Key takeaways
- Budget five cost areas, not one: discovery or paid audit, implementation, infrastructure and model usage, integration licenses, and support and change.
- Model usage scales with volume and with how much text each item carries, and is often a smaller line than the software licenses around it.
- The costs that surprise buyers are data cleanup, exception handling, adoption effort and monitoring, none of which appear on a build quote.
- Hours saved become money only through avoided hiring, reduced overtime or outside spend, or capacity redeployed to measurable work.
- A case that still clears your payback threshold with usage costs doubled and exceptions higher than expected is the one finance will approve.
Ask three suppliers what an AI workflow costs and you will get three numbers that cannot be compared, because each covers a different scope and none of them covers what the workflow costs in year two. The useful question is not the price. It is which cost areas exist, which of them scale with your volume, and which savings survive a finance review.
What are the cost areas in an AI workflow automation project?
Five areas cover almost every engagement. Missing one is how a project comes in on budget and over cost.
| Cost area | One time or recurring | What drives it |
|---|---|---|
| Discovery, or a separately scoped paid audit | One time | Number of teams and workflows reviewed, how well the process is already documented, how many systems are in scope |
| Implementation | One time | Integrations to build, document or request types to handle, exception paths, depth of testing, approval logic |
| Infrastructure and model usage | Recurring | Cloud compute and storage, logging, separate test environments, tokens consumed per item |
| Integration and platform licenses | Recurring | Automation platform seats or task allowances, connector licenses, document processing services, extra API tiers |
| Support and change | Recurring | Monitoring, fixes when connected systems change, retesting when models or document layouts change, exception queue ownership |
Two of these are frequently absent from a quote. The first is your own team's time: process owners in workshops, subject experts labeling test cases, IT running security reviews. The second is the change work after launch, when people run old and new processes in parallel and throughput temporarily drops.
How does model usage scale with volume?
Model usage is billed per token, a unit of text of roughly four characters of English, counted separately for what you send and what the model produces. That makes cost a function of three things: how many items you process, how much text each item carries, and which model you choose.
The spread between models is wide enough to matter. Anthropic publishes Claude Haiku 4.5 at $1 per million input tokens and $5 per million output tokens, Claude Sonnet 5 at $2 and $10, and Claude Opus 5 at $5 and $25 (Anthropic). Two mechanisms reduce the bill further: the Batch API applies a 50 percent discount on input and output tokens for work that does not need an immediate answer, and prompt caching charges a cache hit at 0.1 times the standard input price for most models, which suits workflows that resend the same policy document or instructions on every item.
Put against real volume, the numbers are usually smaller than buyers expect. Anthropic's own worked example puts 10,000 support conversations averaging around 3,700 tokens each at roughly $37.00 in total on Claude Haiku 4.5. Long documents change that arithmetic quickly, because a hundred page contract carries far more input tokens than a support message, and server side tools carry their own charges, for example web search at $10 per 1,000 searches.
Three practical rules follow. Choose the model per task rather than per company, since classification and field extraction often run well on a smaller model. Estimate at your real document length rather than a sample. Then double the estimate in your business case, because volume and document length both tend to grow.
What do the licenses around the model cost?
The software surrounding a workflow is frequently the larger recurring line. Published list prices give a sense of scale.
- Zapier charges by task, counted when Zapier successfully completes a unit of work. Published Professional pricing runs from $19.99 a month for 750 tasks to $129.00 a month for 10,000 tasks on annual billing, with Team plans from $69.00 a month at 2,000 tasks (Zapier).
- Microsoft Power Automate lists Power Automate Premium at $15.00 per user per month paid yearly, Power Automate Process at $150.00 per bot per month and Power Automate Hosted Process at $215.00 per bot per month (Microsoft).
- Document processing services are priced per page. Amazon Textract publishes $0.0015 per page for text detection, $0.015 per page for tables or queries, $0.05 per page for form extraction and $0.01 per page for expense analysis, at the first million pages per month (AWS).
Per task and per seat pricing behaves differently as you grow. Task based platforms are inexpensive at low volume and climb steadily with it, which is one of the clearest signals that a workflow has outgrown a no code tool. Watch for license tiers too, since the API or connector a workflow depends on is sometimes only available on a plan above the one you currently pay for.
Which costs get missed?
Four lines cause most of the overruns, and none of them appear on a build quote.
- Data cleanup. Duplicate supplier records, inconsistent product codes and missing customer IDs cause wrong matches. You do not need a company wide data program, but you do need the specific fields this workflow depends on to be trustworthy, and someone has to do that work.
- Exception handling. Every workflow routes some share of items to a person. If nobody owns that queue, items pile up and staff quietly return to the old process. Budget the time as an ongoing operating cost, not a launch task.
- Adoption. Training, documentation, running both processes in parallel and slower throughput while people learn. Real, temporary and almost always underestimated.
- Monitoring and retesting. Suppliers change invoice layouts, a CRM field gets renamed, a model version is retired. Without monitoring, quality declines quietly and the first person to notice is a customer.
Where do the savings actually come from?
Three savings are defensible, and one common claim is not.
Hours returned, valued at loaded cost. Loaded cost means wages plus benefits, payroll taxes and overhead. The US Bureau of Labor Statistics reports private industry employer compensation costs averaging $46.89 per hour worked in June 2026, of which wages and salaries were $32.82, or 70.0 percent, with benefits making up the remaining 30.0 percent (BLS). Using an hourly wage alone understates the value of returned time by roughly a third.
Fewer errors and less rework. Price each error by what it costs to fix: staff time plus direct costs such as credit notes, redeliveries, late payment fees or missed early payment discounts. Count only the error types the workflow prevents, and subtract any new ones it introduces.
Faster cycle times, where speed changes money. Invoicing two days sooner improves cash timing. Capturing early payment discounts is cash. Answering quote requests faster may raise win rates, but only count it once you can measure the effect.
The claim that does not survive scrutiny is hours saved treated as cash by default. Returned hours become money in three ways only: a hire you planned and no longer make, reduced overtime or outside spend, or capacity redeployed to work with measurable output. Name which applies and agree it with the manager of the team before the case is written.
Illustrative example: Consider a regional distributor automating emailed purchase order intake. Roughly 1,800 orders a month take nine minutes each to key in. After automation, about three quarters go through with a two minute check and the rest take eight minutes as exceptions, giving an average of about 3.5 minutes. That returns around 165 hours a month. At a loaded cost of $45 an hour, that is about $7,425, plus perhaps $1,600 of avoided rework. Against it sits a recurring bill made up of model usage, document processing per page, the automation platform license, hosting and a support agreement, along with the order desk supervisor's time on the exception queue. Every figure here is a placeholder chosen to show the shape of the calculation. None of them is a benchmark, and none is a quote.
How do you build a case finance will accept?
A defensible case is short and testable. Six things belong in it.
- A measured baseline. Two to four weeks of volume, handling time, cycle time and error rate, with the data source and the dates named.
- Every cost line, split by type. One time and recurring, including your own team's time and the exception queue.
- A three year view. Build plus three years of running cost, not a project price.
- A named route to cash. Which hire, which overtime line, which contract.
- Three scenarios. Expected, conservative and downside. The downside should assume exceptions run higher than planned and usage costs double.
- Stop conditions. The results that would pause or end the project, agreed before it starts.
To run the arithmetic on your own figures, use the AI automation ROI calculator, and how to calculate the ROI of AI automation explains the method behind it in more detail. If you have not settled on a workflow yet, start with which business processes to automate first.
How Kastling approaches cost and scope
We do not publish a price list, because a number quoted before anyone has seen your process is a guess. What we can be clear about is the structure. A discovery call is free and carries no commitment. For AI and operations work a separately scoped paid audit is common, though not required for every project, and a well defined brief can move straight to a proposal. Implementation is scoped and quoted against agreed deliverables. Infrastructure and usage are billed according to where the solution runs, in your own AWS or Azure account or in ours. Maintenance is a separate agreement that states what is included.
Keeping those four areas distinct is deliberate. It lets you compare suppliers on scope rather than headline numbers, and it means nobody discovers a recurring cost after go live. Read more about our AI Integration & Automation service, or bring us a workflow and we will tell you what would drive its cost.
Questions
Is the recurring cost or the build cost the bigger risk?
The recurring cost, because it is the one nobody sizes properly. A build is quoted once and negotiated once. Support, model usage, licenses and the internal time spent handling exceptions repeat every month, and together they often exceed the build within two to three years. Ask for a three year total rather than a project price.
Why are quotes for the same workflow so different?
Usually because the scope is different even when the description sounds the same. Two systems or six, a supported API or a screen based workaround, ten document variants or eighty, shallow testing or a tested set of real historical cases. Ask each supplier to list the integrations, the document or request types covered, the exception paths and the testing approach, then compare those rather than the totals.
Should we pay for an audit before committing to a build?
It depends on how well you already understand the process. If you can describe the steps, the volumes, the systems and the exception paths, and you have permitted access to the data, a proposal can often be scoped directly. If the process crosses teams, or nobody can say how many exceptions occur, a paid audit usually costs less than discovering the same facts halfway through a build.
How do we budget model usage when prices keep changing?
Treat it as a variable, not a fixed line. Estimate at your real volume and document length, then run the case again with the figure doubled. Design the workflow so the model can be swapped without rebuilding the process, which protects you when a provider changes its line up or retires a version you depend on.
What ongoing support does an automated workflow actually need?
Monitoring that someone reads, fixes when a connected system changes, retesting when a model version is retired or a document layout changes, and a named person who owns the exception queue. Agree what is included, what is billed separately and how quickly issues are picked up, because hosting a workflow is not the same as maintaining it.