From the Editor’s Desk:
“I enjoyed co-authoring this piece with Vijay Rayapati, co-founder and CEO of Atomicwork, and appreciated his refusal to sugarcoat the part most AI vendors tiptoe around: every business is going to get leaner, and AI will accelerate that shift. But the underlying argument is what has stayed with me. The fight over enterprise AI is not about whose model is smarter. It’s about which budget line a vendor is paid from. The moment a vendor asks to be paid like a contractor rather than a toolmaker, someone has to decide what counts as a completed job and who gets to verify it.
What follows is his perspective. I hope you enjoy reading this!” – Victor Dey
For decades, enterprise software companies have followed a consistent business model. Companies hire employees, employees need software applications to do their jobs, and vendors charge for software access. As organizations grow, they purchase more licenses, creating reliable revenue for some of the largest technology companies.
AI coworkers could fundamentally change that relationship by performing work that previously required employees to operate software. The buyer’s question changes from “how many users need this?” to “how much work can this system complete, and at what cost?”
That puts pressure on an industry whose revenue has grown with headcount. An agent automates a workflow. A coworker is accountable for a job. A single IT role encompasses hundreds, sometimes thousands, of workflows, and this piece is about the difference between automating them individually and taking responsibility for the entire role.
Choosing Between the Software and the Salary Budget
Established vendors such as SAP, Salesforce and ServiceNow are adding AI to applications that manage critical business functions. A second group is building AI-native alternatives to traditional ERP, CRM and IT service management software. The third group is approaching the market through the work itself. Companies such as Sierra, Decagon and Atomicwork are building AI coworkers for customer service and IT operations. These three approaches might look like they have some overlap, but only the third group lets customers purchase completed work instead of employee tools.
Salesforce now bills Agentforce per action and per resolved conversation. ServiceNow said on its first-quarter 2026 earnings call that half its net new business comes from non-seat pricing. So the meter alone does not define the distinction. The distinction lies in which budget the vendor is trying to capture. Incumbents are adding consumption pricing to grow the software budget. The third group is targeting the salary budget and the outsourcing budget alongside it.
Say I need a logo. In one model, I give you Adobe, two weeks and $2,000. In the other, I tell you that I don’t care whether you use Adobe, ChatGPT or Claude, or whether it takes two days or two weeks; I need three logo options, and I’ll pay $1,400 for them. In the first model, you are buying software and time. In the second, you are buying the outcome. A CEO, CFO or CIO ultimately wants the second option. They want the work completed.
There is a related distinction within the AI-native camp between companies trying to replace incumbent software and those trying to replace the work performed around it.
Take Serval. ServiceNow started with tickets and added workflows. Serval is essentially saying it can automate those workflows, while still approaching the opportunity through the software budget. Its pitch to a customer goes something like this: you pay ServiceNow $500,000; pay me $250,000, and I’ll replace the software.
Our pitch is different. You pay ServiceNow $500,000, but you also pay Accenture $5 million. Give me $1 million of that work, and I’ll show you what I can do. That is why our annual contract values are much higher than those of a software substitute. A competitor offering similar capabilities at 30% to 60% of ServiceNow’s price still sees the problem in terms of software. There is nothing wrong with that. It is simply a smaller pool of money.
Paying for Verified Outcomes Instead of Seats
Traditional software pricing tracks headcount as the business model revolves around seat-based pricing. AI coworkers weaken that connection by increasing output without an equivalent seat count. The cost of running AI has also fallen dramatically in the last decade. By my count, token costs have fallen by a factor of nearly a million over the past six years alone.
But cheaper AI does not mean a smaller bill. Jevons paradox is at work here. As the cost of AI work falls, enterprises will do far more of it. Most companies today investigate about 1% of their security events. The world is moving toward investigating 100%. It is a hundredfold increase, and no CFO is going to multiply the security budget by a hundred to pay for it. So the money has to come from somewhere. It moves out of the software and salary lines and into agentic IT and security budgets. Total spending doesn’t necessarily fall. Its composition changes.
What falls is the cost per unit of work, which is the number we track at Atomicwork: cost per verified outcome. Our internal data puts the cost of level-1 support roughly 70% to 80% below the cost of having a person do the same work. Level 2 runs 40% to 50% cheaper. Level 3, where the work becomes more complicated, still comes in 20% to 30% below human delivery. Those are our figures, not an auditor’s, and the advantage narrows as the work gets harder.
Established vendors like ServiceNow face a challenge. Many add AI to products and pricing models built around human users. Customers will increasingly question why they are paying for additional software seats, AI features and the employee time still required to complete work.
We compete with ServiceNow, so weigh what follows accordingly. ServiceNow was built for an era of tickets and tasks. To its credit, it moved from tickets to workflows. The next move is from workflows to work, and that requires a different order of change. Consider what a customer still has to assemble to get value: buy the platform, bring in KPMG or Accenture to integrate it, hire a couple of ServiceNow developers to configure the workflows, then add the AI module and the L1 Service Desk AI Specialist. The specialist handles a portion of the work. It is not responsible for the job. The customer pays for the software, configuration, integration and labor, along with what I’d call the coordination tax: the cost of someone inside the enterprise making it all work together.
An enterprise can start with our AI coworkers for about $24,000 a year. I don’t believe an enterprise can realistically start with ServiceNow at that price.
Can incumbents execute? Of course. Assume ServiceNow executes at 70% on the architectural change, 70% on the business model change and 70% on go-to-market. Multiply those figures and you get roughly 30%. Each change is manageable on its own. Making all three at once while protecting existing revenue is the innovator’s dilemma in its purest form. Siebel launched its own answer to Salesforce. It understood the market better than Salesforce did. That did not save it.
I don’t think customers will rip ServiceNow out. The more likely outcome is quieter. Over the next three to five years, an enormous share of incumbent revenue will come up for renewal, and the conversation will not be about the platform. It will be about the Accenture work on top of the platform or the Infosys work alongside it. Perhaps the Deloitte audit work, too. Those layers come off first. The platform may stay. The work around it changes hands.
Enterprises want predictable costs and measurable results when they hire an AI workforce, so we offer both: credits for variable AI workloads and fixed prices for defined outcomes, at $1 for a resolved knowledge query, $2 for an access request and $3 and up for a service or incident resolution. The hard part is not the price. It is the word “resolved.”
Closing a ticket is not the same as resolving a problem. An incorrect answer, a failed fix or a handoff that a human technician must finish should not count as an autonomous resolution. Customers deserve a clear distinction between paying for AI usage and paying for completed work, and most of the industry has yet to draw it. Is this model 99.99% ready for enterprise use? No. If you hire someone to produce three logos, you still require them to sign an NDA, and you might still run a background check. Confidentiality, security and compliance will slow adoption relative to consumer technology. The economics have already shifted. The paperwork is catching up.
Treating IT as Ground Zero
IT departments handle large volumes of recurring requests, from application access and password resets to device support, system alerts and technical problems. At Atomicwork, we are building an AI-native IT service management platform with AI coworkers that automate service operations. We ask customers to pay for the work those coworkers handle, just as they would pay Infosys, TCS or Cognizant for an outsourced service desk.
One financial services customer that moved from ServiceNow to Atomicwork cut its average resolution time from 29 hours to 7.25 hours and now fully automates about 45% of its requests. Back-office onboarding that once took five business days now takes two hours when no physical hardware is involved. Across customers migrating from ServiceNow, we see total cost of ownership fall by 50% to 75%, which is why we guarantee a 50% reduction.
Reliability remains essential. Granting access to a collaboration tool carries different risks from providing administrative privileges to a production database. Organizations must define which actions AI coworkers can perform independently, which require approval and how employees can investigate mistakes. IT has always governed identity and access for people. It must now do the same for AI coworkers, effectively becoming the HR department for an AI workforce.
What Happens to Existing Enterprise Applications?
As AI coworkers assume more responsibility, employees may interact less frequently with traditional software applications. Someone requesting software access may never open an IT service portal if an AI coworker can complete the process through Slack or Microsoft Teams.
The obvious objection is that whoever owns the system of record owns the value. I don’t think history supports that. Technology has always been a layered cake. Intel could have argued that it controlled the chip, so software didn’t matter. Borland controlled the compiler. Someone else controlled the operating system. You judge a cake by the layer you bite into. Enterprises will judge their technology by the layer they interact with to get work done, and I believe AI coworkers will become that layer. Owning the underlying record still matters. It no longer guarantees ownership of the relationship on top.
SAP, Microsoft, Salesforce and ServiceNow will all build this. SAP was late to the cloud and got there. Microsoft was late to the cloud and built Azure. Both took 10 to 15 years. Earlier transitions gave incumbents that much time. This one won’t. The architecture and business model are changing simultaneously. If you don’t transform faster, you die faster. Every business is going to become leaner in headcount. We should not sugarcoat it.
Twenty-five or thirty years ago, a $1 billion business in the United States might have needed 10,000 people. Today, many run on 4,000 to 5,000. Technology drove that change long before AI arrived. AI will accelerate it. What I’d push back on is the leap from leaner businesses to job destruction. Businesses die because they are unprofitable, and the surest way to stay profitable is to lower operating costs. A profitable company keeps employing people. An unprofitable one employs nobody.
For forty years, enterprise software sold seats to the people doing the work. The next generation of vendors will be judged by the work itself.
The Control Layer Verdict
Vijay asserts that a contract that pays only when the job is done threatens incumbents more than any model does. Consumption meters from Salesforce and ServiceNow narrow the pricing gap but leave the budget distinction intact: they still grow software spending, while AI-native vendors price against salaries and outsourcing. That shift moves the hard question from how many seats are needed to who verifies the work, and the vendor that defines a resolved request most honestly will set the standard against which everyone else is measured. Enterprises are starting to hire software the way they hire people. Now, they might have to govern it the same way: by what it delivers, under rules rigorous enough to trust.
