In short
The AI use cases with the best documented results in portfolio companies are price and terms discipline in proposals, order intake, customer service, invoice processing and finance, and knowledge search. Reported effects: about 1 to 2 points of margin from AI pricing at distributors, order processing time cut by half at Danfoss, customer service and operations cost down 15% at Klarna in 2024, 30% of accountants’ time freed at Eurofins, and a few percent of office time from AI assistants. Most figures are self-reported or vendor cases, and time saved only partly becomes cost saved. Which cases fit depends on each company’s cost structure, data and pricing power.
Key takeaways
- Price and terms discipline is the strongest lever because it flows almost fully to profit; documented gains at distributors are about 1 to 2 points of margin, in one case more than 2.5 points.
- Order intake from e-mail and PDF is a well-documented operational case: Danfoss halved order processing time, Oiki cut time per order from 30 to 12 minutes (vendor cases).
- Customer service shows real cost effects, but also the limits: Klarna’s cost fell 15% in 2024 and rose 2% in 2025, and the company began hiring human agents again.
- Finance cases save a lot of time in payables and receivables; how much of it becomes cost saving depends on what happens to the freed hours.
- Organisation-wide AI assistants save a few percent of office time at most, and there is little evidence yet that it turns into productivity.
- Pick cases per company by frequency, data, checkability and an owner, not by what is fashionable.
How to read the numbers in this article
There is no shortage of AI success stories. There is a shortage of numbers you can put into a value creation plan. This overview uses only results at named or clearly described companies, from filings, owners’ statements, peer-reviewed studies and vendor case studies, and says which is which. Vendor cases describe their best customers; filings and controlled studies are more reliable. Treat every figure as an upper reference for a company that does the work well, not as a forecast.
| Lever | Documented result | Type of evidence |
|---|---|---|
| Pricing and terms | More than 250 bps of margin at a $15bn distributor, over 200 of them from AI price and discount tools in about 18 months; about 1–2 points at other distributors | Consultancy case studies (anonymised) |
| Order intake | Danfoss: order processing time −50%; Oiki: 30 → 12 minutes per e-mailed order | Vendor cases |
| Customer service | Klarna: customer service and operations cost −15% in 2024; Cengage: customer-care cost −15% in early results | Filing; owner via Bain |
| Finance | Eurofins: 30% of accountants’ time freed; Uber: invoice handling time −70%, process cost −25–30% | Vendor case; self-reported |
| Knowledge work | UK government trial: 26 minutes a day; Danish survey: 2.8% of work hours | Self-reported trial; survey |
1. Pricing and proposals: the strongest lever
Price flows almost fully to profit. At a 12% EBITDA margin, one point of price on all revenue adds 8.3% to EBITDA, as long as volume holds. AI helps here less by setting prices than by enforcing discipline: it drafts the proposal from what the company has offered before, checks discounts and payment terms against the rules and flags deviations for a person to decide.
The documented results come from distribution. A $15bn distributor gained more than 250 basis points of margin, more than 200 of them from AI price and discount tools over about 18 months and about 50 from an agentic pricing copilot (McKinsey 2026, anonymised). BCG and McKinsey report about 1 to 2 points at other distributors (BCG 2026; McKinsey 2026). McKinsey also writes that value from generative and agentic AI “has been limited in pricing so far”.
Fits best: many quotes, list prices with discount freedom, a sales team that negotiates. Fits less: revenue on framework contracts or fixed price lists.
2. Order intake: a well-documented operational case
In many mid-sized industrial and distribution companies, orders still arrive by e-mail, PDF or spreadsheet and are typed into the ERP by hand. AI reads the order, matches customer, items and prices against the system and presents the finished order for approval. People only check what stands out.
Danfoss cut order processing time by 50% with AI agents that read B2B order e-mails (Google Cloud case). Oiki cut the average time per e-mailed order from 30 to 12 minutes (IBM case). Both are vendor cases. How we build such a tool, with a specification, test cases and human sign-off, is described in Vibe tooling: AI order entry.
Fits best: many orders in free format, stable item master data, an ERP that stays the system of record. Watch: freed order-entry time often moves into other tasks rather than out of the cost base.
3. Customer service: real effects, real limits
Customer service has a lot of public evidence, and some of it is instructive. Klarna’s customer service and operations cost fell 15% in 2024, the year it launched its AI assistant (F-1 filing); the assistant handled 80% of chats in 2025, yet the same cost line rose 2% that year (20-F), and Klarna began hiring human agents again after its CEO said quality had suffered (Entrepreneur).
In private equity portfolios, Cengage, in which Apollo holds a minority stake, cut customer-care cost by 15% in early results (Bain 2025), and Vista’s Assent reports more than 10% full-time employee savings in support at more than 33% deflection (Vista 2026).
Lesson: a contact handled by AI is not a cost removed. Plan on cost reductions in the low double digits, keep a human route for complex cases and measure customer satisfaction alongside cost.
4. Finance: invoices, close and reporting
Payables and receivables are full of routine: reading invoices, matching them to orders, chasing differences. Eurofins says automating invoice processing freed up 30% of its accountants’ time (Rossum case). Maximus cut invoice processing time by 85% and its AP team from 10 to 8 people through natural attrition (Rossum case). Uber cut invoice handling time by 70% and process cost by 25–30% (Uber 2025).
For smaller companies, a field study is more representative: accountants at 79 small and mid-sized firms using an AI tool moved about 8.5% of their time off routine data entry and cut monthly close time by 7.5 days (Choi & Xie 2025, observational). A faster close is valuable in its own right for a fund: board-ready numbers sooner.
5. Knowledge work: assistants, search and onboarding
Organisation-wide assistants are a common AI investment and among the hardest to turn into EBITDA. In a UK government trial across 12 organisations, users reported saving 26 minutes a day (UK GDS 2025); chatbot users in a survey of 25,000 Danish workers reported saving 2.8% of their work hours (Humlum & Vestergaard 2025). The UK Department for Business and Trade found no evidence that saved time became productivity (UK DBT 2025).
Targeted tools do better than general assistants: a knowledge search that answers with sources from the company’s own policies, an onboarding assistant, a staffing engine that knows skills and availability. They solve a named problem and can be measured.
Why time saved is not money saved
The finance cases show the gap clearly. Uber cut handling time by 70% but process cost by 25–30%; Maximus cut processing time by 85% and its team by 20%. By our conversion, 24–43% of the saved time left the cost base. The rest went into other work, quality or growth, which can be valuable, but is not a cost saving.
How to choose the first use case in a portfolio company
- Frequency: the task occurs daily or weekly, not quarterly.
- Value: it touches price, a large cost pool or cash, and the effect can be named in euros.
- Data: the inputs exist in usable form, and the company has the rights to use them.
- Checkability: a person can check the result before it takes effect. AI does the groundwork; a person decides where there is no clear answer.
- Fit: the task sits inside what AI does well. Outside that frontier, consultants using AI were 19 percentage points less often correct (Dell’Acqua et al. 2026).
- Owner: one manager in the company owns the result and reports it.
- Acceptance: the team that does the work wants the tool and helped to define its test cases. Without that, the other six do not matter (AI change management).
A case that passes all seven is usually worth a controlled MVP, which typically takes two to four weeks of build time. Partners Group reports 14 weeks from concept to production for its first case at Foundation Risk Partners (Partners Group 2026). How to anchor the cases in the value creation plan is the topic of AI in the value creation plan.
We do this work for private equity funds as an impact partner: portfolio scan, calculated value cases, tools on the company’s own systems, leaders who carry them. More on pe.leadbeyond.de.
Frequently asked questions
What are the best AI use cases for private equity portfolio companies?
Those with the best documented results: price and terms discipline in proposals, order intake from e-mail and PDF, customer service, invoice processing and finance, and targeted knowledge search. Which fit depends on each company’s cost structure, data and pricing power.
Which AI use case pays off fastest?
Often order intake or invoice processing, because they are frequent, easy to check and well documented. Pricing usually has the largest effect on EBITDA, but takes longer to show because contracts and price lists change slowly.
How reliable are published AI results?
Filings and controlled studies are the most reliable, then owners’ statements, then vendor case studies, which describe their best customers. Most published figures are self-reported. Use them as an upper reference for a company that does the work well, not as a forecast.
Does AI in customer service reduce cost?
It can: Klarna’s customer service and operations cost fell 15% in 2024, and Cengage cut customer-care cost by 15% in early results. But Klarna’s same cost line rose 2% in 2025 and it began hiring human agents again, so plan on a human route and measure quality alongside cost.
Do AI assistants like Copilot or ChatGPT raise EBITDA?
There is little evidence yet. Users report saving 26 minutes a day in a UK government trial and 2.8% of their hours in a Danish survey, but the UK Department for Business and Trade found no evidence that saved time became productivity. Targeted tools for a named process are easier to measure.
Sources
- McKinsey & Company. (2026). B2B pricing: Navigating the next phase of the AI revolution., McKinsey & Company (2026). Anonymised case: a $15bn distributor gained more than 250 bps of margin, more than 200 of it from AI price and discount tools over about 18 months; also states that value from gen AI and agentic AI “has been limited in pricing so far”.
- BCG. (2026). Why AI in B2B pricing isn’t plug and play., Boston Consulting Group (2026). Cited for margin gains of about 1 to 2 points from AI pricing at distributors.
- McKinsey & Company. (2026). Pricing in distribution: A strategic growth engine., McKinsey & Company (2026). Cited with BCG 2026 for margin gains of about 1 to 2 points from AI pricing at other distributors.
- Google Cloud. Danfoss customer story., Google Cloud. Vendor case: order processing time cut by 50% with AI agents that read B2B order e-mails.
- IBM. (2026). Oiki case study., IBM (2026). Vendor case: average time per e-mailed order from 30 to 12 minutes.
- Klarna Group plc. (2025). Form F-1 registration statement (amendment)., U.S. Securities and Exchange Commission (EDGAR) (2025). Filing: customer service and operations cost fell 15% in 2024, the year the AI assistant launched (the line includes operations).
- Klarna Group plc. (2026). Annual report on Form 20-F for 2025., U.S. Securities and Exchange Commission (EDGAR) (2026). Filing: the AI assistant handled 80% of chats in 2025; the same cost line rose 2% in 2025.
- Entrepreneur. (2025). Klarna CEO reverses course by hiring more humans, not AI., Entrepreneur (2025). Report that Klarna began hiring human agents again after its CEO said quality had suffered.
- Bain & Company. (2025). Field notes from the generative AI insurgency. Global Private Equity Report 2025., Bain & Company (2025). Survey of PE firms (September 2024) on AI use in portfolio companies; cited for the Cengage example (customer-care cost down 15% in early results, Apollo minority stake).
- Vista Equity Partners. (2026). Vista Outlook: Software’s transition to agentic enterprise AI., Vista Equity Partners (2026). Cited only for the Assent example: more than 10% full-time employee savings in support at more than 33% deflection (self-reported).
- Rossum. Eurofins customer story., Rossum. Vendor case: automating invoice processing freed up 30% of accountants’ time.
- Rossum. Maximus customer story., Rossum. Vendor case: invoice processing time cut by 85%; AP team from 10 to 8 people through natural attrition.
- Uber Engineering. (2025). Advancing invoice document processing using GenAI., Uber (2025). Self-reported: invoice handling time cut by 70%, process cost by 25–30%.
- Choi, J. H., & Xie, C. (2025). Human + AI in accounting: Early evidence from the field. Stanford GSB working paper., Stanford Graduate School of Business (2025). Observational data from accountants at 79 small and mid-sized firms: about 8.5% of time moved off routine data entry; month-end close 7.5 days faster.
- UK Government Digital Service. (2025). Microsoft 365 Copilot experiment: Cross-government findings report (June 2025)., UK Government (2025). Trial with 20,000 employees across 12 organisations: 26 minutes a day saved, self-reported.
- Humlum, A., & Vestergaard, E. (2025). Large language models, small labor market effects (BFI Working Paper 2025-56). Becker Friedman Institute, University of Chicago., Becker Friedman Institute for Economics (2025). Survey of 25,000 Danish workers: chatbot users report saving 2.8% of their work hours; firm-led initiatives (encouragement, in-house tools, training) nearly double take-up for the typical worker, from 47% to 83%.
- UK Department for Business and Trade. (2025). The evaluation of the M365 Copilot pilot in the Department for Business and Trade., UK Government (2025). Found no evidence that time saved turned into higher productivity.
- Partners Group. (2026). Partners Group drives EBITDA margin uplift through AI collaboration between private equity portfolio companies. Press release via EQS News., Partners Group / EQS News (2026). Press release of 3 August 2026, owner’s own figures: two AI use cases at Foundation Risk Partners worth $10m of EBITDA, 120 bps of margin; close rates doubled; first use case from concept to production in 14 weeks. Self-reported.
- Dell’Acqua, F., McFowland III, E., Mollick, E., Lifshitz, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2026). Navigating the jagged technological frontier: Field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality. Organization Science, 37(2), 403–423., INFORMS (2026). Pre-registered field experiment with 758 consultants: inside the AI’s frontier 12.2% more subtasks, 25.1% faster, 32% higher graded quality; on a task outside it, 19 percentage points less often correct.