AI in the systems you already use: where it saves time and money

AI saves time and money in the systems a company already runs (an ERP such as SAP, DATEV, a CRM such as Salesforce or HubSpot, Microsoft 365, the help desk) by taking over the work that sits between them: reading invoices and emails, typing what they say into the right fields, sorting requests, looking things up and drafting replies and reports. The gains are real but uneven: in controlled studies, support agents with an AI assistant resolved 15% more issues per hour and consultants finished suitable tasks 25% faster, yet made more mistakes where the AI was weak. The business case is plain arithmetic: hours spent today × the share AI takes over × the fully loaded cost of an hour, plus avoided errors, minus running costs and review time.

#Where the time goes today

The hours rarely disappear into one big task. They leak through hundreds of small hand-offs between systems that do not talk to each other.

• Manual data entry. In a July 2025 survey of 500 US professionals, run by Parseur with QuestionPro, respondents said they spend more than 9 hours a week moving data from emails, PDFs and spreadsheets into systems, at an estimated $28,500 per employee a year. Parseur sells document processing software, so read it as a vendor survey.
• Document handling. Invoices, delivery notes, contracts and certificates are opened, read, renamed, filed and keyed. In APQC's benchmark of 1,485 organisations, the median cost to process one invoice was $5.83, the top quarter spent $2.07 or less and the bottom quarter $10 or more. Much of the gap is how many invoices someone touches.
• Email triage. Shared inboxes where a person reads every message only to decide who should read it next.
• Reporting. The median monthly close in APQC's benchmark of 2,300 organisations took 6.4 calendar days; the top quarter needed 4.8 or fewer.
• Lookups. Which contract version did the customer sign, what did we agree last spring: answered by searching SharePoint, the ERP and old email threads. In the UK government's Copilot trial, over 70% of users agreed the assistant cut the time they spent searching for information.

#Why AI belongs inside the systems you already use

Adoption is climbing but far from universal. According to Germany's Federal Statistical Office, 26% of companies with at least ten employees used AI in 2025: 23% of small firms, 36% of medium-sized ones and 57% of large ones. Among companies that had considered AI but not used it, the reasons given were lack of knowledge (72%), unclear legal consequences (62%), data protection concerns (60%) and incompatibility with existing devices, software and systems (45%).

That last point is the practical one. A standalone AI tool adds one more window to copy and paste between. The value appears when the agent reads from and writes to the system of record: the invoice arrives in the ERP as a posting proposal, the call summary lands in the CRM as an activity, the classified email becomes a ticket with the right customer attached. The system of record stays the single source of truth: AI prepares, a person approves, the system books. Start read only, give the agent a service account with the least rights that work, and let it write only through narrow, logged actions such as "create posting proposal".

Built-in assistant or custom agent

SAP, Salesforce, HubSpot and Microsoft all ship assistants inside their products. They are the quickest start when the work stays inside one product. A custom agent, connected through each system's API, earns its cost when a process crosses systems (email to ERP to DATEV) or when you need control over hosting, model choice and logs.

#What AI does well, and what it does not

The best evidence comes from two field experiments. In a study of 758 BCG consultants, those using GPT-4 on tasks inside the AI's capabilities completed 12.2% more tasks, 25.1% faster, at more than 40% higher quality. On a task just outside them, consultants with AI were 19 percentage points less likely to reach the correct answer than those without. In the Quarterly Journal of Economics study of 5,172 customer support agents, an AI assistant raised issues resolved per hour by 15% on average and by about 30% for less experienced agents, while the most skilled saw little gain in speed and a small drop in quality.

The practical rule follows. Work comes in three kinds. Fixed rules ("invoices over €5,000 go to the CFO") belong to classic automation: cheaper, faster and fully predictable, no model needed. Unstructured input with a result someone can check (an invoice PDF, a free text email, a call transcript) is where AI earns its place. Judgement, accountability and exceptions stay with people, with AI preparing the file.

#Use cases by department

Finance and accounts payable

An agent reads each incoming invoice, whether PDF, scan or structured e-invoice, proposes supplier, GL account, cost centre and tax code, matches it against the purchase order and goods receipt, checks for duplicates and chases the approver. Germany's e-invoicing rules help: since 1 January 2025 every domestic business must be able to receive e-invoices, so more invoices arrive as data (deadlines in our e-invoicing guide). Keep payment release, new suppliers and any change of bank details with people: the last is a classic fraud route.

Customer service

Klarna's assistant handled two thirds of its customer service chats in its first month, 2.3 million conversations, which the company described as the work of 700 full-time agents, with an estimated $40 million profit improvement for 2024. Fifteen months later its chief executive told Bloomberg that Klarna would recruit people again so customers could always reach a human, conceding that cost had weighed too heavily. Both halves are the lesson: automate routine requests, keep a fast route to a person. Our guide to AI agents in customer service covers hand-off rules and metrics.

Sales operations

An agent turns a call transcript or meeting notes into a CRM activity with next steps, updates deal fields, enriches new leads from their website and drafts the follow-up email and the quote from the ERP price list. The salesperson reviews and sends. Cleaner CRM data is often the bigger win, because forecasts depend on it.

HR

Good first uses are answering policy questions from the handbook with the paragraph linked, collecting onboarding documents and preparing contracts from approved templates. Ranking CVs or evaluating staff is a different category: such systems are high risk under Annex III of the AI Act, with heavy duties for the companies using them, as our AI Act guide explains.

IT

For password resets, access requests and how-to questions, an agent classifies the ticket, answers from the knowledge base, collects the details an access request needs and routes it to the approver. Privileged access and changes to production systems stay with an engineer.

#How to calculate the business case

1. Hours today: people doing the work × hours each spends on it a week × 46 working weeks.
2. Hours AI takes over: hours today × the share AI handles. Start cautious and replace the guess with the share you measure in a pilot.
3. Value of that time: hours taken over × fully loaded hourly cost (salary, employer contributions, workplace and overhead).
4. Errors and delays avoided: rework, duplicate payments, missed early payment discounts, late fees, SLA penalties.
5. Running costs: licences, model usage, hosting, maintenance and the human time spent reviewing the AI's work.
6. Payback: one-off cost ÷ monthly net benefit. Full-time equivalents: hours taken over ÷ 1,720 hours a year.

A worked example (hypothetical)

Assumptions, all hypothetical: a company of 250 people receives 18,000 supplier invoices a year. Three accounts payable clerks each spend 25 hours a week keying, coding, matching and chasing approvals. An hour costs €45 fully loaded. Two percent of invoices need correction later, at 45 minutes each. Licences, model usage and hosting cost €30,000 a year. Reviewing the agent's proposals takes 10% of the hours it takes over.

In the base case each €10,000 of one-off cost pays back in about two months; in the cautious case it takes about eight. The share AI takes over moves the result more than any other input, which is why it should be measured in a pilot before anyone signs a large budget. The savings calculator on our process automation page uses the same logic, 46 working weeks and 1,720 hours per full-time role, with your own numbers.

#What the published evidence says

Two patterns run through all of it, and through Klarna's experience above. The gains are largest on high volume, well documented work, especially for people who are not yet experts in it. And the gains hold when a person stays in the loop for what the AI handles badly.

#How to start safely

1. Pick one process with high volume, written rules, a checkable result and cheap mistakes. Invoice capture and ticket triage usually qualify.
2. Measure the baseline for two to four weeks: volume, touch time per item, cycle time, error rate.
3. Run in shadow mode. The agent proposes, people decide, nothing is written without approval. Grade a sample every week.
4. Design the human in the loop: approval thresholds, a confidence level below which items go to a person, a weekly sample review.
5. Widen the scope only when the numbers hold, one document type or request type at a time.

GDPR and EU hosting

Invoices, emails and CRM records contain personal data, so the GDPR applies in full: a data processing agreement under Article 28 with every AI vendor, processing in an EU region, a check whether your data trains the vendor's models, a retention period for prompts and logs, and care with Article 22, which restricts decisions with significant effects on people made by software alone. Hosting options are compared in our guide to GDPR-compliant AI hosting in Germany.

The EU AI Act in brief

For most back-office agents, two duties apply today. Since 2 February 2025, Article 4 requires that staff who use AI systems have sufficient AI literacy. Since 2 August 2026, Article 50 requires that people are told when they are interacting with an AI system, so a customer-facing agent must say so at the latest at the first interaction. HR uses such as CV ranking are high risk. Details, roles and dates are in our guide to the EU AI Act for AI agents.

#Measuring results

Measure the same things, the same way, before and after. Vendor dashboards count what is easy; you want what changed in the work.

Report monthly in money, and say what happened to the hours: absorbed growth, reduced overtime, a vacancy not refilled, or capacity not yet used. That last line keeps the business case honest.

#How we can help

Bespoke IT's free AI audit maps one or two of your processes across the systems you use, estimates the hours and the realistic share for AI, checks your GDPR and AI Act position and gives you a staged plan with the metrics to judge it by. When the process is clear, our process automation team builds the agents into SAP, DATEV, Salesforce, HubSpot or Microsoft 365, with the approval steps your team keeps.