What AI Automation Costs, and How to Work Out Payback

At Bitvyas, an AI automation project has four costs: a one-off build, a monthly run cost, the time a person spends reviewing, and upkeep when things change. The model fee is usually the smallest of the four. Payback is the build cost divided by the monthly saving after run and review costs.

That formula is the whole method. The rest of this article shows how to fill it in for your own process, with three worked examples in rupees, pounds and dollars. The examples use assumptions that are stated in each table, not client results. We do not publish client numbers without permission, and this post does not quote a Bitvyas price list. Replace every assumption with your own before you decide anything.

What are the key facts for costing AI automation?

These are the reference points the worked examples use. All are as of 9 October 2026.

Gartner's stated reasons for cancellation were "escalating costs, unclear business value or inadequate risk controls". Two of those three are costing problems, which is why it is worth doing the sums before the build.

What are the four costs of an AI automation?

Build, run, review and upkeep. Each behaves differently, and most budgets miss at least one.

1. Build (one-off). Mapping the process, connecting it to your tools (Tally, Zoho Books, Xero, QuickBooks, a CRM, a shared inbox), writing the rules and the checks, and running it alongside the current way for two to four weeks. This is the largest number and the one a vendor quotes.
2. Run (monthly). The model fee, hosting, and the monitoring that tells you when something stops working. For document and message work the model fee is small, as the examples below show.
3. Review (monthly, in hours). The time a person spends on what the workflow flags. This is the cost that is easiest to forget and the one that makes the saving real. If review takes as long as the old way, there is no saving.
4. Upkeep (irregular). Rules change, formats change and models are retired. Anthropic's pricing page, for example, marks Claude Sonnet 4.5 as deprecated, so a workflow built on it needs a planned move. A supplier changing its invoice layout or an accounting package changing its import format does the same.

How do you work out payback?

Payback in months equals the build cost divided by the net monthly saving. The net monthly saving is the value of the hours freed, minus the model fee, minus other run costs.

Write it out:

• Hours freed = items per month × minutes per item by hand ÷ 60, minus the hours a person now spends reviewing.
• Value of hours freed = hours freed × your loaded cost per hour (salary plus employer costs, not salary alone).
• Net monthly saving = value of hours freed − model fee − other run costs.
• Payback (months) = build cost ÷ net monthly saving.

This counts time only. It leaves out cash that arrives sooner, errors that no longer happen and work that was simply not being done. Add those if you can measure them, but do not let them carry a case that fails on time alone.

What does the model fee actually come to?

Very little for document and message work. A purchase invoice read at about 4,000 input tokens and 600 output tokens on Claude Sonnet 5.5 costs about $0.014, or roughly ₹1.35, at the prices above. That is 4,000 × $2 ÷ 1,000,000 plus 600 × $10 ÷ 1,000,000. Token counts vary with document length and how many checks the workflow runs, so treat 4,000 and 600 as an assumption and measure your own in the first two weeks.

The model fee grows with volume and with how much text each item carries. It is still the smallest line in all three examples below. The costs that decide payback are build, review and volume.

Example 1: purchase invoices into Tally or Zoho Books (India, INR)

This is the process in our post on purchase invoices into Tally and Zoho Books with AI. The numbers are illustrative.

Assumptions: 4 minutes of manual entry per invoice; the workflow flags 15% of invoices and a person spends 3 minutes on each flag; loaded cost ₹300 per hour; 4,000 input and 600 output tokens per invoice on Claude Sonnet 5.5; other run costs ₹10,000 a month (hosting, monitoring, rule upkeep); build cost ₹3,00,000.

The same build pays back in 4 months at 5,000 invoices and in 31 months at 1,200. Volume is the lever. Our view: at around 1,200 invoices a month, use the automatic capture already inside your accounting software first. Zoho Books, for example, has Autoscan, which pre-fills a bill from an uploaded document. Build a custom workflow when volume, supplier variety or the number of systems makes the built-in feature stop being enough.

Example 2: chasing overdue invoices (UK, GBP)

This is the process in Collections with AI: what it chases, what your team still handles.

Assumptions: 300 overdue invoices a month; each takes three chasing touches of 6 minutes by hand; the workflow drafts and sends the reminders and a person handles the 25% of items that need judgement (disputes, promises to pay, payment plans) at 10 minutes each; loaded cost £25 per hour; 6,000 input and 800 output tokens per item across the touches; other run costs £150 a month.

Here the build cost is the open question, so the table shows three. The cash side is not counted. If the workflow brings payment dates forward, that is a further benefit on top of the time saved.

Example 3: insurance renewals (US, USD)

This is the renewal process in what to automate with AI first.

Assumptions: 120 renewals a month; 45 minutes of preparation by hand (gathering updated details, building the comparison pack); with the workflow a person spends 15 minutes reviewing each pack before it goes to the client; loaded cost $35 per hour; 30,000 input and 3,000 output tokens per renewal; other run costs $200 a month.

The advice on cover, any change in risk and the renewal conversation stay with the agent. The workflow only prepares.

What makes the cost go up?

• More systems. Each extra system to read from or write to adds build and upkeep.
• Messy inputs. Handwriting, photos and rare layouts push more items into review.
• Higher cost of error. Where a mistake reaches a customer or a regulator, review stays at 100% and the saving shrinks. This is the third question in our four-question test.
• No written rules. If two people do the work two ways, someone has to agree one way before anything is built.
• Low volume. A monthly task with a handful of items rarely repays a build.

Should you build, buy a tool or use what your software already has?

Use the built-in feature if it does the job at your volume, buy a specialist tool if one exists for your exact process, and build when neither fits your systems, rules or mix of inputs. The costing is the same in all three cases: put the licence or build cost in the denominator, and count the review time honestly.

Ask any vendor, including us, five questions before you sign. What is the build cost and what does it include? What does it cost to run each month, excluding the model fee? What share of items will a person still review in the first month? What happens when a model, format or rule changes, and who pays? How do I switch it off and keep my data?

How do you check the payback after go-live?

Take a two-week baseline before you start: items per month, minutes per item and who does the work. Track the same measures monthly after launch, plus the share of items a person edits. If review hours stay high after two months, fix the rules or the inputs before adding volume. Reducing approvals comes later, and only where the error rate allows it.

Bitvyas builds this kind of automation inside the tools a business already runs. Each process above has a page on our use cases section, including collections, reconciliation and insurance renewals, and our services page explains how we scope a first workflow.

Sources

• Anthropic, Claude API pricing (undated page; figures read 9 Oct 2026): https://platform.claude.com/docs/en/about-claude/pricing
• Pound Sterling Live, US dollar to Indian rupee history 2026: https://www.poundsterlinglive.com/history/USD-INR-2026
• Pound Sterling Live, British pound to Indian rupee history 2026: https://www.poundsterlinglive.com/history/GBP-INR-2026
• Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", 25 June 2025: https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
• Zoho Books, Documents (Autoscan): https://www.zoho.com/in/books/help/documents/documents.html

Related

• What to automate with AI first: 8 business processes mapped
• Purchase invoices into Tally and Zoho Books with AI
• Collections with AI: what it chases, what your team still handles
• Use cases
• Services

Prices and exchange rates as of 9 October 2026. The worked examples are illustrations built on stated assumptions, not client results or a Bitvyas quote.