Every so often a vendor conversation earns the word "useful," and this one flirts with it. On CRMKonvo #309, Pega's Matt Healy walked into a room of skeptics and, refreshingly, did not try to sell AI as pixie dust. He sold governance. Let me explain why that is the interesting part, and where the pitch still needs a second look.

First, the setup, because it is absurd. According to Matt, ninety-five percent of Fortune 500s still run a mainframe in some capacity. Roughly thirty thousand organizations are still on Lotus Notes. Healy mentions a government claims system running on hardware funded by a grant from the JFK administration, and a separate agency contracting retirees back out of retirement homes to keep the thing alive. This is the installed base that every "AI-native transformation" slide ignores.

TL;DR

If you want to watch the full CRMKonvo, please go ahead here (optimized for smartphones) or here (optimized for tablets/computers).

Else, be my guest and continue to read.

Or do both …

The graveyard problem, now with agents

Modernization has been on the CIO agenda since CIOs were invented. What changed, is that frontier AI does not mix with data and processes trapped inside sixty-year-old systems. But even worse: if AI lets you build faster, it also lets you fill Alan Trefler's famous "application graveyard" faster than ever. The Lotus Notes graveyard of the 2000s simply reopens as an agent graveyard, or a Claude graveyard; take your pick of tombstone.

Healy does not dodge this. He cites the now-familiar numbers on AI-generated code: roughly eight times more duplicated blocks, double the code churn, and about 1.7 times the vulnerabilities versus human-written code. Convenient for a platform vendor to quote? Absolutely. Wrong? Not so much. His conclusion is the sensible one: what works for a hacker building a toy on a weekend does not survive contact with regulated, mission-critical scale.

"Model, don't code," or low-code wearing an AI hat?

This is the core message. Do not let AI generate an application from the foundation up. Use AI to translate business requirements into a model of the business: the processes, the decisions, the rules. Then let a platform run that model consistently, so ninety percent of the application behaves the same across hundreds of apps and only ten percent is specialized.

Revolutionary? Not quite: this is model-driven development, wearing an AI hat. The new and genuinely valuable move is using AI early: analyzing legacy systems, gathering requirements, researching regulations, and doing it grounded on curated best practices rather than, in Healy's words, "who knows where out there on the internet." That is where the productivity actually lives.

Predictability is the whole ballgame

Businesses want predictability, and probabilistic models are by definition not predictable. Healy's stance is the adult-in-the-room part of the episode. Keep orchestration and governance deterministic; confine agents to tasks like summarization, document handling, content generation, and research; and check their work. Crucially, produce visual, explainable models rather than millions of lines of machine-translated code that no auditor can read. When only about ten percent of enterprises have compelling AI in production, and the two roadblocks are cost and risk, "explainable and deterministic" is not a nice-to-have. It is the entire permission slip.

The pricing tell

Healy spells it out. Token-based pricing measures how much thinking the model does, which is not tied to value at all, and this incentivizes vendors to make the model think more. Pega's, as well as other vendors’, counter is outcome-based pricing: pay per claim, or a percentage of your cost per claim, and burn as many tokens as you like. This is the most buyer-aligned idea in the whole conversation. One caveat is of course that outcome pricing is also a lock-in and margin play, and the real negotiation lives in how you define a "claim" or a “resolution” and what today's baseline cost supposedly is.

The three-month miracle

Then the headline claim. An insurance was quoted by a system integrator seven years and roughly $25 million for a lift-and-shift from COBOL to Java. Pega plus it’s AWS tooling, we are told, produced a working application in three months.

The lift-and-shift critique is dead right: machine-translating COBOL into Java that nobody can read just translates old debt to new debt in the cloud. But "a working application based on their mainframe in three months" is an extraordinary claim resting on a single vendor-chosen reference that also happened to speak at Pega's own event. Before anyone budgets around that number, define "working." Which of the seven systems? What stayed on the mainframe? Who maintained it the day after go-live? And note the honest piece Healy volunteers: mainframe-zero is a fantasy. High-volume, low-latency payment processing stays put. The play is to extract the customer-facing, longer-running workloads and leave the transactional core alone.

The debt you cannot see is architectural

Which brings us to the point that vendors would rather skip: Gartner's observation (behind paywall)  that technical debt is increasingly becoming architectural debt. That distinction is important, because swapping COBOL for something modern is a code problem, while re-cutting your solution architecture is a business-continuity problem. You cannot simply stop the enterprise, rebuild the plumbing, and switch it back on. Healy's answer is reasonable if unglamorous: plan top-down for where the business must be in one, three, and five years, let AI do the bottom-up archaeology of what you actually have, and meet in the middle. His genuinely useful framing is that this rationalization work, the sort of thing that used to eat six to twelve months of enterprise-architecture effort, can now be compressed into roughly two weeks. That is the compression worth paying for. Just do not let the same speed refill the graveyard with agents nobody governs.

Before you sign anything: three notes for CX buyers

Buy the archaeology, scrutinize the miracle. The lowest-risk, highest-certainty value is in using AI for legacy analysis, requirements gathering, and regulation research. Fund exactly that. Treat "three months, one platform, done" stories as scope-defined case studies, not as your project plan. Ask what "working" means, what was left running on the mainframe, and who owns maintenance after the confetti settled. A demo is a promise; a reference is a data point; neither is your architecture.

Make explainability and determinism contractual, not aspirational. For regulated CX, the probabilistic parts belong at the task level, boxed in and checked, while the process and governance stay deterministic and auditable. Insist on visual, explainable models. If your vendor cannot show a regulator how a decision was reached, understand that you own that risk, not them. "The AI decided" is not a defense you want to offer an auditor.

Price for outcomes, own the baseline. Outcome pricing beats a token meter hands down, so push for it. Just remember the leverage sits in the definitions: what counts as a claim or resolution, what today's cost really is, and what happens when volumes move. Bring finance and a hard-nosed controller into the room early, not after the first invoice. And govern the citizen-developer and agent sprawl from day one; the alternative is watching your Lotus Notes graveyard reopen under new management.

An unusually grounded hour. Healy sells discipline, not magic, and that alone puts this ahead of most vendor pitches. Just read the three-month case study with your glasses on.

The news

At its PegaWorld conference in Las Vegas on June 8, 2026, Pegasystems announced Pega Infinity 26, which it says will be available in Q3 2026. The principal change is commercial: Pega is moving away from per-token pricing for its AI agents toward a flat charge per completed "case," which it defines as a task carried out from start to finish, such as a customer changing an order, a loan approval, or a claim. Pega frames the move as removing what it calls the "AI token tax".

The pricing change rests on an architecture Pega calls Predictable AI. Reasoning-heavy AI work is concentrated at design time, when workflows are authored in Pega Blueprint and the new Infinity Studio. At runtime, a lighter-weight model identifies the user's intent, selects a pre-approved workflow, and executes it step by step; where an individual step requires a language model, for example to parse a document or summarize a prior interaction, that step is given bounded instructions rather than open-ended latitude. Pega gives two reasons: more consistent outcomes, because agents follow approved workflows rather than re-reasoning each request, and more predictable cost, because the heavier processing happens only once during design rather than on every transaction.

The architecture is not new to this release. Pega introduced Predictable AI Agents in May 2025 and integrated them into Pega Infinity '25, which reached general availability in December 2025. Infinity 26 primarily adds the outcomes-based pricing model, alongside a companion announcement that exposes Pega processes as Model Context Protocol (MCP) servers, allowing third-party agents from Anthropic, OpenAI, Google, and AWS to call them under Pega's governance controls. The release cites no named customer, quotes analyst Liz Miller of Constellation Research. The "more than 20x" savings figure comes from Pega's AI Token Cost Calculator and is qualified as applying "depending on workflow complexity and scale".

The bigger picture

Two industry currents explain the timing of this announcement.

The first is pricing. The customer-service software market has spent the past year and a half moving away from per-seat and per-token models toward charging for outcomes. Intercom Fin charges $0.99 per resolution. HubSpot cut its customer agent to $0.50 per resolved conversation in April. Zendesk runs around $1.50 per automated resolution on committed volume and has been selling outcome-based pricing since 2024. Salesforce launched Agentforce at $2.00 per conversation, a unit so loose that only roughly 8,000 of its 150,000-plus customers adopted it, which forced a pivot to per-action Flex Credits and Agentic Work Units. Sierra, Decagon, and Ada all sell per-outcome on custom enterprise contracts. Gartner, in a March 2026 forecast, projects that the cost of running inference on a trillion-parameter model will fall more than 90% by 2030, while noting that those provider-side savings will not fully reach customers and that agentic models consume between 5 and 30 times more tokens per task than a standard chatbot. Not all of it will reach the buyers, though. The unit price of thinking is falling while the number of units per task climbs, which is the squeeze every vendor in this market is now pricing against. Pega's per-"case" charge belongs to this trend, with its unit defined differently from a customer-service "resolution": a case spans a back-office task such as a loan approval or an insurance claim run end to end, rather than a single support interaction.

The second current is a deep disagreement across the industry about how much freedom an AI agent should have at runtime. One camp ships prompt-based tooling and lets agents reason and plan at each step, treating flexibility as the key point. Another constrains agents to pre-approved workflows and treats unbounded runtime reasoning as a liability, especially in regulated processes. Pega sits firmly in the second camp, and its CEO has said publicly that competitors asking users to write prompts are setting themselves up for trouble. The context underneath the argument is not trivial. A widely cited 2025 MIT study from its NANDA initiative found that roughly 95% of enterprise generative AI pilots produced no measurable return on the profit line, which the authors attributed less to model quality than to a "learning gap" in how organizations integrated the tools. This is the line the market is arguing about right now, and the vendors have started to pick sides.

My point of view and analysis

Start with the part Pega frames as leadership. On price, Pega is not leading, it is catching up, and the per-"case" charge is the same outcome-based move the customer-service vendors made first, just dressed for a different room. Credit where it is due, however, because the chosen unit is better than most: a completed back-office case is harder to game than a support "resolution" and maps to work a CFO already values. That is a real distinction. It is also a modest one, and it is not a first.

On the architecture, Pega's CEO is not entirely wrong about the risk he is arguing against. Letting a model improvise its way through a regulated claims process is asking for trouble, and the graveyard of failed genAI pilots is full of companies that could not audit what their agents did. The trouble is that the cure and the original promise of agentic AI pull in opposite directions.

Here is the question I cannot get my head around. There is real value in customer interactions that follow a rote path, and a great deal of work is exactly that; so Pega serving the rote case cheaply and consistently is a good thing, period. But the value of an agentic system was supposed to be the other case: the request that does not fit the workflow as designed, the genuinely novel situation. Pega's architecture is built to do the opposite of reasoning through those at runtime. So how does the system know it can safely run the rote workflow if it never reasons through the case at the outset? Pega's answer is the lightweight intent query that does the routing, which means the only runtime intelligence in the loop is intent classification, and classification is itself probabilistic and perfectly able to misroute. A request that matches no workflow then has three exits: forced onto the nearest approved path, escalated to a human, or handed to Blueprint to generate a workflow on the fly. However, that third option is the one Pega spends the whole pitch warning against, because runtime generation in a regulated process is precisely what it calls dangerous. You cannot headline determinism and keep on-the-fly generation as the safety valve without owning the contradiction.

There is a distinction underneath all of this. Deterministic guardrails wrapped around a probabilistic system set the boundaries of acceptable action without collapsing the space inside them. The agent still reasons; it simply cannot climb the fence. Pega is doing something else. At runtime, the approved space is the entire space. There is no reasoning inside the fence, because the fence is the answer. That is not an agent operating within guardrails. It is a workflow engine with a probabilistic front desk. For loan approvals and claims that may well be the right trade, and it should simply be named as one. The industry spent two years insisting agents would handle the unscripted long tail, and Pega's bet is that the long tail is where you get hurt, so it designed the long tail out. They may be right about the risk while conceding the promise without saying so. This is BPM, Pega's home turf since 1983, with an AI intake layer on the front. Calling it agentic is generous.

So here is what I would do before believing the deck. Ask Pega for one named production customer, on the record, who has run this at scale and watched the cost curve flatten, because a calculator output is not a reference you can phone. Then get the definition of a billable "case" in writing, including what happens when the workflow misroutes, fails, or escalates to a human, because "resolution" was always a vendor-defined word and "case" is no different, and that ambiguity surfaces on the invoice rather than in the contract. Finally, ask the uncomfortable one: what share of your real request volume does not map cleanly to a pre-approved workflow today, and what does Pega do with that slice? If the answer is "a human takes it" or "Blueprint writes a new one live," you are buying a very capable workflow engine, which may be exactly what you need, as long as you buy it with your eyes open.

The token critique landed because it is true, and the architecture is sensible for the work Pega is aiming at. I am just not convinced the market asked for agents that are forbidden from thinking the moment a request gets interesting, and I would like to know whether buyers are actually asking for this or whether the industry has decided the long tail was a bad idea all along.