Workflow automation: BPM, RPA, iPaaS, low-code or AI agents, and what each costs

Workflow automation approaches compared: BPM, iPaaS, RPA, low-code and AI agents

Anthropic's engineering guidance on agents recommends the simplest solution possible, adding complexity only when needed. In its terms, a workflow runs a language model and tools through predefined code paths.^5,16

GSA's federal process improvement playbook, published June 3, 2026, orders the work in three steps. First, eliminate waste and non-value-add activities. Second, optimize mission-critical operations. Third, automate low-value, costly tasks. Its goal for the technology step is to pick the automation technology that fits the requirements found in process and value stream mapping.^17,18,19

RPA bots: maintenance, security and decommissioning

GSA's Office of Inspector General notes that bots can perform thousands of read, write and deletion actions at high speed, which makes logic and processing errors hard to catch.²⁰

UiPath's Studio documentation says some programs, such as some web apps, have changing layouts and volatile attribute values that Studio cannot predict, so developers may have to write selectors by hand.²¹

The Inspector General's August 2024 audit of GSA's RPA program found:

• GSA had 119 active bots and 24 decommissioned bots by December 2022.^22,23
• None of the 16 system security plans reviewed had been updated, as GSA's RPA policy required, to address how bots accessed the systems. Seven of the 16 did not mention bots at all.^7,24
• RPA program management removed or modified the security requirements instead of addressing the gaps.²⁵
• No process existed to remove access for decommissioned bots, which left GSA systems and data at risk of exposure.²⁶

Low-code governance for citizen-built automations

In Microsoft Power Platform, every employee has access to the default environment. Microsoft tells administrators to secure it while keeping it open for makers' personal productivity.^27,28

Data policies in the Power Platform admin center let administrators control access to connectors. Each policy sorts connectors into three groups: business, non-business and blocked. A new policy places every connector in the non-business group until an administrator moves it.^29,30,31

Microsoft's Center of Excellence Starter Kit is no longer actively maintained. Its core functions now sit in the admin center, whose inventory view covers the apps, flows and agents built across a tenant.^32,33

TIGTA reported in September 2020 that the IRS RPA program had not established an effective governance structure or a suitable development methodology for automation projects.^34,35

AI agents vs. workflow automation

Anthropic suggests agents for open-ended problems where the steps cannot be predicted in advance. It warns that their autonomy brings higher costs and the potential for compounding errors, and recommends extensive testing in sandboxed environments with guardrails.^15,36

OWASP names the matching security risk excessive agency: damaging actions taken in response to unexpected, ambiguous or manipulated output from a language model. Its root causes are excessive functionality, excessive permissions and excessive autonomy.^37,38

Microsoft sells Copilot Studio as tenant-wide packs of 25,000 Copilot Credits at $200 a pack per month, and each completed agent action or response bills a varying number of credits.^39,40

Workflow automation pricing: published list prices

Under Microsoft's terms, a bot assigned to a machine runs one unattended desktop flow at a time, and concurrent runs on a machine need more bots.⁴⁸

In November 2023, GSA's Inspector General told GSA to track the cost of developing each bot so the program could produce objective statistics such as return on investment.^49,50

Workflow automation results: what federal audits found

The 2023 audit asked GSA for a performance evaluation process to confirm that bots work as intended, with objective, auditable measures of the hours they save. A review dated June 18, 2026 found GSA had taken the corrective actions and needed no further action.^{49,54,55,56,57}

Where workflow automation goes wrong

• Automating a broken process. GSA's playbook puts eliminating waste and optimizing operations ahead of automation.¹⁸
• Designing from the documented process instead of the real one. Process mining analyzes event data from various sources to show how processes are actually executed.⁵⁸
• Claiming savings without evidence. GSA's Inspector General found no evidence for the savings GSA reported from its bots.⁶
• Bots that break when the screen changes. Selectors for apps with changing layouts may have to be written by hand.²¹
• Bots that outlive their purpose. At GSA, decommissioned bots kept access to systems because no removal process existed.²⁶
• Shadow automations. Every employee can build in the default Power Platform environment, and connectors start in the non-business group until a policy says otherwise.^27,31
• Agents with too much reach. OWASP traces excessive agency to excessive functionality, permissions or autonomy.³⁸