Stop framing this as build versus buy and decide it layer by layer. The rent worth cancelling is the per-bot execution license, which prices a commodity. The rent worth paying is the connector library and the vendor's attestations, at least until your own control plane can produce the same evidence.
The case for cancelling platform rent is usually made on the license and usually loses on the labor. The case for keeping the platform is usually made on governance and usually understates how much governance you still have to build yourself. Both sides are pitched at the wrong layer. What a commercial automation platform sells is not mainly execution: it is a credential vault, several hundred maintained connectors, a scheduler with retry semantics that survive partial failure, role-based access control, an immutable run log, tenant isolation, a support contract with a response time in it, and a stack of attestations your auditor already knows how to read. Rebuilding execution is a weekend. Rebuilding the rest, to a standard an auditor accepts, is the actual project, and it is the part AI-assisted coding helps with least. Cancelling the platform does not cancel the control plane. It transfers it to you, along with the maintenance obligation that came with it.
Ten findings in the report, each sourced and labeled by evidence type. Six of them carry the commercial argument.
A mid-size program of forty unattended automations and 250 licensed makers runs roughly $117,000 a year at Power Automate list prices. Two additional engineers cost about three times that. Under the five-year model in this report the owned stack lands about 24 percent more expensive than the licensed one, because engineering payroll dominates both columns and the owned path needs more of it.
Two extra engineers cost roughly $360,000 a year. At Microsoft's published $150 per bot per month for unattended RPA, that is 200 automations. At the $215 hosted rate, 140. At a negotiated $75, 400. Below that count an owned stack is a strategic choice you are funding, not a saving you are booking, and the business case should say so rather than dress it up as cost avoidance.
Google's DORA program found near-universal adoption alongside a negative relationship between AI use and delivery stability, calling AI an amplifier of whatever system it lands in. GitClear's telemetry over 623 million code changes shows duplicated blocks up 81 percent and refactoring line moves down 70 percent since 2023. Veracode found models picked the insecure option in 45 percent of security-relevant tasks, with no improvement as models got larger.
UiPath closed fiscal 2026 on 31 January 2026 with $1.853 billion in ARR, up 11 percent year over year, full-year revenue of $1.611 billion up 13 percent, and dollar-based net retention of 107 percent. A retention rate that close to flat means the installed base is expanding barely faster than it is contracting. Vendors in that position discount to protect renewals.
IBM's CEO study found roughly 25 percent of AI initiatives delivered the expected return and 16 percent had scaled enterprise-wide. Later IBM work puts the share of executives who can confidently measure AI return at 29 percent. Without a pre-automation baseline there is no attribution, only testimonial, and finance discounts a testimonial to zero at the first budget challenge.
Vendor-sponsored counts put non-human identities at 45 to 1 against humans in the average enterprise, 80 to 1 by KPMG's figure, and 144 to 1 in cloud-native environments. The definitions differ and none of the instruments are published. What is consistent is the gap: roughly nine in ten organizations say their identity tooling cannot manage AI agent identities, and fewer than half have any policy for them.
Treating UiPath, Power Automate and n8n as alternatives to each other, and jointly as an alternative to writing code, is where most of these programs go wrong. Split the category by layer and the build-versus-buy answer stops being one verdict and becomes four, pointing in different directions.
Screen and API interaction with legacy systems that have no usable integration surface. This is the one place per-bot pricing is worth paying, because you are buying somebody else's obligation to keep up with UI changes in software you do not control.
You take on: screen-scraping maintenance · OS image management · per-application breakage
Moving structured data between SaaS products on triggers and schedules. The connector library is the most underestimated line in every build case: a connector is not an HTTP client, it is auth flows, token refresh, pagination, rate-limit backoff and a maintenance commitment that recurs forever.
You take on: connector authorship · credential rotation · every upstream API version bump
Guaranteeing a long-running, multi-step process survives a restart, a network partition and a deploy. Temporal, Restate and Inngest all offer self-hosted paths, and this layer changes slowly, which is the profile that rewards owning it.
You take on: state store operation · versioning of in-flight workflows · replay determinism
Model calls, tool selection, prompt versioning, evaluation and trace capture. This is the layer where your requirements diverge most from every other buyer's and where the switching cost of getting it wrong is highest. It is also where teams start last, because execution is more visible and more satisfying to build.
You take on: trace schema drift · evaluation tooling · prompt release process · model deprecation churn
The full thirty-two pages carry the five-year cost model with its inputs exposed, the seven components of a control plane worth building and why only two of them are dashboards, the eight-step baseline capture that survives an audit, the control objectives table naming the artifact an auditor will actually accept, the failure modes ranked by what ends funding rather than by what gets complained about, and four scenarios to end-2028 with tripwires any reader can watch without private information.
This isn't a vendor summary. Every sentence is labeled by what stands behind it: verified fact, vendor claim, third-party estimate, my assessment, hypothesis, or scenario. Sources are numbered and clickable. Forward-looking sections use scenarios with observable tripwires, not forecasts. It's the same method behind every market assessment I write.
Thirty-two pages, built from public sources with no client brief and no interviews. Read it in the browser or take the PDF.
Each report here answers a real question, directed and researched against public sources and evaluated against a stated assumption, then delivered as Word and PDF. If you're weighing a platform, sizing a category, or defending a number to a board, tell me the decision behind it and I'll tell you honestly whether a report is the right tool.
Commission an assessment