AI automation payback and ROI calculator
Find out how quickly an AI project pays for itself by reducing manual operations.
An estimate from published figures, checked on 10 October 2026. Not legal, tax or financial advice.
On this page
What the result means
This is the time it takes for your upfront build cost to be entirely covered by the monthly labour savings.
- Payback counts from the day the automation runs. With months to go live, the line under the result counts from signing.
- Monthly net savings are the manual cost the automation removes, less what it costs to run. If the share of hours removed is below 100%, only that share counts.
- The return over three years is the net cash after three years of running, as a share of what had to be paid back.
- The target line works backwards: the most the build can cost and still pay back within the months you aim for. Use it to judge a quote.
- Where the saved hours go decides whether the saving is cash. Fewer roles adds the exit cost the law sets in your country to what has to pay back.
How it works
Payback period is how long an investment takes to return what it cost from the savings it makes. For an automation, the investment is the build (and any exit cost when roles go) and the saving is the manual work it removes, less what it costs to run.
The tool calculates current manual costs by annualising weekly hours and dividing by 12. It subtracts the monthly AI running cost to find your net monthly savings. The upfront build cost is then divided by the net savings to determine the payback period.
- Monthly manual cost = hours per week × 52 ÷ 12 × hourly rate.
- Manual cost removed = monthly manual cost × share of the hours removed.
- Monthly net savings = manual cost removed − monthly running cost.
- Upfront cost = build cost, plus the exit cost set by law when roles go.
- Payback = upfront cost ÷ monthly net savings. If net savings are zero or less, it never breaks even.
- Return over N months = (monthly net savings × N − upfront cost) ÷ upfront cost.
- Most the build can cost for a target = monthly net savings × target months − exit cost.
- Model calls a month = requests × (tokens in × input price + tokens out × output price) ÷ 1,000,000, converted from US dollars at reference rates.
Assumptions
- A year has 52 weeks of the process, so a month is 52 ÷ 12 weeks. A full-time role is 40 hours a week; fewer roles counts only whole roles' worth of removed hours.
- Savings start the month the automation goes live and stay level. The build is paid at signing; the exit cost when roles go.
- The fully loaded hourly rate adds the employer's costs set by law at that pay, from the same country data as the site's other tools; benefits come on top.
- Model prices are each provider's published list price per million tokens, checked 10 October 2026, without caching or batch discounts unless noted. Prices outside the US are converted at the European Central Bank's reference rates of 9 October 2026; the UAE dirham at its fixed rate to the dollar.
- Example figures change with the country. The build cost is Stafflo's pilot price where Stafflo sells in that currency, and otherwise the US price at reference rates; running cost and hosting keep the brief's ratios.
- Payback shows to one decimal, in months up to three years and in years beyond.
Model prices
Every current model the running-cost estimate can use, with its provider's list prices and what the provider does with API data. Prices are US dollars per million tokens.
Anthropic
- Claude Fable 5.1 (claude-fable-5-1): $10.00 in, $0.25 cached in, $50.00 out; 1M context, 128K output; batch 50% off.
- Claude Opus 5.5 (claude-opus-5-5): $4.00 in, $0.20 cached in, $20.00 out; 1M context, 128K output; batch 50% off.
- Claude Sonnet 5.5 (claude-sonnet-5-5): $2.00 in, $0.10 cached in, $10.00 out; 1M context, 128K output; batch 50% off.
- Claude Haiku 5.5 (claude-haiku-5-5): $0.10 in, $0.01 cached in, $0.50 out; 1M context, 128K output; batch 50% off; $0.50 in and $2.50 out above 100K of prompt.
API data: doesn't train the provider's models by default. Inputs and outputs deleted within 30 days. Available on eligible models and features; the most capable models (Covered Models) keep 30-day retention. A HIPAA business associate agreement is offered. US-only inference available at 1.1× the price. (Anthropic; Anthropic; Anthropic Privacy Center)
OpenAI
- GPT-6 Astra (gpt-6-astra): $10.00 in, $1.00 cached in, $50.00 out; 1.05M context, 128K output; batch 50% off; $20.00 in and $75.00 out above 272K of prompt.
- GPT-6.1 Sol (gpt-6.1-sol): $2.00 in, $0.10 cached in, $10.00 out; 1.05M context, 128K output; batch 50% off; $4.00 in and $15.00 out above 272K of prompt.
- GPT-6 Luna (gpt-6-luna): $0.10 in, $0.01 cached in, $0.50 out; 1.05M context, 128K output; batch 50% off; $0.20 in and $0.75 out above 272K of prompt.
API data: doesn't train the provider's models by default. Abuse-monitoring logs kept for up to 30 days. Zero data retention needs OpenAI's prior approval. A HIPAA business associate agreement is offered. In-region processing in the US, EU and UAE, at 10% more. (OpenAI; OpenAI; OpenAI; OpenAI)
- Gemini 3.1 Pro (gemini-3.1-pro-preview): $2.00 in, $0.20 cached in, $12.00 out; 1.05M context, 66K output; batch 50% off; $4.00 in and $18.00 out above 200K of prompt.
- Gemini 3.8 Flash (gemini-3.8-flash): $0.75 in, $0.08 cached in, $3.75 out; 1.05M context, 66K output; batch 50% off.
- Gemini 3.5 Flash-Lite (gemini-3.5-flash-lite): $0.30 in, $0.03 cached in, $2.50 out; 1.05M context, 66K output; batch 50% off.
API data: doesn't train the provider's models by default. On the paid Gemini API, abuse-monitoring logs kept for 55 days; the free tier's data can train Google's products. No guaranteed zero retention on the Gemini API; available on Google Cloud's Vertex AI. No region choice on the Gemini API; Vertex AI's regional endpoints cost 10% more and are covered by Google Cloud's BAA. (Google; Google; Google; Google; Google)
xAI
- Grok 4.7 (grok-4.7): $2.00 in, $0.50 cached in, $6.00 out; 500K context, no published output limit; $4.00 in and $12.00 out above 200K of prompt.
- Grok 4.3 (grok-4.3): $1.25 in, $0.20 cached in, $2.50 out; 1M context, no published output limit; batch 20% off; $2.50 in and $5.00 out above 200K of prompt.
API data: doesn't train the provider's models by default. Requests and responses stored encrypted for 30 days, then deleted. Zero data retention can be switched on for a team in the console. A HIPAA business associate agreement is offered. A US-only endpoint for Grok 4.7 and 4.6, at 1.1× the price. (xAI; xAI; xAI; xAI)
Mistral AI
- Mistral Large 4 (mistral-large-4): $0.68 in, $0.07 cached in, $2.09 out; 1M context, no published output limit; batch 50% off.
- Mistral Large 3 (mistral-large-2512): $0.50 in, $0.05 cached in, $1.50 out; 256K context, no published output limit; batch 50% off; open weights (Apache 2.0).
- Mistral Medium 3.5 (mistral-medium-3-5): $1.50 in, $0.15 cached in, $7.50 out; 256K context, no published output limit; batch 50% off; open weights (Modified MIT).
- Mistral Small 4 (mistral-small-2603): $0.15 in, $0.02 cached in, $0.60 out; 256K context, no published output limit; batch 50% off; open weights (Apache 2.0).
- Ministral 3 14B (ministral-14b-2512): $0.20 in, $0.02 cached in, $0.20 out; 256K context, no published output limit; batch 50% off; open weights (Apache 2.0).
- Ministral 3 8B (ministral-8b-2512): $0.15 in, $0.02 cached in, $0.15 out; 256K context, no published output limit; batch 50% off; open weights (Apache 2.0).
- Ministral 3 3B (ministral-3b-2512): $0.10 in, $0.01 cached in, $0.10 out; 256K context, no published output limit; batch 50% off; open weights (Apache 2.0).
API data: can train the provider's models unless you opt out. Inputs and outputs kept for 30 rolling days to monitor abuse. Zero data retention on request, for supported calls on paid plans. Hosted in the EU by default; processing kept in the EU or the US on the regional endpoints at 1.1× the price. (Mistral AI; Mistral AI; Mistral AI; Mistral AI; Mistral AI)
DeepSeek
- DeepSeek V4 Pro (deepseek-v4-pro): $1.32 in, $0.04 cached in, $3.96 out; 1M context, 384K output; open weights (MIT).
- DeepSeek V4.1 Flash (deepseek-flash): $0.30 in, $0.01 cached in, $1.20 out; 1M context, 384K output; open weights (MIT).
API data: can train the provider's models unless you opt out. Inputs kept for as long as the account exists. No zero-retention option published. Personal data is collected, processed and stored in the People's Republic of China. (DeepSeek; DeepSeek)
Amazon (Bedrock)
- Amazon Nova Pro (amazon.nova-pro-v1:0): $0.80 in, $0.20 cached in, $3.20 out; 300K context, 10K output; batch 50% off.
- Amazon Nova 2 Lite (us.amazon.nova-2-lite-v1:0): $0.33 in, $0.08 cached in, $2.75 out; 1M context, 64K output; batch 50% off.
- Amazon Nova Micro (amazon.nova-micro-v1:0): $0.04 in, $0.01 cached in, $0.14 out; 128K context, 10K output; batch 50% off.
API data: doesn't train the provider's models by default. Bedrock doesn't store model inputs or outputs by default. A no-retention mode can be set for an account or project. A HIPAA business associate agreement is offered. In-region endpoints keep data in the region; US profiles stay in the US; global profiles route worldwide for about 10% less. (Amazon Web Services; Amazon Web Services; Amazon Web Services; Amazon Web Services; Amazon Web Services)
Meta (Model API)
- Muse Spark 1.3 (muse-spark-1.3): $1.25 in, $0.15 cached in, $4.25 out; 1.05M context, no published output limit.
API data: doesn't train the provider's models by default. Kept as needed for the purposes in Meta's terms; no fixed period. Zero data retention for qualified accounts, through sales. No region stated. (Meta; Meta; Meta)
Example
Example: the brief's process
40 hours a week at $25 an hour, a $15,000 build and $500 a month to run, removing all of the hours:
- Monthly manual cost: 40 × 52 ÷ 12 × $25.00 = $4,333.33.
- Monthly net savings: $4,333.33 − $500.00 = $3,833.33.
- Payback: $15,000.00 ÷ $3,833.33 = 3.9 months.
Automating a process that takes 40 hours a week at $25 an hour costs $4,333 a month manually. If the AI build costs $15,000 and runs for $500 a month, the net savings are $3,833. The build pays for itself in 3.9 months.
Example: the same process, more realistically
The same inputs, but people keep a quarter of the work (exceptions and review), and it takes two months to go live:
- Monthly manual cost: 40 × 52 ÷ 12 × $25.00 = $4,333.33.
- Manual cost removed: $4,333.33 × 75% = $3,250.00.
- Monthly net savings: $3,250.00 − $500.00 = $2,750.00.
- Payback: $15,000.00 ÷ $2,750.00 = 5.5 months.
It removes $3,250 of manual cost a month and saves $2,750 net, so it pays back in 5.5 months of running, 7.5 months after signing.
When this is wrong
This calculation treats time saved as cash saved. In reality, unless headcount is reduced, that time must be redirected to revenue-generating work to see the financial return on your profit and loss statement.
- Partial automation: few automations remove every hour. Set the share removed to what stays automatic after exceptions, review and fixes, and read the table of smaller shares.
- Running costs that grow: model prices change, volumes rise and prompts get longer. Recheck the running cost every quarter; the model prices here are dated.
- Build overruns and change: the build cost assumes the quote holds. Add the time your own team spends on testing, training and changing how the work is done.
- Exit costs above the minimum: the law sets a floor; employment contracts, collective agreements and settlement terms often pay more, and notice periods add salary.
- Tax: the payback is before tax. Whether the build is expensed or capitalised, and any relief for software or research, changes the after-tax figure.
- Exchange rates: model prices are in US dollars and move with the rate you pay.
It assumes the process can be automated at all. Check that first with the process viability checker.
Rules by country
What the law sets as the minimum cost of ending a role for redundancy, used when the saved hours go as fewer roles. Contracts and collective agreements often promise more.
United States
No federal law requires severance pay. An employer with 100 or more employees must give 60 days' written notice of a plant closing or mass layoff (the WARN Act), or owe back pay and benefits for each day missed; some states add their own rules. (US Department of Labor; US Government Publishing Office)
Canada
In Ontario, after five years' service at an employer with a payroll of $2.5 million or more, severance pay is a week's wages for each year, up to 26 weeks; it comes on top of notice of up to eight weeks, or pay in lieu. (Government of Ontario; Government of Ontario)
United Kingdom
After two years' service, statutory redundancy pay is a week's pay for each full year aged 22 to 40 (half a week under 22, one and a half from 41), with weekly pay capped at £751 and service at 20 years. Twenty or more redundancies at one site need 30 days' consultation first (45 for 100 or more). (GOV.UK; GOV.UK)
Germany
There is no general statutory right to severance. An employer can offer half a month's pay per year of service in the dismissal letter (§ 1a KSchG), but needn't. Notice rises with service to seven months after 20 years, the works council must be heard first, and larger dismissals must be notified to the Agentur für Arbeit. (Bundesministerium der Justiz; Bundesministerium der Justiz; Bundesministerium der Justiz)
India
A retrenched worker with a year's service is owed 15 days' average pay for every completed year (or part over six months), one month's notice or wages in lieu, and the employer pays a further 15 days' wages into the re-skilling fund. Managers, and supervisors earning over ₹18,000 a month, are outside these rules. (Gazette of India)
Singapore
The law doesn't require a retrenchment benefit. The Ministry of Manpower's norm is two weeks to a month's salary per year of service for staff with two years or more, and employers with ten or more staff must notify the Ministry within five working days. (Ministry of Manpower, Singapore; Ministry of Manpower, Singapore)
Philippines
A role made redundant is owed at least a month's pay, or a month's pay for every year of service if more (a part year of six months counts as a year), with written notice to the worker and the labour department at least a month before. (Department of Labor and Employment, Philippines)
Malaysia
After a year's service, termination benefits are 10 days' wages per year (15 from two years, 20 from five), but not for staff earning over RM4,000 a month. Retrenchment must be reported to the Labour Department at least 30 days before. (Department of Labour, Peninsular Malaysia; Attorney General's Chambers, Malaysia)
United Arab Emirates
There is no separate redundancy payment. The end-of-service gratuity (21 days' basic wage a year for five years, then 30) is owed on any exit after a year and is already counted in the loaded rate; notice is 30 to 90 days, or a wage in lieu. (UAE Government portal (u.ae))
Australia
After a year's service, redundancy pay runs from 4 weeks' base pay (1 to 2 years) to 16 weeks (9 to 10 years), then 12 weeks from 10 years. Employers with fewer than 15 employees are exempt, and notice of 1 to 5 weeks comes on top. (Federal Register of Legislation, Australia)
Sources
Every source this page cites (96)
Each one was opened and checked on the date beside it.
- European Central BankEuro foreign exchange reference rates9 October 2026
- UAE Government portal (u.ae)Fact sheet2024
- AnthropicGlossary: tokens10 October 2026
- AnthropicClaude API pricing10 October 2026
- OpenAIAPI pricing10 October 2026
- OpenAIModels10 October 2026
- GoogleGemini Developer API pricing9 October 2026
- GoogleGemini models9 October 2026
- xAIModels and pricing10 October 2026
- xAIModels10 October 2026
- Mistral AIPricing10 October 2026
- Mistral AIModels10 October 2026
- DeepSeekModels and pricing10 October 2026
- Amazon Web ServicesAmazon Bedrock pricing6 October 2026
- Amazon Web ServicesAmazon Nova model cards10 October 2026
- MetaModel API pricing and rate limits10 October 2026
- MetaModel API models10 October 2026
- AnthropicAPI and data retention10 October 2026
- Anthropic Privacy CenterIs my data used for model training?18 August 2026
- OpenAIYour data10 October 2026
- OpenAIDeprecations1 October 2026
- GoogleGemini API additional terms of service28 April 2026
- GoogleZero data retention in the Gemini API14 September 2026
- GoogleGemini API deprecations9 October 2026
- xAISecurity FAQ10 October 2026
- xAIRegional endpoints10 October 2026
- Mistral AIPrivacy policy3 September 2026
- Mistral AIRegional inference10 October 2026
- Mistral AIDo you use my user data to train your artificial intelligence models?31 August 2026
- DeepSeekPrivacy policy10 February 2026
- Amazon Web ServicesAmazon Bedrock FAQs10 October 2026
- Amazon Web ServicesData retention in Amazon Bedrock10 October 2026
- Amazon Web ServicesHIPAA eligible services reference3 September 2026
- MetaModel API terms of service9 October 2026
- IRSPublication 15 (2026), Employer's Tax Guide2026
- IRSTopic no. 759, Form 940, employer's annual federal unemployment (FUTA) tax return24 September 2026
- US Department of LaborUnemployment insurance tax topic11 September 2026
- US Bureau of Labor StatisticsEmployer Costs for Employee Compensation, June 2026 (table 1)9 September 2026
- US Department of LaborSeverance pay2026
- US Government Publishing Office29 U.S.C. chapter 23, Worker Adjustment and Retraining Notification2024
- Canada Revenue AgencyCPP contribution rates, maximums and exemptions31 October 2025
- Canada Revenue AgencySecond additional CPP contribution rates and maximums31 October 2025
- Canada Revenue AgencyEI premium rates and maximums16 September 2025
- Government of OntarioEmployer Health Tax (EHT)18 September 2026
- Government of OntarioYour guide to the Employment Standards Act: severance pay18 March 2026
- Government of OntarioYour guide to the Employment Standards Act: termination of employment23 July 2026
- HMRCRates and thresholds for employers 2026 to 20271 September 2026
- GOV.UKWorkplace pensions: what you, your employer and the government pay2026
- The Pensions RegulatorEarnings thresholds2026
- HMRCPay Apprenticeship Levy6 April 2026
- GOV.UKRedundancy: your rights, redundancy pay2026
- GOV.UKMaking staff redundant: redundancy consultations2026
- Bundesministerium für Arbeit und SozialesDas ändert sich im neuen Jahr22 December 2025
- Bundesministerium für Arbeit und SozialesSozialversicherungsrechengrößen 20268 October 2025
- Bundesministerium der Justiz§ 241 SGB V, allgemeiner Beitragssatz2026
- Bundesministerium für GesundheitBeiträge3 September 2026
- Bundesministerium für GesundheitFinanzierung der Pflegeversicherung14 September 2026
- Deutsche Gesetzliche UnfallversicherungZahlen 202522 September 2026
- Bundesministerium der Justiz§ 1a KSchG, Abfindungsanspruch bei betriebsbedingter Kündigung2026
- Bundesministerium der Justiz§ 622 BGB, Kündigungsfristen2026
- Bundesministerium der Justiz§ 17 KSchG, Anzeigepflicht2026
- Gazette of IndiaS.O. 5109(E): wage ceiling for Chapter III of the Code on Social Security17 September 2026
- Gazette of IndiaG.S.R. 525(E): Employees' Provident Fund Scheme, 202629 June 2026
- EPFOContribution rates12 June 2018
- ESICCoverage4 April 2025
- Gazette of IndiaThe Industrial Relations Code, 2020: retrenchment (sections 70 and 83)29 September 2020
- CPF BoardHow much CPF contributions to pay7 October 2026
- CPF BoardCPF-related changes taking place in 20267 October 2026
- CPF BoardSkills Development Levy2 July 2026
- CPF BoardCan I make CPF contributions for foreign employees?20 April 2026
- Ministry of Manpower, SingaporeS Pass levy and quota requirements19 February 2026
- Ministry of Manpower, SingaporeEmployment Pass: key facts4 May 2026
- Ministry of Manpower, SingaporeResponsible retrenchment14 July 2026
- Ministry of Manpower, SingaporeMandatory retrenchment notifications14 July 2026
- Social Security System, PhilippinesPay contribution2025
- PhilHealthAdvisory 2025-0002: premium contribution rate2025
- Department of Budget and Management, PhilippinesCircular Letter No. 2024-2: Pag-IBIG Fund contributions1 February 2024
- Supreme Court of the Philippines E-LibraryPresidential Decree No. 851: 13th-month pay1975
- Department of Labor and Employment, PhilippinesLabor Code of the Philippines, 2022 DOLE edition (Article 298)2 February 2022
- KWSP (EPF Malaysia)Mandatory contribution2026
- PERKESO (SOCSO)Rate of contribution1 October 2024
- PERKESO (SOCSO)Employment Insurance System contribution rates (Act 800)15 November 2024
- Department of Labour, Peninsular MalaysiaHow is the termination benefit payment calculated?8 August 2025
- Attorney General's Chambers, MalaysiaP.U.(A) 262/2022, Employment (Amendment of First Schedule) Order 202212 August 2022
- UAE Government portal (u.ae)End of service benefits for employees in the private sector12 August 2026
- UAE Government portal (u.ae)Taxation2026
- UAE Government portal (u.ae)Labour rights in the private sector8 October 2026
- Dubai Health Authority (ISAHD)Health insurance FAQ25 March 2026
- Department of Health, Abu DhabiHealthcare Insurers Manual2017
- UAE Government portal (u.ae)Federal Decree-Law No. 33 of 2021 regulating labour relations (English text): Articles 43, 51 and 532022
- Australian Taxation OfficeSuper guarantee17 April 2026
- Revenue NSWPayroll tax thresholds and rates1 July 2026
- State Revenue Office VictoriaPayroll tax current rates29 September 2026
- WorkSafe VictoriaIndustry rates and key dates29 June 2026
- Federal Register of Legislation, AustraliaFair Work Act 2009, sections 117 to 123 (notice and redundancy pay)7 July 2026
Frequently asked questions
What are typical AI running costs?
Three parts: model calls, hosting, and the time someone spends keeping it running. Model calls depend on the model and the volume: 2,000 requests a month of 3,000 tokens in and 500 out cost $0.35 on Amazon Nova Micro and $110 on GPT-6 Astra at list prices checked 10 October 2026. In the tool's US example, four hours of upkeep ($100) cost more than the model calls on Claude Sonnet 5.5 ($22.00); the running-cost estimate adds all three.
Should I use burdened rates or base pay?
Always use the fully burdened rate (including taxes and benefits) to get an accurate picture of what the manual process actually costs your company. The part the law sets differs by country: at the tool's example pay it adds 7.7% in the US, 15.4% in the UK and 21.3% in Germany, before benefits. The button under the rate field adds it for the country you pick.
What is a good payback period for automation?
One well inside the time the process and its systems will stay as they are. A process that changes every year needs a payback of months; one that has run the same way for a decade can carry a longer one. Compare the figure with the hurdle your finance team sets for internal projects, and with the payback the tool shows when the automation removes less of the work.
How do I estimate the build cost?
Get a fixed quote for the specific process: the cost depends on how many systems it touches, how messy the inputs are and how much checking it needs. For reference, Stafflo's pilot, one tool built around one process and deployed in four to six weeks, is priced from $15,000 in the US, from ₹4 lakh in India, from £10,000 in the UK and from €12,000 in Europe.
Is the time saved real money?
Only when the hours turn into cash, or into work that earns it. If roles go, the saving is cash, less the exit cost the law sets: for one role with three years' service at the tool's example pay, nothing in the US, £2,160 in the UK and ₹60,000 in India. If you avoid a hire, the saving is the hire's cost. If the hours go to other work, the return depends on that work.
Which AI model should the running cost assume?
The cheapest one that does the task well enough on your own examples. For the same workload, current list prices run from $0.35 to $110 a month, so test two or three models on real documents before choosing. The estimate in the tool compares every current model at your volume.
Related
See exactly where your process bleeds money.
We map out your manual workflows to show you the true cost and the automation ROI.
Buy an auditBuilt and maintained by Stafflo. Questions or a wrong number: contact. Last updated 11 October 2026.