Over the past few years, enterprises experimenting with AI agents grew accustomed to a comforting illusion that intelligence was cheap. That illusion is now breaking.
Ed Keisling, Chief AI Officer (CAIO) at Nasdaq-listed Progress Software, says the withdrawal of subsidised AI pricing by frontier labs is forcing companies to confront what it actually costs to keep AI agents running around the clock, and the numbers are changing behaviour fast.
“We used to have some agents at Progress that were running 24/7, and they were running thousands of dollars a month,” Keisling tells AIM in an interview, describing internal engineering agents his teams had deployed to research issues and assist with coding. “Once we started to understand what that cost relative to the value we were getting out of it, the economics really no longer made any sense.”
Keisling’s point cuts against the industry’s dominant narrative of AI as a near-free productivity multiplier.
He says that the business model of frontier labs such as Anthropic and OpenAI has, until recently, nudged enterprises towards their most powerful—and most expensive—models by default, even for routine, repetitive knowledge work that does not require that level of capability.
“These models are extraordinarily powerful. They are trained on biology, physics and every language known to man, and coding. But the reality is, for a lot of knowledge work on a day-to-day basis, people do not need all of those capabilities,” he says. “You are paying for that power, and you are incurring that cost just because it can do those things, even though it’s not doing them on a day-to-day basis.”
With model providers now pulling back on subsidised plans, a shift Keisling says has visibly accelerated over the past few months, he expects enterprises to move decisively towards smaller, cheaper, and often locally run language models, alongside a renewed appreciation for already-proven software systems that can accomplish the same tasks more economically.
That, in turn, is reshaping the conversation around so-called ‘AI deflation’ in enterprise software. Rather than frontier models displacing existing applications wholesale, Keisling expects a more selective, cost-driven sorting of workloads between AI and traditional software systems.
Founded in 1981 and listed on Nasdaq, Progress Software builds infrastructure and application-development software that underpins business-critical systems for enterprises globally.
Over the last seven years, the company has pursued an inorganic growth strategy, acquiring IT management firm Ipswitch, multi-model NoSQL database and semantic metadata management platform MarkLogic, file management platform ShareFile and agentic retrieval-augmented generation (RAG) and AI search solutions platform Nuclia, alongside its existing Chef infrastructure-automation business, a run of deals that Keisling says has taken Progress from roughly $300 million to nearly $1 billion in revenue.
Today, the company says more than four million developers and technologists across hundreds of thousands of enterprises rely on its products.
A Role Born of Scale
Keisling stepped into the Chief AI Officer role, a newly created position at Progress, around 18 months ago, after previously running engineering for the company’s Infrastructure Management portfolio.
The role, he says, exists because Progress runs 15–20 products, each innovating on AI independently, without a proper mechanism to share learnings or reusable components across teams.
Beyond coordinating that innovation, Keisling has also driven an enablement push across Progress’ nearly 1,000-strong engineering organisation, including structured training with Microsoft on tools like GitHub Copilot, to avoid what he called a common industry mistake of handing engineers AI tools without teaching them how to use them well.
The Pilot Problem
Much of the conversation returned to familiar enterprise complaint. AI pilots that dazzle in testing but stalled before reaching production.
Keisling says this difficulty is real and structural, not merely a case of poor execution.
“Implementing AI at an enterprise scale is proving to be much more difficult than people were expecting,” he adds, pointing to security, governance and observability requirements that don’t exist in a personal-use pilot.
As evidence, he cited the scramble by Microsoft, Amazon, OpenAI and Google to hire large numbers of forward-deployed engineers (FDEs) to help enterprise customers operationalise AI, a sign, in his view, that scaling agents safely requires specialised expertise that most organisations don’t yet have in-house.
Progress itself does not employ forward-deployed engineers, a gap Keisling believes creates an opening for systems integrators and IT services firms, including Indian majors such as TCS, Infosys and HCL, to move up the value chain.
Rather than viewing FDEs as a threat to the IT services business model, he framed long-standing client relationships as a durable advantage. “If you were giving me the choice of working with a forward-deployed engineer that really doesn’t understand my business, or a partner that I have worked with before, that has proven value, that I know is going to be there for me in the future, I think that choice is obvious.”
The India Opportunity
Progress’ India presence through GCCs, spread across Bengaluru, Hyderabad and Delhi, has roughly doubled in headcount over the past five years to nearly 800 employees, spanning engineering, marketing, sales and legal functions.
Keisling described the offices as full innovation centres rather than delivery arms, adding that generative AI’s relative newness has flattened the playing field for talent in India to lead global product work rather than simply execute it.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.