Agentic BPM: When Your Workflows Start Making Decisions

Forty per cent of enterprise apps will have task-specific AI agents baked in by the end of 2026. Last year? Under five per cent. That’s not incremental growth — that’s a cliff edge.

Here’s the uncomfortable part. Ninety-two per cent of decision-makers say these agents need rules-based guardrails to operate safely. Only 48% actually have them. So the real question isn’t whether workflows will start making autonomous calls. They already are. It’s which decisions get delegated, within what limits, and on whose platform — and that’s exactly where Agentic BPM enters the picture.

Twenty years of rules, then a reasoning shift

Business process management used to run on one idea: write the rule, follow it. Invoice over EUR 10,000? Senior approval. Tier 1 customer? Priority queue. Contract expiring in 90 days? Fire off a renewal notice.

Simple. Predictable. And it works fine, right up until it doesn’t.

Once the rules pile into the thousands — once exceptions start outnumbering the “normal” path — rule-based BPM becomes the very bottleneck it was built to prevent. IT can’t reconfigure the workflow engine fast enough. The business has already moved on.

Gartner’s numbers are stark: by 2028, AI agents will make 15% of daily work decisions autonomously. In 2024, that figure was effectively zero. Gartner calls it the fastest enterprise tech adoption curve they’ve tracked since cloud infrastructure took off.

There’s a name for this shift now: Agentic BPM. Not workflows that follow rules — workflows that perceive context, weigh options, act inside set boundaries, and learn from what happens next.

What “agentic” actually means (not the buzzword version)

Let’s cut through the marketing fog for a second.

Stanford HAI defines agentic AI as systems built to set or interpret goals, sequence actions, use tools, and adapt over time — largely without a human clicking “approve” at every step. That’s a real departure from the reactive chatbots most people picture when they hear “AI.”

Inside a BPM context, there are basically three tiers:

Rule-based BPM. Humans design every path. The system just executes. Anything unexpected gets kicked upstairs to a person.

AI-augmented BPM. Machine learning suggests a decision. A human still signs off. The system recommends; the person decides.

Agentic BPM. Agents reason on their own, inside governance boundaries — routing, classifying, approving, escalating. No script required. The human’s job shifts from operator to architect of the guardrails.

A 2026 paper from Dumas, Milani, and Chapela-Campa lays out the architecture: intelligent exception handling, adaptation to shifting conditions, learning from execution history, and orchestration that’s proactive rather than reactive.

MIT Sloan adds a detail that gets glossed over constantly: 80% of the actual implementation effort in agentic systems has nothing to do with the model. It’s data engineering. Governance design. Workflow integration. The AI is the easy part — the scaffolding around it is what makes or breaks the deployment.

Where it’s already working — four industries, real numbers

This isn’t theoretical. It’s running in production, with results you can measure.

Banking — NatWest Group (UK). Cora AI fielded 12.9 million retail banking conversations in 2025, up from 11.2 million the year prior. AI-drafted call summaries and complaint handling saved 70,000 employee hours. Twenty-one customer journeys now run on generative AI agents that resolve routine cases without a human handoff.

Banking — CaixaBank (Spain). Four specialised agents — market evaluator, financial health analyser, document reviewer, submission agent — handle mortgage processing end to end, each passing context to the next. The payoff: loan approvals moved 25-40% faster, and manual effort on trade finance dropped 45-65%.

Insurance — Allianz (Project Nemo). Seven agents cover claims from coverage verification through payout calculation and audit logging. Claims that used to take days now close in under five minutes, filing to human review. They built it in under 100 days.

Manufacturing — General Mills. Agents optimising over 5,000 daily shipments have banked more than USD 20 million in savings since FY2024. Waste dropped 20-30%, worth another USD 20 million a year. The number that actually matters, though: 70% of AI-generated recommendations now get accepted and executed automatically. That’s not a pilot anymore. That’s trust.

Government — Singapore GovTech (VICA). Agents classify and route more than 800,000 citizen inquiries a month across 60-plus agencies. Multi-agency decisions that once took weeks now clear in hours — no predefined category menu required.

The governance problem nobody wants to solve first

Notice something about every example above? None of them turned agents loose without boundaries. And that’s exactly where most companies trip.

Harvard Business Review Analytic Services found that just 6% of companies fully trust AI agents with core business processes, unsupervised. Twenty-five per cent are actively using agentic AI. Sixty-two per cent are still evaluating. The trust gap isn’t closing on its own.

The paradox, in one stat: 92% say guardrails are necessary. Only 48% have built them. Half of respondents named “better defining the rules and guardrails AI must follow” as their top organisational priority, showing where the real work lies.

Deloitte’s 2026 State of AI report backs these findings up. Seventy-four per cent plan to adopt agentic AI within two years. Only 21% have a governance model mature enough to actually manage it.

And the warning shot from Gartner: over 40% of agentic AI projects will be scrapped by the end of 2027 — due to cost overruns, murky ROI, and weak risk controls. Forrester puts it even more bluntly: fewer than 15% of firms will actually switch on agentic features in their automation suites this year.

The takeaway isn’t to “wait longer.” It’s that governance has to come first. Decision boundaries, escalation paths, audit trails, and human override are designed before agent one goes live, not patched in after something breaks.

Why the platform matters more than the point solution

An AI agent with no process fabric around it is just an experiment. It can classify a document, maybe draft a reply. What it can’t do is orchestrate a process that spans five departments, three legacy systems, and a stack of compliance requirements.

That’s the gap low-code BPM platforms fill. They bring the workflow context, the compliance rails, the legacy integration, the human-in-the-loop checkpoints — the things that raw AI can’t provide on its own.

Take Uniksystem’s platform. Its generative AI agents aren’t bolted onto existing workflows — they’re embedded in production, handling over 90,000 non-compliance cases a year across seven Portuguese banks. Document intelligence runs with near-100% accuracy on unstructured data. Compliance with Banco de Portugal, NIS2, and the EU AI Act is part of the governance layer itself, not a separate checkbox.

The numbers back it up: 280% ROI on process automation, a 93% jump in workflow efficiency, and new process rollouts in two to four weeks — because the governance fabric is already there, not built from scratch each time.

Gartner’s 15%-by-2028 figure isn’t really the headline anymore. The real question is shifting: not “Should we use AI agents?” but “Whose platform governs the decisions once we do?” Get that wrong, and Agentic BPM becomes just another cancelled IT initiative. Get it right, and you have the operating model for the next decade.

Bonus: Five-Point Agentic BPM Readiness Checklist

1. Map your decision landscape. Sort process decisions into purely rule-based, purely human judgment, and everything in between. That middle category? That’s where agents actually pay off.

2. Audit your data readiness. Agents can’t reason without context. If your process data is scattered across spreadsheets and email threads, that’s step zero — fix it before anything else.

3. Define decision boundaries first. For every candidate process: What can the agent decide alone? What triggers a human review? What never gets delegated, period?

4. Start small — high volume, low risk. Document classification. Routine approvals. Standard routing. Build trust there before letting agents near anything higher-stakes.

5. Pick a platform built for compliance, not bolted onto it. Audit trails, decision explainability, GDPR/NIS2/AI Act alignment, and human override — these aren’t just nice-to-haves. They’re the floor, not the ceiling.

This was written in collaboration with Uniksystem, a Portuguese company that specialises in low-code BPM platforms and AI-powered automation.