What does it cost to implement AI in a business? A 2026 guide

Short answer: There is no single price, but there is a structure. One AI workflow costs a one-off build, a monthly run cost and your own people’s time. Model fees are usually the smallest line. In our hypothetical worked example, one customer-service workflow costs about €18,000 to build and about €1,000 a month to run.

There is no honest single price for AI, but there is an honest structure. One AI workflow in a company of 20 to 500 people costs a one-off build (scoping, integration, testing, training), a monthly run cost (model usage, hosting, monitoring, maintenance) and the time of your own people.

Model fees are usually the smallest line: in Bitkom’s 2026 survey of German companies, only 8 percent of AI users called token consumption a large cost item, against 50 percent for data preparation and 41 percent for integration. In the hypothetical worked example below, one customer-service workflow comes to about €18,000 to build and about €1,000 a month to run.

Below: where the money goes, what models cost as of October 2026, and a worksheet to fill in before you talk to any vendor.

What are the three cost buckets of an AI project?

Bitkom asked 603 German companies with 20 or more employees where their AI costs arise. Among companies already using AI, 51 percent named infrastructure such as cloud and computing power as a large cost item, 50 percent data preparation and 41 percent integration into existing systems. Software subscriptions and licenses (21 percent) tied with external consulting (21 percent), and training staff was a large cost for 19 percent. The survey is self-reported and covers Germany only. It still corrects a common assumption: AI cost is mostly not model fees.

Bitkom president Ralf Wintergerst put it plainly: “Anyone who only buys a license has bought AI, but has not changed anything yet” (our translation).

1. Build and integration (one-off)

• Scoping and baseline: choosing the workflow, defining “done,” measuring today’s numbers.

• Data access and preparation: getting the documents and cleaning the knowledge the system must use.

• Integration: connecting the helpdesk, ERP, CRM or document system. Each system adds access rights, a test environment and error handling.

• Workflow build: instructions, rules, escalation paths, approval steps.

• Evaluation: a test set of real past cases with known correct answers, and a way to rerun it.

• Rollout: a parallel run, training and handover.

What moves this number most: how many systems you connect, the state of your data, how much output a person must check, and how regulated the work is.

2. Monthly running costs

• Model usage: priced per token (prices below).

• Hosting: where the workflow runs, plus storage and logs.

• Monitoring: error rate, review rate and cost per case, watched every week.

• Maintenance: new edge cases, updated instructions, model migrations.

• Seat licenses, if you also buy assistant tools. Anthropic’s Claude Team plan lists a standard seat at $20 a month billed annually or $25 billed monthly. Its Enterprise plan is $20 a seat plus usage at API rates.

3. People, change and training

Your own people are a cost too. Someone owns the workflow. Someone reviews outputs, and that review time must be subtracted from any time saved. The team needs training: Bitkom found that 91 percent of German companies using AI train their staff. Acceptance is work as well. For 42 percent of AI users in the same survey, lack of employee acceptance is an obstacle.

How much do AI models cost per month?

Model providers charge per million tokens, with separate prices for input (what you send) and output (what the model writes). As a rough guide, Anthropic’s pricing page says one token is about 4 characters or 0.75 words in English. Standard prices as of October 2026, from the providers’ own pages:

Sources: Anthropic, OpenAI (standard, short-context rates), Google (paid tier, page updated October 1, 2026). All three offer a batch mode at 50 percent off for work that can wait.

For most workflows, the monthly model bill is cases × tokens per case × price. At list prices, 10,000 support conversations of about 3,000 input and 700 output tokens each come to about $65 on Claude Haiku 4.5 ($1 input and $5 output per million tokens). Agents that run for long sessions use far more tokens than a single draft. Anthropic’s managed agents also add $0.08 per session-hour of runtime.

Compare cost per finished case, not price per token. Tokenizers differ: Anthropic notes that its newer models produce about 30 percent more tokens for the same text.

How much does an AI consultant cost?

The largest published rate survey we found for German-speaking markets is freelancermap’s Freelancer-Kompass 2026, with about 5,400 freelancers in the DACH region surveyed between November 17, 2025 and February 8, 2026. Its IT breakdown gives a median rate of €95 an hour across IT disciplines, €100 for data and analytics, €120 for SAP and ERP specialists and €90 for software and web development. That puts a median IT freelancer day at about €760.

Read this with care. The rates are self-reported by users of one platform, and they describe freelancers, not firms that carry project management, testing and warranty. We found no official statistic for AI consultant rates in the US or the UK. Rather than compare day rates, ask each vendor for a fixed scope: which workflow, which systems, what is tested, and what happens after go-live.

At the other end of the scale, Gartner estimated in 2024 that deployment approaches which use generative AI to change a business model cost $5 million to $20 million. Gartner’s Rita Sallam added that with generative AI, “costs aren’t as predictable as other technologies.” Scope, not the technology, sets the bill.

What do different types of AI projects cost?

We do not give averages here, because we found no credible survey that publishes them for companies of 20 to 500 people. What we can say is what you pay for in each type and what moves the price:

Gartner has also predicted that over 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. It notes that many use cases sold as agentic do not need an agent. Start with the cheapest design that does the job.

Worked example: what one AI workflow costs

This example is hypothetical. It shows how the numbers fit together. It is not drawn from any client.

A 60-person consumer health brand answers 3,000 customer emails a month. Each takes about 6 minutes of hands-on time: 300 hours a month. The new workflow reads each email, pulls the order status from the shop system and drafts a reply from approved answers. An agent reviews, edits and sends. Anything about health questions, large refunds or complaints goes to a person without a draft. We assume 70 percent of emails get a usable draft and that reviewing one takes 2.5 minutes.

Build cost

External days are priced at €760, the freelancer median above. Internal hours are priced at €45, the average cost of an hour worked in German industry and services in 2025, according to Destatis.

Monthly run cost

Dollars are converted at the ECB reference rate of October 2, 2026 (1 EUR = 1.1225 USD). The model fee is under a tenth of the run cost.

AI cost vs salary: what are the freed hours worth?

After go-live, 2,100 drafted emails take 2.5 minutes each (87.5 hours) and 900 still take 6 minutes (90 hours). Add 8 hours a month for the owner’s spot checks. Total: 185.5 hours, down from 300. About 115 hours a month are freed. At €45 an hour, that is about €5,150 a month. In the US, the Bureau of Labor Statistics puts private-industry employer compensation costs at $46.89 per hour worked in June 2026, which gives about $5,370.

Time freed is not money saved. Those 115 hours turn into money only if something changes:

• A hire you no longer need. If growth would have required another agent, the avoided cost is real. Then the build pays back in about four months: €18,000 ÷ (€5,150 − €1,000).

• Different work. The team spends the hours on returns, retention calls or faster replies. The value depends on that work, and you should measure it.

• Nothing changes. Same team, same hours, no new work. The cash saving is zero, and you have added about €1,000 a month in cost. Faster answers may still be worth it, but say so honestly.

Decide which case you are in before you build, and write it into the business case.

Which costs do companies forget?

• Data access. Contracts, permissions and cleanup take time. Bitkom found that 40 percent of German companies not using AI lack the data they need.

• Approvals. Review time is a running cost. Count it, as the example does.

• Evaluation. Every change to instructions, data or model needs a rerun of the test set.

• Model changes. Models retire. Anthropic’s deprecation page promises at least 60 days’ notice and set November 30, 2026 as the retirement date for Claude Sonnet 4.5, announced on September 30. OpenAI’s deprecation page promises at least six months for generally available models. Each migration means retesting.

• Price changes. Google’s Gemini 3.8 Flash doubles in price on January 1, 2027. OpenAI says promotional pricing for the older GPT-5.6 Sol runs at least through November 21, 2026. The table above lists GPT-6.1 Sol. If you build on any model with a promotional price, budget for the price after the promotion.

• Regional processing. OpenAI charges a 10 percent uplift for regional (data residency) endpoints on models released on or after March 5, 2026. For Claude on Amazon Bedrock and Google Cloud, regional endpoints carry a 10 percent premium over global ones.

• Integration with older systems. Gartner warns that connecting agents to legacy systems can be technically complex and require costly modifications.

What changes for companies in the EU and DACH?

AI use in Germany has passed a threshold: 57 percent of companies with 20 or more employees now use it, up from 36 percent a year earlier, according to Bitkom. Cost uncertainty is a barrier on both sides. Among companies without AI, 63 percent cite costs that are hard to calculate, and 54 percent of companies already using AI say the same. Even so, 45 percent of all companies are spending more on AI in 2026 than in 2025.

Across the EU, adoption is lower. Eurostat reports that 19.95 percent of enterprises with 10 or more employees used AI in 2025. Among those that had considered AI but did not use it, the most common reason was lack of relevant expertise (70.89 percent), followed by unclear legal consequences (52.52 percent).

Two EU items belong in the budget. Data protection review takes time: 66 percent of German AI users name data protection requirements as an obstacle. And under the AI Act, AI-literacy obligations have applied since February 2, 2025, according to the European Commission, and since July 2026 the duty is to take measures that support your staff’s AI literacy, so budget for some training. This is general information, not legal advice.

AI cost worksheet: fill in your own numbers

Copy this into a spreadsheet. The second column shows the hypothetical example.

If H is no, stop at F and write down what the hours will be used for. That is your business case.

Where does Postmodel’s pricing sit?

We build custom AI applications and workflow automation inside the tools a team already uses, with a person approving where it matters. Then we train the team and measure the result after go-live. An implementation project starts at €10,000, or from $12,000 in the US. Ongoing partnership is available on request. What a specific project costs depends on the drivers above: the systems involved, the state of the data and the review the work needs.

Where to start

Pick one workflow, measure its baseline and fill in the worksheet. If you need help choosing, read how to choose your first AI workflow. If you are weighing a product against a custom build, see build vs buy AI. Filled in the worksheet? Book a fit check and bring your numbers. We’ll tell you whether the workflow is worth a diagnostic.

Frequently asked questions

What are the cost buckets of an AI project?

There are three. Build and integration is one-off: scoping, data preparation, integration, the workflow build, evaluation and rollout. Running costs come every month: model usage, hosting, monitoring and maintenance. The third is your own people: an owner, reviewers and training. Review time has to be subtracted from any time saved.

Are model fees the biggest cost of AI?

Usually not. In Bitkom’s 2026 survey of German companies, only 8 percent of AI users called token consumption a large cost item, against 50 percent for data preparation and 41 percent for integration. In our worked example, the model fee is under a tenth of the monthly run cost.

Which AI costs do companies forget?

Data access, review time and a rerun of the test set after every change. Also model retirements, price changes after a promotion, uplifts for regional processing and integration with older systems. Every model migration means retesting, so put it in the budget.

Do the hours AI frees up save money?

Only if something changes. Freed hours become money when you avoid a hire or move the team to work with a result you can measure. If the team and its hours stay the same, the cash saving is zero and you have added a monthly cost. Decide which case you are in before you build.

What does a Postmodel project cost?

An implementation project starts at €10,000, or from $12,000 in the US. Ongoing partnership is available on request. What a specific project costs depends on the systems involved, the state of the data and the review the work needs.

Related reading

• Who should implement AI: partner, agency, consultant or your own team?

• How to measure AI ROI

• An AI implementation roadmap for the first 12 months

Sources

1. Bitkom, Erstmals nutzt die Mehrheit der Unternehmen KI, September 14, 2026. Telephone survey of 603 companies with 20 or more employees in Germany, weeks 28 to 33 of 2026.

2. Anthropic, Pricing, accessed October 4, 2026.

3. Anthropic, Claude plans and pricing, accessed October 4, 2026.

4. OpenAI, API pricing, accessed October 4, 2026.

5. Google, Gemini Developer API pricing, last updated October 1, 2026.

6. freelancermap, Freelancer-Kompass 2026, survey of about 5,400 freelancers, November 17, 2025 to February 8, 2026.

7. freelancermap, IT-Freelancer Stundensatz 2026, July 2, 2026.

8. Gartner, Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025, July 29, 2024.

9. Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, June 25, 2025.

10. Statistisches Bundesamt (Destatis), Eine Arbeitsstunde kostete im Jahr 2025 durchschnittlich 45,00 Euro, press release no. 148, April 29, 2026.

11. U.S. Bureau of Labor Statistics, Compensation costs for private industry workers averaged $46.89 per hour worked in June 2026, September 17, 2026.

12. European Central Bank, Euro foreign exchange reference rates: US dollar, rate of October 2, 2026.

13. Anthropic, Model deprecations, accessed October 4, 2026.

14. OpenAI, Deprecations, accessed October 4, 2026.

15. Eurostat, Use of artificial intelligence in enterprises, data extracted December 2025.

16. European Commission, AI Act, page updated August 3, 2026.