For a single fixed UI process over 12–24 months, desktop RPA is usually the cheapest and most predictable option. Cloud model calls only pay when they remove measured reviewer hours or handle messy text reliably. Hosting a model locally removes per‑call fees but shifts cost into ops and servers.
- For deterministic, UI-driven work, desktop RPA usually has the lowest 24‑month total cost and the fewest operational outages.
- Cloud model calls win only when they measurably remove reviewer hours or handle heavy unstructured text better than humans do.
- Hosting a model locally removes per‑call fees but typically costs more once you budget continuous ops, energy and hardware.
- Compare options by the human hours they free, not by headline per‑call or license prices alone.
- Ongoing failure modes — UI drift, nondeterministic outputs and outages — create the majority of long‑term cost.
- Run a 24‑month worksheet that includes development, licenses, monitoring, incident fixes and ops before you decide.
Signs this is happening in your business
- Several people spend multiple hours each week copying identical fields between systems.
- Automations break after app layout changes and remain unfixed for days.
- You pay per‑use cloud fees while a person still rechecks many results.
- IT adds servers for experimental models and the project stalls for months.
- Finance cannot explain the monthly automation bill without a spreadsheet.
- Operations keeps a single person who 'knows how it works' to babysit the tool.
- Managers reroute cases to humans when automation output looks wrong or inconsistent.
If several of these are true at the same time, the problem is usually the process rather than the people running it.
What this is costing you a year
(Team hours per week x hourly cost of employees x 52) = annual payroll loss
- People doing this work4
- Hours each per week6
- Hourly cost of employees$45
- Weeks52
the annual payroll cost of four people spending six hours weekly on the task
Find the number before you spend anything
Run the formula above on your three most repetitive processes. If any of them clears a year of a salary, that is the one to bring to a call.
You are losing payroll on a fixed task and must choose the cheaper, safer automation
Four people spending six hours each weekly on a predictable data‑copy task costs real payroll. I compare 24‑month totals for desktop automation, cloud model calls, and local hosting so you can pick the option that reduces payroll and operational outages for a single fixed process.
Desktop RPA vs Cloud LLM cost comparison: short answer
Desktop RPA typically costs less and causes fewer operational failures over 12–24 months for one deterministic process. Cloud model calls are a variable cost that only justify themselves when they remove measured human review or handle messy text; local hosting usually shifts spend into ops and hardware.
A 24‑month worked TCO: desktop automation, cloud model calls, local model hosting
I present three 24‑month cost models with the inputs visible so you can swap numbers for your case: desktop automation, cloud model calls, and local model hosting. Change runs per week, reviewer time and hourly rates to see which option wins for your operation.
Desktop automation: where the costs are
Desktop automation costs concentrate in the initial build, an unattended license and routine monitoring and fixes; those are mostly fixed over two years. I list the developer hours, license and ongoing monitoring so you can see the arithmetic and how the total is reached.
- Development: 80 hours × $120/hr = $9,600
- License: $2,000 per year × 2 = $4,000
- Monitoring: 2 hours/week × $45/hr × 52 weeks × 2 = $9,360
- Endpoint maintenance: $500/year × 2 = $1,000
- Incident fixes contingency: $1,200/year × 2 = $2,400
- 24‑month total: $26,360
A project without a budgeted weekly maintenance owner does not fail slowly; it dies on a date and still costs payroll.
Cloud model calls: where the costs are
Cloud model calls raise initial integration and monitoring effort, add per‑call charges, and often require human review for nondeterministic outputs. Below I assume a run volume so you can replace it with your own figure and observe the result.
- Development and connector work: 120 hours × $120/hr = $14,400
- Model usage: $0.06 per run × 10,400 runs/year × 2 years = $1,248
- Monitoring: 2 hours/week × $45/hr × 52 × 2 = $9,360
- Human review for uncertain outputs: 10% of runs × 3 minutes each × $45/hr × 2 years = $4,680
- Outage and incident handling: $3,000/year × 2 = $6,000
- 24‑month total: $35,688
Local model hosting: where the costs are
Hosting a model locally removes per‑call fees but adds larger upfront hardware, heavier development and continuous ops staffing. If data residency forces local hosting, budget ops staff, energy and server amortisation first—those items dominate the 24‑month bill.
- Development and model adaptation: 160 hours × $120/hr = $19,200
- Hardware purchase and amortisation: $12,000 over 24 months = $12,000
- Ops staffing: 5 hours/week × $60/hr × 52 × 2 = $31,200
- Energy / colocated hosting: $200/month × 24 = $4,800
- Support and software: $3,600/year × 2 = $7,200
- 24‑month total: $74,400
How I ran these numbers and what to change for your business
I summed 24 months of measured inputs: developer hours, license fees, monitoring time, reviewer minutes and runs per year so you can swap values. Measure runs per week, reviewer minutes per run, reviewer hourly cost and how often UI changes require a fix, then rerun the sheet for your business.
Run economics: how volume and human‑review change the tipping point
Desktop automation front‑loads fixed costs; cloud calls scale with usage. Calculate the cloud premium over desktop, divide it by your reviewer hourly rate, and you get the annual reviewer hours cloud must save to break even.
- Using the worked totals above, cloud costs $35,688 over 24 months vs desktop $26,360 — an extra $9,328 over two years, or $4,664 per year.
- If your reviewer cost is $45/hr, cloud must save about 104 reviewer hours per year (2 hours/week) to break even.
- If cloud reduces reviewer time by more than that, cloud can be economical; otherwise desktop wins.
Per‑call fees look cheap until you count the continuous human review they often require.
Reliability, delays and error economics
The real cost of automation appears when it fails: rework, SLA penalties, manual overrides and loss of trust. Quantify that cost by measuring how many runs per month need manual correction and how long each correction takes in staff hours.
- UI change: automation stops and needs a 2–8 hour fix, reducing throughput for days.
- Nondeterministic output: reviewers must check a share of runs continuously; cost equals reviewer time times review rate.
- Cloud outage: interruption forces fallback to manual processing; cost equals people hourly rate times outage duration.
Identify your biggest operational bottleneck
Most owners already know which process is the problem. What they do not have is the annual figure attached to it. That is a 15-minute conversation, not a project.
Maintenance and ops: who does the work after go‑live
Someone must own weekly maintenance. Desktop automations need selector updates and small fixes; cloud calls need monitoring and prompt rule or prompt tuning; local models require dedicated ops for patching, backups and performance tuning. Count those weekly hours in your TCO—maintenance dominates SMB projects after month one.
- Desktop: expect 1–3 hours/week for selector updates and small fixes.
- Cloud: expect 2–5 hours/week for monitoring, tuning and reviewer coordination.
- Local model: expect 5+ hours/week for patching, backups and performance tuning; budget a technical owner.
Security, data residency and compliance costs
Sending sensitive data to a cloud model can trigger legal, logging and redaction costs; keeping data on desktops or on local hosting reduces external exposure but raises ops costs. Ask your compliance person what additional logging, encryption or legal review a cloud path requires and include that effort in your TCO.
- Cloud model calls: may require redaction, anonymisation, additional logging or a paid contract feature; include legal review time.
- Desktop automation: data stays on company machines and behind existing controls, simplifying auditability.
- Local model hosting: keeps data in‑house but adds encryption, backup and patching effort to ops.
Vendor lock‑in and portability
Desktop automations are literal UI instructions and are usually easier to hand back to humans or move to another tool. Prompts, training data and model weights in cloud or local setups can be costly to extract. Document inputs, outputs and a manual fallback before you choose a vendor.
- Exportability: desktop steps are straightforward to document and hand off.
- Cloud models: prompts and curated examples are the IP you must extract on exit.
- Local models: moving model weights and training data requires engineer time and licensing steps.
Hosting a model locally turns a predictable software job into an infrastructure job that management forgets to staff.
How this fails in practice
Projects die when teams ignore maintenance and human review; the usual failure path is an initially successful build, then growing selector or prompt drift, rising tickets, reduced monitoring and eventual abandonment. That sequence creates hidden payroll costs and lost trust faster than people expect.
- Week 0–4: Automation built and deployed; initial success raises expectations.
- Month 1–3: UI change or new format causes intermittent errors; tickets accumulate.
- Month 3–6: Reviewers increase manual checks; owner is reassigned to urgent tasks.
- Month 6+: No one fixes drift regularly; the tool becomes half‑working and manual work resurfaces.
Decision checklist: which choice for which situation
Use these rules of thumb to match business attributes—sensitivity, variability, run volume, accuracy—to a recommended path. Run your numbers in the worksheet, then book a short cost review at How to work together if you want outside help.
- If data is sensitive and the process is deterministic: choose desktop automation.
- If the process needs messy text understanding and reduces reviewer hours by 2+ hours/week: consider cloud model calls.
- If you must keep data entirely on‑prem: prepare for the local model ops bill before choosing that path.
- If run volume is low to medium and you need predictability: desktop automation is usually safest.
- If you need rapid iteration with language, budget cloud calls and extra monitoring.
Next step you can take tomorrow
Run the 24‑month worksheet with your developer rate, runs per week, reviewer hourly wage and expected monitoring hours. That single exercise shows which option wins for your case. If you want help running the sheet, I offer a short cost review at How to work together and advisory services at Services.
Sources and further reading
- Two AI Chipmakers Make Up 34% of America’s Earnings Boom — Google News
- UiPath (PATH) Gets A Buy Rating, Is It Still A Bargain? — Google News
Where businesses waste the most time
- Re‑keying customer detailsPeople spend hours copying identical fields between systems each weekAutomate the typing on the desktop so existing systems remain unchangedA person still checks: A person spot‑checks new or suspicious records
- Reading inbound messages for tagsSomeone scans messages to decide which queue they belong to, taking time from higher value workUse a cloud model call to suggest tags while leaving the inbox software intactA person still checks: A reviewer confirms suggested tags for edge cases
- Manual invoice matchingStaff match invoices to orders by looking up references and copying numbersDesktop automation fills forms and highlights mismatches inside the existing accounting systemA person still checks: An accountant approves flagged mismatches
How this compares with what you already run
| Operational area | The manual way | The automated way | Annual business impact |
|---|---|---|---|
| Data entry | People copy fields between screens | Desktop automation types fields directly | Most manual hours eliminated; payroll saved |
| Unstructured text handling | People read messages and decide outcome | Cloud model suggests outcomes; humans review some | Reviewer time reduced if model is accurate |
| Scaling volume | Add headcount as volume grows | Desktop uses fixed licences; cloud costs scale per run | Predictable fixed cost vs rising variable spend |
| Outages and fixes | Staff do manual work during downtime | Automation falls back to manual or queues; cloud outage may block | Downtime causes SLA breaches and overtime costs |
| Data residency | Data stays inside company systems | Cloud sends data externally; local keeps it internal | Cloud may require compliance spend; local adds ops spend |
| Portability | A person knows the steps and can revert | Desktop steps are documentable; cloud prompts are harder to extract | Lower exit and reversion costs for desktop automation |
What the end-to-end workflow looks like
- Email arrivesA new email with customer info lands in the shared inboxStart a desktop automation when the mail client shows the message
- Extract key fieldsSubject, name, reference and attachment are identifiedAutomation reads the displayed fields and captures them
- Decide actionDecide which CRM record and stage the data belongs toCloud model call suggests the CRM stage when text is messy
- Enter data in CRMFields are populated in the CRM formDesktop automation types the fields and submits the record
- Human reviewA person checks records marked uncertainStays human
- Audit log and reportingAll actions are logged for compliance and auditsAutomation writes timestamped entries to a spreadsheet or log file
The metrics that tell you it worked
| Metric | What it tells you |
|---|---|
| Manual hours per week | Total team hours on the process; measures payroll at risk. |
| Runs per week | Count of discrete process runs each week; drives cloud variable costs. |
| Reviewer minutes per run | Average reviewer time per run; shows gains from models or automation. |
| Mean time to repair | Hours from a failure being noticed to it being fixed; converts outages into cost. |
| Incidents requiring manual fallback per month | How often automation cannot complete and humans must step in; indicates operational risk. |
| Cost per run | Variable cost per automated run combining human review, model fees and support time. |
| Automation uptime | Percentage of business hours the automation runs without failing; impacts SLA costs. |
When this is the wrong move
- When the task requires frequent subjective judgment and you cannot measure reviewer time savings.
- When your compliance policy forbids external services and you lack ops budget to host locally.
- When you cannot staff a weekly maintenance owner to keep the automation working.
- When run volume is tiny and the build cost exceeds payroll saved.
- When the process changes every few months and you should fix the process before automating.
The mistakes that cost the most
- Automating a broken processTeams automate before fixing upstream variability and the automation reproduces errors, creating extra tickets.
- Ignoring maintenanceOwnership is unclear after build; selector failures pile up and the automation is switched off months later.
- Counting only headline feesDecision‑makers compare per‑call or license prices and forget continuous human review and ops costs that dominate SMB TCO.
Frequently asked questions
How much human review must cloud model calls save to be worth the extra cost?
Divide the extra annual cloud cost by your reviewer hourly rate to get break‑even reviewer hours. For example, an extra $4,664 per year divided by $45/hr equals about 104 reviewer hours per year, or roughly 2 hours per week; cloud must save more than that to pay off.
Will hosting a model locally remove all variable costs?
No. Hosting locally removes per‑call fees but adds hardware amortisation, continuous ops staffing, energy and support costs that often exceed cloud fees for SMBs. You must budget a technical owner for patching and performance and include those hours in your TCO.
What is the single biggest hidden cost vendors forget to show?
Ongoing human time for monitoring and incident fixes. Small automations need a weekly owner and cloud solutions add reviewer time for uncertain outputs; those hours compound into real payroll spend and usually dominate TCO.
If my process is UI‑driven and deterministic, should I ever pick cloud?
Generally no. Desktop automation is cheaper and more predictable for deterministic UI work. Choose cloud only if you need substantial understanding of unstructured text or rapid iteration that desktop automation cannot handle.
How do I get this into a decision I can present to finance?
Run the 24‑month worksheet with your actual inputs: developer rate, license offers, runs per week and reviewer wage to produce a straightforward TCO. If you want help, I offer a cost review at How to work together.
Stop paying salary for robotic work
Bring your messiest process to a 15-minute audit and we will map where the hours actually go, live on screen. You leave with the number whether or not you work with me.