriate outputs. DPD said a system update caused the issue and deactivated the problematic AI component while it investigated.

Meta also killed a visible AI project, but it was more consumer-social than enterprise workflow. Meta removed AI-generated Instagram and Facebook profiles after renewed attention to old AI personas and technical issues, including users being unable to block some accounts. This is a real AI project shutdown, but not a classic enterprise agent doing back-office or support work.

At aggregate level, the rollback problem is much bigger than the named examples. Sinch’s 2026 survey says 74% of enterprises had rolled back or shut down an AI customer-communications agent after deployment because of governance failures. The main reasons included data exposure, hallucinations, and lack of auditability.

Who brought humans back after using agents?

Klarna is the clearest example in the AI agentic market.

After loudly saying its AI assistant did the work of 700 full-time customer-service agents, Klarna later resumed recruiting human customer-service workers.

CX Dive reported in May 2025 that Klarna was turning back to people for more customer-service work, while still saying the AI chatbot handled about two-thirds of inquiries.

This is the most important negative case because it does not say “agents failed completely.” It says something more useful: agents worked for volume, but humans had to come back for quality, trust, and complex support.

So, clearly, agents can cut cost very visibly, but the companies that push too far toward “AI-only” support tend to rediscover that customer service is not just ticket resolution. It is also exception handling, confidence, and brand repair.

If you want more recent data on this point, please see our latest agentic AI market report.

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.