RPA development company: the licence maths and the 1 November 2026 credit cliff
Summary. Most failed RPA programmes fail on arithmetic, not on robots. In Microsoft Power Automate, a Premium user licence at $15 per user per month carries 40,000 actions per day and exactly one attended bot; running anything unattended requires a separate Power Automate Process capacity licence at $150 per bot per month, which carries 250,000 actions per day. Up to 10 Process licences can be stacked on a single cloud flow, each adding 250,000 actions, or one licence can be assigned to a flow group to share 250,000 actions across up to 25 flows, with stacking unavailable in that mode. Those numbers, published by Microsoft Learn and last updated on 20 July 2026, decide the architecture. One more date decides the AI parts: the seeded 5,000 monthly AI Builder credits that ship with these licences are removed on 1 November 2026. eCorpIT scopes RPA against that maths before writing a single flow, because the wrong licence shape costs more than the wrong code.
Where bots still beat agents
The interesting question in 2026 is not "RPA or AI agent". It is which steps of a process want a deterministic script and which want a model. Getting this wrong in either direction is expensive.
A rule-based bot is still the right tool when the step is high-volume, the UI or API is stable, the decision is fully specified, and someone will eventually audit the run. A bot that logs into the same portal 4,000 times a month and types the same eleven fields does not need reasoning, and paying per token for reasoning it does not need is a permanent tax on a solved problem. Deterministic also means reproducible, which is what an auditor asks for.
A model-driven agent earns its cost when the input is unstructured, the decision depends on judgement, or the target application changes shape often enough that selector maintenance eats the savings. Invoice lines that arrive in forty formats, exception handling that needs a human's reasoning written down, a supplier portal that redesigns twice a year: those are the cases where a brittle script costs more to maintain than it saves.
Most real back-office processes are a mixture, and the design job is drawing the line, then wiring the two together. We wrote about the cost model for the agent side of that boundary in our analysis of replacing RPA with computer-use agents, and the conclusion there holds: the choice is per step, not per programme.
The licence boundary is the architecture
This is the part that surprises teams, and it is documented rather than folkloric. In Power Automate, attended and unattended are not two settings on the same licence. They are two different products.
| Premium (user) | Process (capacity) | Hosted Process (capacity) | |
|---|---|---|---|
| Daily action limit | 40,000 per user | 250,000 per licence | 250,000 per licence |
| Attended RPA | One attended bot | Not included | Not included |
| Unattended RPA | Not included | One unattended bot | One unattended bot |
| Hosted RPA | Not included | Not included | One hosted bot |
| Process mining | Included | Not included | Not included |
| Dataverse database storage | 250 MB per licence | 50 MB per licence | 50 MB per licence |
Three consequences follow, and each one has sunk a project somewhere.
You cannot buy capacity licences alone. Allocating a Process licence to a machine, which unattended mode requires, still presupposes that the machine was registered by a Power Automate Premium user. And to trigger attended or unattended desktop flow runs from cloud flows operating under a Process licence, the user whose connection runs the desktop flow needs a Premium licence. A "we will just buy bots" plan does not work.
Capacity is bought in units of 250,000 actions per day, and the shape of the purchase matters. Stacking up to 10 Process licences on one cloud flow gives that flow up to 2,500,000 actions per day. Assigning a single licence to a flow group instead shares 250,000 actions across up to 25 cloud flows, and stacking is not available for flow groups. Microsoft's own guidance is to estimate actions per run multiplied by runs per day and buy to cover that volume. That estimate belongs in the discovery phase, not in the first billing surprise.
Process mining is a Premium feature. It is not included with Process or Hosted Process licences, and its storage allowance is 50 MB per Premium licence, capped at 100 GB per tenant. If the plan is "mine the process first, then automate it", the mining work needs user licences that the automation itself does not.
The 1 November 2026 credit cliff
Any RPA design that quietly depends on AI Builder needs a budget line dated 1 November 2026.
Today each Power Automate Premium, Process and Hosted Process licence includes 5,000 AI Builder credits per month, and those credits are what pay for form processing, object detection and text classification inside a flow. Microsoft's end of AI Builder credits guidance removes the seeded credits on 1 November 2026 for all new and existing customers, including those on an Enterprise Agreement. After that date the seeded allocation is gone and only purchased AI Builder add-on capacity remains active, with Copilot Studio credits as the route for continued access.
The practical effect: a document-processing flow that has been free at the margin for two years acquires a per-unit cost in November. Extraction volumes that nobody has ever measured suddenly need measuring. We build the credit consumption estimate into the design now rather than discovering it in a November invoice.
The other platform, and why the maths differs
Power Automate is not the only option, and the licensing model is genuinely different elsewhere. UiPath's Unified Pricing licensing framework collapses multiple consumables into a single Platform Unit, with a separate Flex Plan as the alternative, and VM Automation Cloud Robots consume Robot Units on an annual basis rather than a per-bot monthly fee. UiPath's pricing page lists entry plans starting at $25 per month.
The comparison that matters is not the sticker price on either platform. It is which consumption model matches your load. A Power Automate Process licence is a flat monthly cost with a hard daily action ceiling, which suits steady, predictable volume. A unit-based model absorbs spiky volume more gracefully and punishes idle capacity less. A process that runs 200 times on the last two days of every month and never otherwise prices very differently under the two.
We are Microsoft partners and most of our RPA work lands on Power Automate, but the platform recommendation comes out of the volume profile, the existing Microsoft 365 estate and the security model, not out of a preference.
What eCorpIT builds
Back-office process automation across finance, HR and operations: three-way invoice matching, bank reconciliation, payroll input assembly, employee onboarding and offboarding, vendor master maintenance. Data movement between systems that have no usable API, which is the original and still the best justification for RPA. Report assembly and distribution on a schedule. Regulatory and compliance filings where the submission portal is a web form and the deadline is fixed. Legacy application bridging, where a bot is genuinely cheaper than an integration project against a system that will be retired in three years.
Alongside the bots we do the parts that decide whether a programme survives its first year: exception handling and human-in-the-loop design, a work queue model, run logging that an auditor accepts, and monitoring that tells someone the bot failed before the business notices.
Where an API exists, we will usually recommend building against it instead. A bot driving a UI is a workaround for a missing interface, and our API development company page covers the alternative. Honest advice at the scoping stage is cheaper than a screen-scraper that breaks on the next release.
How we run an RPA engagement
1. Measure the process before automating it. Actual volumes, actual cycle times, actual exception rates, and the actions-per-run figure that the capacity purchase depends on. Processes that turn out to run 30 times a month do not justify a bot; the discovery is meant to find those and stop.
2. Fix the process, then automate it. Automating a broken process makes it fail faster. Where a step exists only because a form was designed badly in 2014, we say so.
3. Choose the licence shape deliberately. Attended or unattended, user licences or capacity licences, stacked Process licences or a flow group, and the AI Builder consumption estimate for after 1 November 2026. This is a written decision with numbers attached, not a default.
4. Build with failure as the first-class case. Every bot gets a work queue, an exception path with a named human owner, idempotent retries, and structured logs. The happy path is the easy half.
5. Hand over the operating model. A runbook, a monitoring dashboard, an owner for selector maintenance, and a review cadence. RPA without an owner degrades quietly, and the first sign is usually a month-end that did not close.
Our stack
Power Automate cloud flows and desktop flows, with Dataverse for state and work queues, and on-premises gateways where the target system is inside the network. UiPath Studio and Orchestrator where an estate is already there. Python for the transformation steps that are cleaner in code than in a designer. Azure for hosting and identity, since most of this work sits inside a Microsoft tenant already. Structured logging into whatever the client's observability stack already is, rather than a second dashboard nobody opens.
Test automation discipline carries over directly from our QA and test automation service work, because a bot driving a UI has the same fragility profile as a UI test and benefits from the same defences. Where the automation grows into a product rather than a set of scripts, the custom software development company page describes that path.
Why eCorpIT
eCorp Information Technologies Private Limited has built software from Gurugram since 2021. We are CMMI Level 5 appraised, ISO 27001:2022 certified and MSME registered, and we partner with Microsoft, AWS, Google, Shopify, Kaspersky and Desktime. Engagements are senior-led.
The reason to bring in a partner for RPA is not the scripting. It is that the licence model, the capacity maths and the exception design are where the money goes, and all three are decided in the first two weeks by whoever is in the room. A team learning the platform on your process will get the flows right and the licence shape wrong.
What drives the price
Four variables. The number of distinct applications a bot must touch, because each one is a separate integration surface and a separate failure mode. Whether the target systems have APIs, which can turn a three-week bot into a one-week integration. The exception rate, since a process with 30% exceptions needs far more human-in-the-loop design than one with 2%. And whether unattended execution is required, which changes the licence bill rather than the build effort.
We scope in three shapes: a fixed-price process assessment that returns volumes, a licence recommendation with numbers and a build estimate; a per-process build, priced per automation; and a retained engagement covering monitoring, selector maintenance and platform changes such as the November credit change. Ongoing ownership matters more here than in most software work.
India-specific considerations
Two things shape RPA economics in India specifically. Labour arbitrage is real, so the business case has to be honest: for a low-volume process, three people doing the work manually in India can be cheaper than a bot plus its licence plus its maintenance. We would rather say that during assessment than build something that never pays back. The processes that do pay back here are the high-volume, deadline-bound ones, especially statutory filings and reconciliation, where the cost of a missed deadline dominates the labour cost.
Second, bots handle personal data by definition, since they read the same screens a clerk reads. Under the Digital Personal Data Protection Act 2023, that means a bot's logs and screenshots are themselves personal data and need retention limits, access control and a deletion path. We design automations aligned with DPDP requirements: no screenshots of personal data in default logging, credentials in a vault rather than in a flow, and a documented retention window. Teams building a wider delivery footprint here can start from our software development company in India page.
FAQ
Do we need a separate licence for unattended RPA?
Yes. In Power Automate, a Premium user licence includes one attended bot and no unattended capability. Unattended runs need a Power Automate Process or Hosted Process capacity licence, each of which includes one unattended bot. The machine still has to be registered by a Power Automate Premium user first.
How many actions does a Power Automate licence allow per day?
A Premium user licence carries 40,000 actions per day. Each Process or Hosted Process capacity licence carries 250,000 actions per day. Up to 10 Process licences can be stacked on one cloud flow, adding 250,000 actions each, giving a maximum of 2,500,000 actions per day for that flow.
What changes on 1 November 2026?
Microsoft removes the seeded 5,000 monthly AI Builder credits that currently ship with Power Automate Premium, Process and Hosted Process licences. The removal applies to new and existing customers, including Enterprise Agreement customers. After that date only purchased AI Builder add-on capacity remains active, with Copilot Studio credits as the alternative route.
Is RPA obsolete now that AI agents exist?
No, but the boundary moved. Deterministic bots remain cheaper and more auditable for high-volume steps with stable interfaces and fully specified rules. Model-driven agents earn their cost where inputs are unstructured, judgement is required, or the target application changes often enough that selector maintenance exceeds the savings.
Should we automate a process or build an integration?
Build the integration where a usable API exists. A bot driving a user interface is a workaround for a missing interface and inherits every fragility of that interface. RPA is the right answer when there is no API, when the system is being retired anyway, or when the integration project would take longer than the process has left.
How does UiPath licensing compare with Power Automate?
Differently enough that volume profile decides it. UiPath's Unified Pricing framework consolidates consumables into a single Platform Unit, with a Flex Plan alternative, and VM Automation Cloud Robots consume Robot Units annually. Power Automate charges a flat monthly fee per bot with a hard daily action ceiling, which suits steady predictable volume better than spiky volume.
What makes an RPA programme fail after go-live?
No owner for selector maintenance, no exception path with a named human, and logging that nobody monitors. Target applications change, and a bot that silently stops is usually discovered at month-end close. We hand over a runbook, a monitoring dashboard and a review cadence for exactly this reason.
How eCorpIT can help
We start RPA work with a process assessment that returns real volumes, an actions-per-day figure, a licence recommendation with the numbers attached and an AI Builder consumption estimate for after 1 November 2026. From there we build per process, with work queues, exception paths and audit-grade logging as part of the definition of done, then keep the automations running under a retained engagement if you want ownership handled. Senior engineers run the work. Tell us which process is costing you the most at /contact-us/.
References
- Types of Power Automate licenses - Microsoft Learn, updated 20 July 2026
- End of AI Builder credits - Microsoft Learn
- Licensing and AI Builder credits - Microsoft Learn
- Power Automate licensing FAQ - Microsoft Learn, updated 14 August 2026
- Premium RPA features - Power Automate, Microsoft Learn
- Power Automate pricing - Microsoft Power Platform
- Power Automate pay-as-you-go pricing - Microsoft Licensing Resources
- Power Platform licensing FAQs - Microsoft Learn
- Automation Cloud unified pricing licensing plan framework - UiPath documentation
- UiPath plans and pricing - UiPath
- Orchestrator unified pricing cloud robots - UiPath documentation
- Commercial licensing plans, Unified Pricing - UiPath documentation
- Power Automate pricing and plans for 2026 - Zapier, August 2026
Last updated: 17 August 2026.
Top comments (0)