Somebody in your organisation has a slide claiming RPA is dead and agents are the future. Somebody else maintains 300 bots that quietly run payroll. Both of them will be in the same meeting on Tuesday.
Here is a position that survives contact with reality: RPA is not dead, it is being surrounded. The global RPA market was still growing through 2025, and research suggests nearly 80% of RPA deployments will incorporate generative or cognitive AI. That is not replacement, that is absorption.
The Difference That Decides Everything
A bot follows a fixed path. An agent decides the path. That single distinction tells you which tool fits which job.
If the process is stable, high volume and deterministic, the bot is not just adequate, it is better. It is cheaper, faster, fully auditable and it does not hallucinate. If the process breaks whenever the input varies slightly, the bot is a maintenance tax and an agent is the upgrade.
What to Migrate First
- Bots that break on layout changes. Screen-scraping automations that fail whenever a vendor updates their UI. Agents handle variation, which is the whole point.
- Anything with an exception queue nobody empties. If 30% of runs end in a human review pile, that queue is the actual process and it was never automated.
- Document-heavy work. Reading unstructured invoices, contracts and emails is where the capability gap is widest.
- Processes with too many if-branches. When your flowchart has forty decision nodes, you have encoded judgement badly. Let a model make the judgement and keep the bot for the clicking.
What to Leave Completely Alone
- Stable, high-volume, regulated transactions. If it runs a million times a month and an auditor reads the log, do not introduce probability.
- Anything where the bot is already invisible. Nobody complains about it, it costs nothing, it works. That is a solved problem, not an opportunity.
- Scheduled data movement. A cron job and a script beat an agent on cost, speed and reliability. Every time.
- Processes about to be retired. Migrating automation for a system being decommissioned next year is expensive theatre.
The Hybrid Shape That Actually Wins
The pattern working in production is not agents replacing bots. It is agents, bots, APIs, workflow tools and humans inside the same process.
The agent reads the messy input and decides what should happen. The bot performs the deterministic clicking. The API does anything with a proper integration. The human approves whatever is expensive to reverse. Each component does the thing it is genuinely best at, and the audit trail survives.
How to Sequence the Work
| Phase | Do this | Avoid |
|---|---|---|
| Month 1 | Inventory bots by failure rate and maintenance hours | Starting with the most visible bot |
| Month 2 | Pick the top exception queue, add an agent in front of it | Replacing the whole bot |
| Month 3 | Measure exceptions cleared without human touch | Declaring victory on volume alone |
| Month 4+ | Expand only where maintenance hours fell | Migrating stable bots for consistency |
Conclusion
Rank your bots by how much maintenance they consume, not by how important the process feels. Put agents in front of exception queues rather than replacing working automation, and leave the boring reliable bots exactly where they are. The goal is fewer maintenance hours and fewer manual exceptions, not a cleaner architecture diagram.
Frequently Asked Questions
Should we stop buying RPA licences?
Not yet, but renegotiate. Your volume mix is shifting, and licence models built for high bot counts do not fit a hybrid estate.
Can agents use our existing bots as tools?
Yes, and that is often the fastest route. Expose the bot as a callable tool so the agent decides when to fire it and the bot keeps doing the deterministic part.
How do we keep audit trails intact?
Log the agent decision and the bot execution separately. Auditors want to see what was decided, on what basis, and what then happened. Two records, not one.