--- title: "AI Implementation Cost: What Companies Pay For | Murarum" description: "What AI implementation costs a company of 10 to 250 people: six cost components, model usage from real volumes, hidden costs and how to calculate ROI." url: https://murarum.ro/en/insights/ai-implementation-cost language: en alternate_ro: https://murarum.ro/articole/cost-implementare-ai --- 7 min read · AI integration # How much does AI implementation cost for a company? The cost of implementing AI in a company is the sum of six components: diagnosis, the integration build, model usage, the surrounding system, your people's time and ongoing evaluation and maintenance. For companies of 10 to 250 people, model usage is often the smallest line. The build, the people time and the upkeep decide the total, and every one of them is driven by process volume and data readiness. By Mircea Muraru · 29 September 2026 · Updated 7 October 2026 ## Why there is no single price Ask ten vendors what AI implementation costs and you will get ten numbers, because each one is pricing a different thing. One quotes a chatbot subscription, another a prototype, a third a production system connected to your ERP. None of them is wrong. They are answering different questions. A better question is: what will this specific workflow cost to build, run and keep correct for a year, and what will it save? That question has an answer, and it can be estimated before any code is written. The rest of this article is how we break it down. We do not publish our own prices; this is the method, not a quote. ## The six cost components | Component | What you pay for | What drives it up | What drives it down | | -------------------------- | ----------------------------------------------------------------------------------- | --------------------------------------------------------------------- | ------------------------------------------------------------------------- | | Diagnosis | Mapping the process, checking the data, ranking use cases, a written go or no-go | Several candidate processes, scattered data, no process owner | One named process, exportable data, an owner in the room | | Integration build | Workflow design, connections to your systems, human approval points, guardrails | Many systems to connect, many exceptions, no API on the core tool | One or two systems with APIs, a narrow scope, a defined correct output | | Model usage | Tokens or requests billed by the model provider every month | Large models, long inputs, agent loops with many calls | The smallest model that passes evaluation, short inputs, batch processing | | Surrounding system | Hosting, logging, monitoring, alerts, storage, access control | Strict data residency, high availability, long log retention | Existing cloud accounts, modest volume, clear retention rules | | People time | Your team's hours in design, testing, review and training | Unclear ownership, many stakeholders, output reviewed by hand forever | One owner, a small pilot group, review only where it matters | | Evaluation and maintenance | Re-running the evaluation set, fixing errors, adapting to model and process changes | Frequent process changes, model retirements, no evaluation set | A stable process, a kept evaluation set, a named internal supervisor | ### 1\. Diagnosis Diagnosis is the cheapest component and the one that most often saves money, because it is where weak use cases are dropped. It answers three things in writing: which process, whether the data can be reached and what a correct result looks like. Our [AI readiness checklist](https://murarum.ro/en/insights/ai-readiness-checklist-smes) covers the same ground. We run it as [The Diagnosis](https://murarum.ro/en/approach/diagnosis), a scoped step whose written output the client keeps whether or not a build follows. ### 2\. Integration build This is usually the largest one-time cost, and the model is a small part of it. Most of the work is connecting the model to the systems where the data lives, handling the exceptions a person handles today without thinking, and building the points where someone approves or corrects an output. The number of systems and the number of exceptions predict the effort better than the choice of model does. ### 3\. Model usage Providers bill per token (a token is a fragment of text) or per request, separately for input and output. It is a running cost that scales with volume, so it can be estimated from your real numbers. The next section shows how. ### 4\. The surrounding system A model call in production needs a place to run, logs of every input and output, monitoring for accuracy, latency and cost, alerts when a budget limit is reached, and access control. These are ordinary cloud costs, often small, but they exist every month and they are the part most first estimates leave out. ### 5\. Your people's time The people who do the task today have to explain the exceptions, check outputs during testing and learn the new workflow. Someone has to own it. This time rarely shows up on an invoice, which is why it is the most underestimated component. It is also the one that decides whether the system is used after launch. ### 6\. Evaluation and maintenance An AI workflow is not finished at launch. Processes change, providers retire models and release new ones, and every change needs the evaluation set run again before it goes live. Budget for this as a recurring cost from the start, not as a surprise in month six. ## How to estimate model usage from real volumes Four numbers are enough for a first estimate: how many times the task runs per month, how many input tokens each run sends (instructions plus the document or message), how many output tokens it returns, and the provider's price per million tokens. As a rough guide, Anthropic's pricing page puts one token at about 0.75 words of English text and notes that the count varies by language, so measure on your own documents rather than guessing. Illustration only: the volumes below are hypothetical, chosen to show the arithmetic. The prices are Anthropic's published API list prices, checked on 29 September 2026: Claude Haiku 4.5 at $1 per million input tokens and $5 per million output tokens, Claude Sonnet 5.5 at $2 and $10\. Say a company processes 2,000 supplier invoices a month, and each extraction sends 2,500 input tokens and returns 400 output tokens. That is 5 million input tokens and 0.8 million output tokens a month. | Setup (illustrative) | Monthly input | Monthly output | Model usage per month | | -------------------------------------------------------------------------- | ----------------- | ------------------- | --------------------- | | One call per invoice, Claude Haiku 4.5 | 5M tokens x $1 | 0.8M tokens x $5 | $9 | | One call per invoice, Claude Sonnet 5.5 | 5M tokens x $2 | 0.8M tokens x $10 | $18 | | One call per invoice, Haiku 4.5 via Batch API (50% off) | 5M tokens x $0.50 | 0.8M tokens x $2.50 | $4.50 | | Agent with 8 calls per invoice and 6,000 input tokens per call, Sonnet 5.5 | 96M tokens x $2 | 6.4M tokens x $10 | $256 | Two things stand out. First, for a well-scoped extraction task at this volume, model usage is small next to the build and the people time. Second, design choices move it far more than the price list does: the agent version of the same task costs about 28 times the simplest one. That is why we choose the smallest model that passes the evaluation set, keep inputs short and set a monthly budget limit with alerts before launch. Prices change, so the estimate is repeated whenever the provider updates its list. For invoices and other documents, dedicated extraction services are an alternative priced per page; see [how much automatic invoice reading costs](https://murarum.ro/en/insights/invoice-reading-cost). ## Hidden costs - Data cleanup. If the inputs are scanned PDFs, inconsistent spreadsheets or records spread across tools, getting them into a usable form can cost more than the AI work itself. - Exceptions. The cases a person resolves by instinct each become a rule, a route to human review or a known limitation. Each one costs design time. - Human review. Outputs that touch money, customers or legal documents need a person to approve them. That review time is a running cost, and it only shrinks as evaluation results justify it. - Data residency. Keeping processing in a given region can cost more. On Amazon Bedrock and Google Cloud, for example, Anthropic lists a 10% premium for regional endpoints over global ones. - Compliance work. A data flow document for your DPO, GDPR decisions on what the model may see, and retention rules. Cheap when done in design, expensive when done after a legal review stops the launch. - Model changes. Providers retire older models. Moving to a new one means re-running the evaluation set and sometimes adjusting prompts. - Lock-in. A workflow built so that only the vendor can change it turns every future adjustment into a new invoice. Ask for documentation and hand-off as part of the scope. The last point is the one we care about most. On every build we ship, the system is documented so the client's own team can run it without us, as on the [WHO mobile application](https://murarum.ro/en/work/who-mobile-application). For AI work, hand-off also includes the evaluation set, the budget limits and the runbook for when the model is wrong. ## How do you calculate the ROI of an AI project? The ROI of an AI project is the net benefit over a period divided by the total cost over the same period: what the workflow saves, minus what it costs to build and run, divided by what it costs. The calculation needs the same inputs as the cost estimate above, plus a measured baseline of how the process runs today. - Hours saved, valued at the loaded cost per hour (salary plus taxes and overheads), net of the time people still spend reviewing outputs. - Errors and rework avoided: corrections, duplicate payments, credit notes, the time spent finding and fixing mistakes. - Faster cycle time, counted only where it has a money value, such as early-payment discounts or orders that no longer wait. - Minus the costs: the one-time build (including diagnosis and your team's time), model usage, the surrounding system, maintenance and the supervision of the workflow. Illustration only: the figures below are hypothetical round numbers, not a client result. Take the 2,000 invoices a month from the example above. Say each one takes 3 minutes by hand, which is 100 hours a month, and after automation people still spend 20 hours reviewing and approving. That leaves 80 hours saved, at a loaded cost of €20 an hour. | Line (illustrative) | Per month | First 12 months | | --------------------------------------------------- | --------- | --------------- | | Hours saved: 80 h x €20 | €1,600 | €19,200 | | Errors and rework avoided | €200 | €2,400 | | Faster cycle time (not counted here) | €0 | €0 | | Model usage and hosting | \-€50 | \-€600 | | Maintenance and supervision | \-€350 | \-€4,200 | | Build, one-time (including diagnosis and team time) | | \-€14,000 | | Net benefit | €1,400 | €2,800 | In this illustration the payback period, the build cost divided by the monthly net benefit, is €14,000 / €1,400 = 10 months. The first-year ROI is €2,800 net benefit divided by €18,800 total cost, about 15%. In the second year there is no build cost, so the same workflow returns €16,800 on €4,800 of running costs. Change one input, such as 40 hours saved instead of 80, and the payback stretches to almost two years, which is why the inputs have to be measured rather than assumed. What to measure before you start: monthly volume, minutes per item (time a sample of real cases), the error and rework rate, the cycle time from arrival to completion, and the loaded cost per hour of the people doing the work. Agree the target number with the process owner before the build starts, and compare against the baseline after launch. ## When AI is not worth the cost The simple test is the ROI calculation above with conservative inputs. If the numbers only work under optimistic assumptions, the answer is no, or not yet. In our experience it is usually one of these: - The task runs a few times a month. Volume is what pays back the build. - A rule or a formula would do. If the logic fits in a spreadsheet or a plain integration, that is cheaper and fully predictable. See [process optimization and rule-based automation](https://murarum.ro/en/services/process-optimization). - The process is not defined. AI makes a broken process fail faster and at a higher cost. - Nobody owns it. A workflow without an owner is abandoned, and its full cost becomes a loss. - An error is expensive and hard to detect. If every output must be checked in full by a person, the savings may never appear. None of this makes AI a poor investment. It makes it a specific one. Romanian companies are at an early stage of adoption, as our [2026 statistics](https://murarum.ro/en/insights/ai-adoption-romania-2026) show, which means most first projects are still being chosen. Choosing the workflow with the clearest payback matters more than choosing the model. When a workflow passes that test, our [AI integration work for companies of 10 to 250 people](https://murarum.ro/en/services/ai-integration) scopes it against the six cost components above before anything is built. ## Summary - AI implementation cost has six components: diagnosis, integration build, model usage, the surrounding system, people time, and evaluation and maintenance. - The build and the people time usually decide the total; model usage is often the smallest line for a well-scoped task. - Estimate model usage from four numbers: runs per month, input tokens, output tokens, price per million tokens. - Design choices such as agent loops and model size move the running cost more than provider prices do. - ROI is net benefit divided by total cost; payback is the build cost divided by the monthly net benefit. Both depend on a baseline measured before the build. - If the payback only works on optimistic assumptions, fix the process first and come back to AI later. ## Frequently asked questions ### How much does AI implementation cost for a small or mid-sized company? It depends on six components: diagnosis, the integration build, model usage, the surrounding system (hosting, logging, monitoring), your team's time, and ongoing evaluation and maintenance. The build and people time usually dominate. Any honest figure comes after a diagnosis of the specific process, its volume and its data, which is why we scope projects after diagnosis and do not publish prices. ### How much does it cost to run an AI model every month? Multiply the monthly number of runs by the input and output tokens per run, then by the provider's price per million tokens. As an illustration, 2,000 invoice extractions a month at 2,500 input and 400 output tokens each come to $9 a month on Claude Haiku 4.5 at Anthropic's list prices checked in September 2026\. Agent designs with many calls per task cost much more. ### What are the hidden costs of implementing AI? Data cleanup, handling exceptions, human review of sensitive outputs, data residency premiums, GDPR and compliance work, re-evaluation when providers retire models, and vendor lock-in if the system is not documented for hand-off. Most of them are cheap to plan for in design and expensive to discover after launch. ### How do I calculate the ROI of an AI project? ROI is the net benefit divided by the total cost over the same period. The benefit is hours saved times the loaded cost per hour, plus errors and rework avoided; the cost is the build plus model usage, hosting, maintenance and supervision. Payback is the build cost divided by the monthly net benefit. Measure the baseline (volume, time per item, error rate) before the build starts. ### When is AI not worth the cost? When the task runs rarely, when a simple rule or integration would do the job, when the process is not defined, when nobody owns the workflow, or when every output must be fully checked by a person. In those cases process optimization or a plain systems project pays back faster. [Still have questions? Write to us](https://murarum.ro/en/contact) ## Related - [AI Integration](https://murarum.ro/en/services/ai-integration) - [All case studies](https://murarum.ro/en/work) [WHO LENA: guidance for hard conversations with patients](https://murarum.ro/en/work/who-mobile-application): World Health Organization · Global health · 2021 [A new teamsunday.com, in five languages](https://murarum.ro/en/work/sunday-website): Sunday · Branded merchandise platform · 2026 ## Want a cost estimate for one workflow? Tell us the process, the monthly volume and where the data lives. We will tell you which cost components matter and whether AI belongs in it at all. [Let’s talk](https://murarum.ro/en/contact) [office@murarum.ro](mailto:office@murarum.ro) ## Contact - Email: office@murarum.ro - Phone: +40 761 339 149 - Let’s talk: https://murarum.ro/en/contact