Three platform lineages converging
RPA platforms excel at UI-level automation where APIs are missing. They remain relevant for legacy desktops and citizen-developer programmes, but UI selectors are fragile and scale poorly without strong governance. Most major RPA vendors have added document AI, process mining, and agentic features to stay in the AI narrative.
iPaaS and workflow engines focus on API integrations, event triggers, and durable orchestration. They offer clearer audit trails than screen scraping and better fit cloud-first estates. AI arrives as content transformation steps, classification nodes, and hooks to external model providers.
AI-native automation platforms treat language models and tools as first-class citizens: prompt nodes, retrieval steps, agent loops, eval hooks, and human approval inboxes designed for unstructured work. They can move faster on cognitive tasks but may be thinner on decades of ERP connector depth.
Where to place intelligence in a flow
Insert model calls where inputs are messy (emails, PDFs, chat), decisions are fuzzy but low-to-medium risk, or content must be generated for humans to edit. Prefer rules, decision tables, and deterministic code where policies are explicit, calculations must be exact, or regulators expect unambiguous logic.
A robust pattern is “extract → validate → decide → act → record.” Models help extract and draft; schemas and business rules validate; decision services or humans decide; systems of record act; observability records. Skipping validation is how confident nonsense enters finance systems.
Feature comparison lens
| Capability | Why operators care | Questions to ask vendors |
|---|---|---|
| Durable execution | Long-running processes survive failures and restarts | How are retries, idempotency, and poison messages handled? |
| Connector depth | Time-to-value hinges on maintained integrations | Who owns connector SLAs when APIs change? |
| Human task UX | Exceptions are the product, not an afterthought | Can reviewers see model rationale and source snippets? |
| Policy controls | Prevents uncontrolled agent spend and actions | Are there step budgets, allow-lists, and environment locks? |
| Observability | Needed for ops, security, and continuous improvement | Can traces export to your SIEM and data lake? |
Pricing and packaging realities
RPA often bills by bot or attended/unattended capacity. iPaaS may bill by recipes, tasks, or data volume. AI-native tools layer model consumption—sometimes marked up—on top of platform seats or workflow runs. Process-mining add-ons and premium AI packs can double initial quotes.
Build a unit economic model on successful business transactions (invoice posted, employee onboarded, claim routed), not on platform abstractions. Include failure rework and licence shelfware; unused bot licences are a famous silent tax. Negotiate flexibility to reassign capacity across departments as priorities shift.
Pros and cons by approach
Extend existing RPA/iPaaS with AI steps
Pros: leverages skills, connectors, and governance already paid for; smoother audit continuity. Cons: AI features may lag specialists; modelling long agent loops can feel bolted on; licensing surprises when AI packs unlock.
Adopt an AI-native automation fabric
Pros: better primitives for prompts, tools, and evaluations; faster experimentation. Cons: migration cost, connector gaps, cultural clash with existing CoE standards, younger compliance track records.
Compose best-of-breed
Pros: right tool per job; avoids megavendor lock-in. Cons: integration overhead, fragmented identity and logging, more vendor management.
Risk controls that belong in the platform, not the prompt
- Environment segregation with promotion workflows for automation releases.
- Secrets management that never places long-lived credentials in prompt text.
- Action allow-lists by role, system, and monetary threshold.
- Rate and spend caps on model endpoints per workflow.
- Mandatory human approval states for externally visible communications.
- Immutable logs linking business case ID to each tool invocation.
Operating model: CoE 2.0
Automation centres of excellence must expand from bot development standards to evaluation datasets, prompt/version control, and model risk tiering. Citizen developers can still build, but AI steps should pass automated tests analogous to code linting: schema checks, toxicity/policy screens where relevant, and golden-task regressions on deploy.
Pair every production workflow with an owner metric—cycle time, error rate, cost per transaction, employee hours returned—and a quarterly kill-or-refactor review. Automation that cannot prove value becomes liability.
For the cognitive components inside these platforms, see leading AI models and intelligent agents. For market packaging of full solutions, see agentic services.