Pega: How the Low-Code Firm Is Betting on AI

Pega’s AI bet inside low-code: why “build for change” now means models, not menus

Pega spent years selling a simple promise: enterprises can ship process-heavy software without hand-coding every step. That story still matters, but the pitch has shifted. The company’s current bet is that generative AI will not replace workflow platforms. Instead, it will change what “low-code” means, and what buyers should demand from vendors. If you run software teams in a regulated business, the real question is not whether a chatbot can draft a user story. The question is whether AI can be stitched into case management, decisioning, and operations without turning production into a science fair.

Start with what low-code looks like in a large Pega shop. A bank or insurer may have hundreds of “cases” in flight: a dispute, a loan application, a claim, a fraud review. Each case has stages, routing rules, SLAs, data checks, human approvals, and audit logs. In the past, Pega’s value was that those rules lived in one place, editable by teams that sit closer to the business than a Java backend team. Generative AI changes the pressure points. Leaders now want AI to draft the workflow, generate test data, summarize customer interactions, and even propose decision logic. The risk is obvious: if AI writes the logic, who owns the logic?

That’s why Pega’s AI narrative has leaned into “AI that builds, not just talks.” Blueprint-style experiences (the branding varies across releases and demos) aim to turn a plain-English idea into a starter app. In practice, that means you describe the process, the personas, and the data objects; the system drafts a case type, UI skeleton, and suggested steps. The buyer-friendly angle is speed. The engineering angle is control: the generated artifacts still land inside Pega’s governed framework, where you can review each rule, run tests, and track changes through environments. The bet is that AI-assisted generation only matters if it ends inside a system that already knows how to handle audits, versioning, and production releases.

A concrete way to evaluate this: imagine a mid-size health insurer, “Harborline Health,” that wants to modernize prior authorization. The old process lives in email threads and a legacy portal. Harborline’s product team wants a working flow in weeks, not quarters. With a prompt-to-app style tool, a business analyst can draft the stages (intake → clinical review → provider outreach → decision → appeal) and list required documents. The AI can propose screens and validations. Then a lead engineer reviews the generated rules, adds integration steps to the claims system, and gates anything risky behind roles. In that scenario, the AI is not the system of record. Pega remains the system of work, with AI acting like an accelerated scaffolding layer. ✅ The insight to keep: the win is not “AI wrote an app,” it’s “AI drafted an app that still ships through the same governance pipeline.”

GenAI in Pega Infinity and GenAI Studio: what it actually changes for builders

Pega’s GenAI messaging often gets summarized as “AI inside the platform,” but that phrase hides the practical shifts. For builders, the meaningful changes land in three places: authoring, operating, and testing. Each one has a different risk profile and a different ROI story.

On the authoring side, GenAI-style assistants can draft user stories into case types, suggest data objects, propose decision rules, and generate UI text. The most credible use is the boring one: converting half-formed requirements into a structured starting point. A product manager might paste a policy doc and ask for a step-by-step intake flow. The assistant drafts stages, assignments, and required fields. The team then trims it down. If you have lived through requirements churn, this matters because it reduces the “blank page” cost. It does not remove the need for a lead architect to make the hard calls about data contracts and integration boundaries.

On the operating side, Pega has pushed GenAI into customer service workflows, where the system can summarize conversations, propose next-best actions, and draft knowledge-base responses. That’s where low-code meets contact-center reality: high volume, tight compliance, and huge variance in customer intent. A good summary feature can cut handle time. It can also create liability if it invents facts. The best pattern is to treat summaries as drafts with citations. If the agent can click to expand “why this summary says that,” you get a workflow tool, not a magic trick. If you can’t, you get a compliance headache. ⚠️

Testing is the sleeper feature. Enterprises struggle to create realistic test data because production data is sensitive, and synthetic data often fails edge cases. A GenAI-based generator can draft test records that match a schema and a set of constraints: “create 200 claims with these diagnosis codes, include 15 with missing attachments, include 5 with conflicting dates.” That kind of targeted data generation speeds QA and helps teams reproduce bugs. The important detail is determinism: you want seedable generation and traceability, so the same test suite can rerun during CI. If the platform can log prompts and outputs as artifacts tied to a build, it starts to look like a real engineering tool.

How to judge “prompt-to-app” beyond the demo

Vendor demos often show a single prompt producing a shiny app. In real shops, prompts are not the hard part. Requirements are messy, and systems are connected. To evaluate a GenAI Studio-style feature, ask for evidence that it handles real constraints.

• 🧭 Round-tripping: After developers edit the generated rules, can the assistant still help without overwriting work?
• 🔒 Governance hooks: Can generated artifacts be reviewed, diffed, and approved like any other change?
• 🧪 Test integration: Does the platform generate tests or at least suggest assertions tied to business outcomes?
• 🧾 Auditability: Can you trace which prompt produced which rule, and who accepted it?
• 💵 Cost predictability: Can you forecast usage, or does model activity explode during peak volume?

Pega’s platform story has increasingly acknowledged the cost question, since enterprises got spooked by unpredictable token billing in early generative AI rollouts. The direction many buyers want is simple: tie AI costs to completed work or other business measures, not to invisible model chatter. The insight that matters: the best GenAI feature is the one your finance team can approve without a three-month fight. 🧩

What follows from this is a bigger question: if AI drafts the work and also helps run it, where does accountability live? That leads straight into Pega’s longer-running strength in decisioning and process analytics.

Pega Process AI and decisioning: the quieter AI layer enterprises actually buy

Generative AI gets the headlines, but many Pega customers pay the bills for a different class of AI: real-time process intelligence and decisioning tied to operational events. This matters because enterprise transformation lives or dies in the “middle mile” between front-end requests and back-end systems. That middle mile is routing, triage, compliance checks, and exception handling. If you can’t improve those loops, a fancy interface just hides the same delays.

Pega’s Process AI framing is about ingesting events as they happen and making decisions quickly: which work queue gets the item, what data gets updated, what can be straight-through processed, and what needs a human. In practical terms, this can look like a claims intake where low-risk claims get paid automatically, while suspicious ones get flagged for investigation. It can also look like customer onboarding where identity checks and document collection are orchestrated across vendors. The models here are often not LLMs at all. They can be rules, scorecards, or classic ML models. The “AI” is the orchestration plus feedback loop, not a single magic endpoint.

Return to Harborline Health. Prior authorization has a backlog problem. Some requests are missing documents; others are duplicates; a subset are urgent. Harborline wants to cut average turnaround time. A Process AI approach starts with instrumentation: events from the provider portal, EHR integrations, call center notes, and document intake. Then the system classifies requests into lanes. The easy ones get auto-routed and checked. The ambiguous ones get a human assignment with a checklist. Over time, the organization tunes the policy rules and the model thresholds based on outcomes. Generative AI can help at the edges: summarizing provider notes or extracting fields from PDFs. But the core value is that the workflow engine can act on events and enforce policy consistently.

A practical comparison: GenAI features vs process decisioning

Buyers often blend these together. It helps to separate them, because you will staff and govern them differently. Here is a buyer-focused table that matches common asks to the risks and controls.

The big operational lesson: generative AI features tend to live in the authoring and assist layers, while process AI lives in the execution layer. If you treat them the same, you either over-govern the assistant or under-govern the engine. ✅ The insight to keep: the AI that changes your metrics is usually the AI that changes your routing decisions, not the AI that writes your meeting notes.

With that foundation, the next piece is money. Pega and its peers learned that AI pricing can kill adoption faster than any technical issue.

Once cost is predictable, security and compliance become the gating factors. That is where Pega’s enterprise posture gets tested hardest.

AI pricing and unit economics: why Pega moved away from token anxiety

Enterprise teams learned a painful lesson during the first wave of LLM adoption: token-based pricing is easy to start and hard to scale. Costs spike when usage spikes, which is exactly when your business needs the system most. That is why the market has moved toward pricing tied to outcomes: per agent seat, per workflow run, per document processed, or per “unit of work” completed. Pega has leaned into that direction in public messaging, including shifting away from token-metered agent pricing in favor of flatter models tied to completed work. That shift fits the buyer reality: finance leaders can budget for volume, but they won’t budget for “how many tokens did the model feel like using today.” 💸

If you are evaluating Pega’s AI add-ons in 2026, the right approach is to ask for a pricing model that maps to your operational metrics. A contact center can estimate conversations per day. A claims org can estimate submissions. A bank can estimate disputes. If the contract ties AI costs to those numbers, you can run a real forecast. If the contract ties costs to model internals, you can’t. That difference shapes adoption. Teams will avoid using features they can’t afford to turn on by default.

Harborline Health again makes a good lens. Suppose prior authorization volume is steady most months, then spikes during open enrollment and flu season. If AI summarization is priced per token, Harborline’s bill rises at the exact time call volume rises. That creates the wrong incentive: managers tell agents to stop using the tool when they need it most. If the same capability is priced per case completed, or per conversation processed within a set limit, Harborline can keep it on and focus on outcomes. The pricing model becomes a product feature.

What to ask for in an AI commercial model

Procurement teams can get stuck in abstract language. The best questions are concrete and tied to failure modes seen across the industry.

1. 📊 What is the bill tied to? A completed case, a routed item, a seat, or model usage?
2. 🧯 What happens in a spike? Are there caps, tiers, or throttles you control?
3. 🧾 Can usage be audited? Can you trace cost back to teams, apps, and features?
4. 🔁 Can you swap models? If a cheaper model works for summaries, can you choose it?
5. 🧪 Is non-prod included? Can you test safely without paying production-like rates?

Pega’s pitch is helped by the fact that many of its customers already run high-volume workflows. They have strong instincts about unit costs. The challenge is that GenAI features can create hidden volume. A prompt-to-app assistant might call a model many times per build. A summary tool might rerun if an agent edits a note. Without clear metering, teams feel the burn later.

Pricing also intersects with architecture. If Pega supports multiple LLM options, teams can match model choice to task. Use a cheaper model for routine summaries. Use a stronger one for complex extraction. Keep sensitive workloads on a private deployment if required. The insight to keep: model choice is a cost-control knob, and platforms that expose that knob will win more production rollouts.

Cost clarity brings attention back to trust: security, compliance, and change management. That is the final test for any AI inside a low-code platform.

Governance and accountability in Pega’s AI stack: shipping safely in regulated enterprises

AI inside a workflow platform creates a specific governance problem: the same system that runs customer-impacting decisions may also be generating parts of itself. That is a tighter loop than most enterprises are used to. The control plane needs to cover both runtime behavior and build-time generation. If your organization has ever been audited, you already know why this matters.

Start with the build-time side. If GenAI drafts a case type, rules, or data fields, you need a record of what changed and why. Classic software solves this with pull requests, diffs, and approvals. Low-code platforms solve this with ruleset versioning and deployment pipelines. An AI assistant must plug into that same machinery. Otherwise, you end up with “ghost changes” where the team cannot explain how a workflow evolved. The practical requirement is straightforward: generated artifacts should be first-class objects in the repository, reviewable by humans, and tied to an identity. If an auditor asks, “who approved this rule,” the answer cannot be “the model.”

Runtime governance is different. Here the focus is on outputs that touch customers or compliance records: generated summaries, drafted emails, knowledge answers, and extracted fields. If you have watched regulated teams roll out LLMs, the safest pattern is “draft + human accept,” at least for high-risk actions. For lower-risk actions, you can move toward automation with monitoring and fallback paths. Pega’s workflow heritage helps because you can design these gates as part of the process: a summary appears, the agent edits it, then clicks accept, and that click is logged. That turns a probabilistic model into a controlled step.

A mini case study: a bank’s dispute workflow with AI assist

Picture “Northlake Bank,” a regional bank modernizing card dispute handling. The bank wants faster resolutions, but it cannot afford wrong letters or sloppy notes. The team adds an AI summary step that reads call transcripts and chat logs, then drafts a dispute narrative for the case file. The case worker can accept, edit, or reject. If rejected, the system asks for a reason tag like “missing key fact” or “incorrect merchant.” Those tags feed a weekly quality review.

This design does two things. First, it protects customers, because humans remain accountable for what goes into the official record. Second, it creates a feedback dataset tied to real failures, not synthetic benchmarks. Over time, Northlake can tune prompts, swap models, or narrow the context window. The workflow engine stays stable while the AI component evolves. ✅ The insight to keep: the safest path to ROI is to treat AI as a controlled step, not a free-floating assistant.

One more governance angle is cultural, not technical. Low-code platforms already blur the line between “developer” and “business user.” GenAI blurs it further. If anyone can generate a flow, you need guardrails on who can deploy a flow. The right control is role-based access plus environment separation. Let many people generate drafts in a sandbox. Restrict promotion to production to a smaller group with on-call responsibility. That is boring, and it works.

Finally, keep an eye on the “TitleResults for ‘How to create a CRG?’” type of internal search problem. Enterprises often have tribal acronyms and half-documented processes. GenAI can surface answers faster, but only if the underlying knowledge base is curated and permissioned. If your internal content is messy, the model will reflect that mess. The platform can help by attaching sources and limiting retrieval to approved docs. Without that, you get confident nonsense in corporate language, which is the worst kind.

The next step after governance is execution: once you trust the system, you will want it to run more work automatically. That is where Pega’s bet on process-centric AI meets the reality of production operations.

Ellie Turner

I’m a Brooklyn tech journalist who spent a decade covering software, cloud and developer tooling. I started this magazine in 2023 to cover generative AI without the hype or the cynicism: testing tools on my own subscriptions and citing primary sources.