What ROI AI delivers in knowledge management, process automation and agents

AI pays off where it takes over or substantially shortens a whole transaction. In knowledge management the return comes from less searching and faster onboarding, in process automation from a lower cost per transaction, and with AI agents from transactions that run through without anyone touching them. How high the ROI turns out depends on the process you choose, the success rate and how you run the model: on premises, in a data center or through an API.

This article goes through each area in turn. For each one you get where the return comes from, what research shows, a worked example and the metrics to measure success. After that we compare the ways of running AI by cost and data protection, because the same agent can cost many times more depending on how it runs, or be ruled out for sensitive data altogether.

What ROI does AI deliver in a company?

Double-digit productivity gains on tasks that suit the technology are well documented. Controlled experiments show them in customer support, writing, consulting and software development:

Much less of this reaches the company level. According to McKinsey ("The State of AI", 2025), 88 percent of organizations use AI, but only 39 percent attribute any effect on EBIT to it. IBM found in 2025, surveying around 2,000 CEOs, that 25 percent of AI initiatives had delivered the expected ROI. At the same time, according to the Wharton School, 72 percent of companies measure their ROI formally, and three out of four report a positive return.

The gap between experiment and balance sheet has a simple cause. A chat assistant that everyone uses as they please saves minutes in many places. In a study of 25,000 workers in Denmark (Humlum and Vestergaard, 2025), self-reported time savings came to 2.8 percent of working hours, with no measurable effect on hours or wages. A process that a system handles end to end saves whole transactions, and those show up in cost accounting. That is why the MIT report "The GenAI Divide" (2025) found the biggest ROI not in marketing, where more than half of AI budgets go, but in back-office automation.

AI ROI comes from four sources:

• less working time per transaction, for invoices, orders or requests,
• more throughput with the same team, such as more quotes per day,
• fewer errors and less rework, such as incorrectly entered orders,
• shorter onboarding, because the knowledge of experienced colleagues is available in the system.

Against that stand the build, the usage costs of the models, operations and team adoption. The formula is the same as for any investment: ROI equals benefit minus cost, divided by cost. You will find the full set of formulas in our guide to AI project ROI.

What ROI does AI deliver in knowledge management?

AI in knowledge management means a system finds the answer in your own documents, tickets, contracts, product data and manuals and phrases it with a source reference. The technical term is retrieval augmented generation, or RAG. If the system keeps going on its own, for example answering a customer request directly in the email thread or opening a follow-up in the ticket system, it becomes a knowledge agent.

The return comes from search time and onboarding. The McKinsey Global Institute estimated, based on IDC data, that knowledge workers spend about 19 percent of their working time looking for information. In Atlassian's "State of Teams 2025" survey, leaders and teams put it at 25 percent. Both are estimates, but the order of magnitude has held for more than ten years: roughly a fifth of working time.

The strongest evidence comes from the Brynjolfsson customer support study. Agents with two months of tenure performed with AI support as well as colleagues without AI who had more than six months. The system makes the knowledge of the best people available to everyone. Demographics make this more urgent, especially in Germany: according to the German Economic Institute (IW), 14.1 million baby boomers will reach retirement age by 2036, while only 9.8 million young people reach working age. Whatever lives only in one person's head leaves with them. How public authorities tackle this is covered in our article on knowledge transfer in public administration.

Worked example: knowledge management

Say an inside sales team of 20 people loses 45 minutes per person per day to searching and asking around, less than half the study figures. With 220 working days and 45 euros an hour, that comes to 3,300 hours or 148,500 euros a year.

A knowledge system halves the search time (assumption). Because saved time only counts if it gets used, we count just half of it as a return: 37,125 euros. Add three new hires a year who reach full output four weeks sooner, at half productivity during that period, which is worth 10,800 euros. Usage costs for 20 people with 30 requests a day at 6,000 tokens each come to about 790 million tokens a year. With a frontier model through Amazon Bedrock's EU profile, that is around 3,500 US dollars, which we put at 3,500 euros for simplicity.

In knowledge management, measure before and after the start:

• time to answer from an incoming question to a complete answer,
• share answered without asking an expert,
• onboarding time until new staff reach the usual case volume,
• measured search time from usage data, not from surveys.

What ROI does process automation deliver?

Process automation pays off when a transaction is frequent, follows recognizable rules and today gets stuck at the handover between systems or people. There are three tools to choose from: a fixed workflow that moves data from system A to system B by rules, RPA, a software robot that imitates screen input where no interface exists, and an AI agent that understands unstructured input such as emails or PDFs. The process decides which one fits.

Classic automation has a long series of surveys behind it. In Deloitte's survey "Automation with intelligence" (2022), 479 executives expected an average cost reduction of 31 percent in the affected areas. More than half, however, had never calculated the actual reduction. Whoever measures has the advantage, including with their own bank.

Worked example: invoice intake

Say a company enters 12,000 incoming invoices a year by hand, six minutes each, at 45 euros an hour. That is 1,200 hours or 54,000 euros. An automation reads the invoices, checks them against purchase order and goods receipt and prepares the posting. 70 percent pass through without anyone touching them (the straight-through processing rate), accounts payable handles the other 3,600 as before, and we allow half a minute of spot-checking for each automated invoice. The new effort is 430 hours or 19,350 euros, and we put model usage at 600 euros a year. The net annual saving is 34,050 euros.

The last line shows what matters. Profitability hinges on one figure you can test in advance on your own documents. How to put a euro figure on the current effort of a process is explained in our guide to calculating the cost of manual processes.

The right metrics for process automation:

What ROI do AI agent systems deliver?

An AI agent is a system that is given a goal and chooses the steps to get there: it reads an order, finds the customer in the ERP, checks prices and stock, creates the order and only reports back on an exception. A multi-agent system spreads the work across several specialized agents, for example one for reading, one for checking and one that coordinates. The difference from a workflow: the workflow follows fixed rules, the agent decides for itself within its guardrails.

Agents are the fastest-growing area. According to BCG ("The Widening AI Value Gap", 2025), agents deliver 17 percent of the total AI value of the companies surveyed today, and they expect 29 percent by 2028. According to McKinsey, 62 percent of organizations are experimenting with agents and 23 percent are already scaling them somewhere in the business. In Germany, according to Bitkom (September 2026), 11 percent of companies that use, plan or are discussing AI already run agents, and another 29 percent plan to.

Independent, measured ROI figures for agents are still scarce. Most success stories come from vendors. What research does show is the lever that decides ROI: the success rate. In Carnegie Mellon University's TheAgentCompany benchmark, the best agent completed 30 percent of realistic office tasks fully on its own. In Salesforce's CRM test CRMArena-Pro, leading agents reached about 58 percent on single-step tasks and about 35 percent in multi-turn conversations, while executing fixed workflows succeeded more than 83 percent of the time. That leads to a clear design rule: where the procedure is known, give the agent a fixed sequence with guardrails, and let it decide freely only where the inputs vary.

The cost side is different for agents too. According to Anthropic, single agents use about four times and multi-agent systems about fifteen times as many tokens as a simple chat. More agents are not automatically better: a controlled study by Google Research, Google DeepMind and MIT (2025) found 80.8 percent better results from multiple agents on finance tasks that split up well, but drops of up to 70 percent on tasks that must run strictly in sequence. A multi-agent system therefore pays off for parallel work such as research or checking many documents, while an order chain is usually best served by a single agent with a clear sequence.

A field experiment in Alibaba's customer service (2026) shows why the human in the loop belongs in the ROI calculation: the agents shortened chats, but customer ratings for those chats dropped markedly. When staff stepped in, they mainly held quality on technical problems, less so with upset customers, and early intervention worked better than late. Gartner expects more than 40 percent of agentic AI projects to be cancelled by the end of 2027 because of rising costs, unclear value and missing risk controls. All three can be settled before anything is built. What a clean human-in-the-loop approval step looks like is covered separately.

Worked example: an AI agent for order entry

Say a wholesaler enters 30,000 orders a year from customer emails, eight minutes per order, at 45 euros an hour. That is 4,000 hours or 180,000 euros. An agent reads the emails, matches customer and items and creates the order. Whatever it cannot match with confidence it hands, prepared, to inside sales, who then need six minutes instead of eight. We put usage costs at 15 cents per order because of the higher token use, and errors that slip through at 0.5 percent of orders with 40 euros of follow-up cost each. For orders the agent creates automatically, we allow half a minute of spot-checking.

The agent pays off in both cases because the escalation is well designed: the person gets a prepared case instead of an empty form. The success rate shifts the return by more than 40,000 euros a year. That is why a test with real emails belongs before the investment decision.

For agent systems, these metrics belong in the ROI calculation:

• cost per successfully completed transaction, including tokens, operations and human review,
• success rate across many runs, not from a demo,
• escalation rate and effort per escalation,
• error and follow-up costs,
• customer satisfaction as a counterweight to handling time.

On premises, data center or API: which way of running AI pays off?

The same agent can run in four ways, and the choice changes both running costs and which data it may process. Prices are taken from the providers' own price lists, as of November 2026. They change quickly.

When does your own hardware pay off?

Assuming three parts input to one part output, Claude Sonnet 5.5 through Bedrock in the EU averages about 4.40 US dollars per million tokens. A rented server at Hetzner for 999 euros a month becomes cheaper from about 260 million tokens a month, provided an open model is good enough for the job. Compared with the open model gpt-oss-120b through the same API (about 0.35 US dollars per million tokens on average), the break-even point only comes at about 3.3 billion tokens a month. For scale: 500 users with 20 requests of 3,000 tokens each on 22 working days add up to about 660 million tokens.

The upshot: at small and medium volumes, an API is almost always cheaper, costs grow with use and there is no upfront investment. Your own hardware rarely pays off on token price. It pays off on data sovereignty, as in hospitals, and staff for running and updating it come on top. How an agent's usage costs arise is covered in our article on AI agent running costs, and the basic choice in self-hosted vs cloud AI.

What does the GDPR say about the four options?

On premises, there is no external service provider and no third-country transfer. In its guidance on AI and data protection (2024), the German Data Protection Conference explicitly calls technically closed systems preferable. The other obligations, such as a legal basis and possibly a data protection impact assessment, remain.

In a European data center, the provider acts as a processor and a data processing agreement is sufficient. If the provider belongs to a group outside the EU, the Data Protection Conference calls for a careful case-by-case review, but there is no blanket ban.

Through AWS in an EU region, processing stays in the EU. According to AWS, model providers have no access to prompts and responses, and storage can be switched off. Claude Sonnet 5.5 is only available on Bedrock through the EU profile, which processes in several EU regions, not only Frankfurt. Because the parent company is American, a residual risk from US access rights remains and belongs in the risk assessment. Since January 2026, AWS also runs the European Sovereign Cloud in Brandenburg, operated exclusively by EU-resident staff. So far only a small selection of models is available there, such as Amazon Nova and open models like Gemma 4.

The direct API of a US model provider is legally permissible with a data processing agreement and a recognized basis for transferring data to the US, and common practice for non-critical data. For health data, data under professional secrecy or strictly regulated areas, however, it is usually ruled out, because the data can leave the EU and the transfer agreement with the US (the Data Privacy Framework) is being challenged before the European Court of Justice. Providers also differ: according to its own documentation, Anthropic's direct API processes and stores data only in the US or globally, and an EU option exists only through AWS or Google. OpenAI offers EU data residency for API projects at a 10 percent premium.

For hospitals and medical practices in Germany, professional secrecy and the cloud rules of social law add further requirements (§ 393 SGB V: processing in Germany, the EU or a country with an adequacy decision, a branch in Germany, a current C5 attestation). For this data, on-premises operation or a data center in Germany is the direct route. More on our hospitals page.

How do you measure AI ROI properly?

With a baseline, a comparison and a metric that reaches the bottom line. Brynjolfsson and his co-authors could only prove the effect in customer support because the tool reached different teams at different times, and the teams without it served as the comparison. For your company that means:

• Take a baseline before anything is built: cost per transaction, throughput time, error rate, ideally from timestamps in the ERP, ticket system or inbox.
• Pick one metric that counts: cost per invoice, orders per day, requests answered without escalation, not "minutes saved".
• Compare instead of believing: one team, site or document type with the system, a comparable one without, or before and after in the same season.
• Log running costs: token costs per transaction go on the same dashboard as the savings.
• Plan the freed-up time: clear backlogs, cut overtime, avoid filling an open position externally.

Self-reports are not enough. In an experiment by the research group METR (2025), experienced developers were 19 percent slower with AI but believed they had been 20 percent faster. Justify the investment with hard ROI, meaning figures in euros, and mention soft ROI such as satisfaction or less dependence on individual people as a bonus.

How do you introduce AI and connect it to your systems?

The studies agree to an unusual degree here. According to McKinsey, the most successful companies are almost three times as likely to fundamentally redesign their workflows rather than putting AI on top of old procedures. According to BCG (2024), the leading 26 percent of companies pursue about half as many use cases as the rest and, following a 70-20-10 rule of thumb, put most of their resources into people and processes, not algorithms. According to MIT, purchased solutions and partnerships reach production much more often than purely internal builds.

The bottleneck is almost always integration. According to Bitkom, 54 percent of German logistics companies lack the expertise to bring AI into their processes. Connecting AI to your infrastructure means concretely: access through documented interfaces or, where none exist, a workaround (see connecting systems without an API), sign-in through your existing user management, a log of every step and a clear rule for who takes over when an exception occurs. Autonomy rises in steps: first the system prepares and a person decides, and only once the error rate is measured and low does it decide routine cases on its own. Each step raises the straight-through rate and with it the ROI.

As for the legal framework, the EU AI Act has largely applied since 2 August 2026, including the transparency obligations for AI-generated content. With the so-called Digital Omnibus, the EU moved the obligations for stand-alone high-risk systems to 2 December 2027. Typical office automation such as invoice intake, order entry or quote preparation is not a high-risk system.

Where does AI pay off in which industry?

When insurers implement AI, the same success factors decide as everywhere else: a well-defined, high-volume process such as claims handling, a baseline for handling time and cost per case, approval rules for borderline cases and a way of running the model that fits the data. In the EIOPA survey published in February 2026, 49 percent of European insurers had their own AI policy. More on our insurance page.

What does this mean for your decision?

Start with the area that has the clearest transaction. In most companies that is a process automation or an agent with high volume, such as invoice intake, order entry or quotes. Knowledge management pays off where many people ask the same questions about the same documents and onboarding is expensive. Measure the baseline, test with real transactions and choose how to run the model according to the data: non-critical through an API, sensitive in a European data center or on premises.

That is also how Sophera Consulting works. In the free Automation Check, we pick the process that makes sense as a starting point and clarify which access we need. The quote comes with a cost-benefit analysis at a fixed price, with no subscription. Depending on the process, we build a workflow, an RPA solution or an AI agent; an agent is built in one to two days, and the test with real cases runs the same week. The models run GDPR-compliant on European infrastructure or, on request, locally on your own hardware. After that you only pay the usage costs of the models, billed directly to you, and a maintenance agent included in the fixed price watches the interfaces. Depending on the process, we save on average 20 to 80 percent of process costs. That is our own figure, not a study, which is exactly why we measure it against your baseline.

Sources

• Afshar, M. et al.: ambient AI documentation in a pragmatic trial, NEJM AI 2(12), 2025
• Alibaba/Taobao: Agentic AI and Human-in-the-Loop Interventions, field experiment, arXiv 2605.14830, 2026
• Anthropic: How we built our multi-agent research system, June 2025
• AWS: Amazon Bedrock pricing (eu-central-1) and documentation on data protection and data retention, accessed November 2026; launch of the AWS European Sovereign Cloud, January 2026
• Becker, J. et al. (METR): Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, 2025
• BCG: Where's the Value in AI?, October 2024; The Widening AI Value Gap, September 2025
• Bitkom: AI adoption survey, September 2026; AI in logistics, September 2026; Industry 4.0, March 2025
• Brynjolfsson, E.; Li, D.; Raymond, L.: Generative AI at Work, NBER Working Paper 31161, 2023; Quarterly Journal of Economics 140(2), 2025
• Choi, J.; Monahan, A.; Schwarcz, D.: Lawyering in the Age of Artificial Intelligence, Minnesota Law Review 109, 2024
• Cui, Z.; Demirer, M. et al.: The Effects of Generative AI on High-Skilled Work, Management Science
• German Data Protection Conference (DSK): guidance on AI and data protection, May 2024; resolution on extraterritorial access, January 2023
• Dell'Acqua, F. et al.: Navigating the Jagged Technological Frontier, Organization Science, 2026
• Deloitte: Automation with intelligence, 2022
• EBA: Rising application of AI in EU banking and payments sector, 2025
• EIOPA: survey on generative AI among insurers, February 2026
• Gartner: press release on agentic AI, 25 June 2025
• Google Research, Google DeepMind, MIT: Towards a Science of Scaling Agent Systems, 2025
• Hetzner, OVHcloud, Nvidia, Anthropic, OpenAI: price lists, accessed October and November 2026
• Humlum, A.; Vestergaard, E.: Large Language Models, Small Labor Market Effects, NBER Working Paper 33777, 2025
• IBM Institute for Business Value: CEO Study 2025
• German Economic Institute (IW): IW-Kurzbericht 39/2026 on baby boomers reaching retirement age
• Lukac, P. et al.: Ambient AI Scribes in Clinical Practice, NEJM AI, 2025
• McKinsey Global Institute: The social economy, 2012
• McKinsey: The State of AI 2025
• MIT NANDA: The GenAI Divide, State of AI in Business 2025
• Noy, S.; Zhang, W.: Experimental evidence on the productivity effects of generative artificial intelligence, Science 381, 2023
• Salesforce AI Research: CRMArena-Pro, 2025
• Xu, F. F. et al. (Carnegie Mellon University): TheAgentCompany, 2024/2025
• Regulation (EU) 2024/1689 (AI Act) and Regulation (EU) 2026/1744 (Digital Omnibus on AI)
• Wharton School and GBK Collective: Accountable Acceleration, October 2025