SI Integration in the Services Industry: A Guide to Preparing for 2027

There is a large difference in responsibility between answering a customer’s question “When is my appointment?” and cancelling that appointment or approving a refund amount. The first task requires access to current data and a clear answer. The second requires identity verification, explicit authorisation, a transaction record and error handling.

SI integration in the services industry is taking shape around this difference. Customer service, professional services, hospitality, transport and back-office operations can draw on similar technologies. In each of them, however, the cost of an error, the sensitivity of the data and the role of human expertise are different.

In the Webtures approach, a strong services SI strategy addresses four areas together: the accuracy of the information given to customers, the efficient use of employee capacity, control over transaction authority and the measurability of the outcome. Model selection is one part of this structure. Process design, information quality and human handover are at least as decisive.

Preparing for 2027 requires setting out clearly, before automating every interaction, which tasks belong to an assistant, which to an agent with limited authority and which to a human expert. This guide is based on global sources and is not a substitute for individual legal advice.

Executive summary

What is SI integration in the services industry?

The services industry covers activities that deliver work, information or an experience rather than producing physical goods. In the World Bank’s definition, services include wholesale and retail trade, hotels and restaurants, transport, storage and communication, finance, real estate, business services, public administration, education, health and other personal services.

SI integration, in turn, means that capabilities such as prediction, content generation and tool use work in connection with the organisation’s real data, business rules and applications. An employee asking a general-purpose chat tool a question and a system that generates answers from the organisation’s approved sources, respects access permissions and records human handover are different levels of maturity.

Four approaches stand out in services applications:

These structures can work together. A forecasting model can anticipate busy hours while a language model classifies incoming requests, and an agent prepares appointment options within approved rules. The role and responsibility of each component should be defined separately. In clear and stable processes, rule-based automation can also be a more understandable and economical option.

2026 data: how widespread is SI in the services industry?

The scale and diversity of the services industry show why the SI transformation cannot be described with a single rate. The indicators below are based on official statistics; each measures a different population and a different question.

These data measure different populations. Eurostat covers EU enterprises with 10 or more employees, while the US Census Bureau examines US businesses with a different question and period. Adding the two rates together, averaging them or drawing a trend from one to the other would not be correct.

The shared message is still clear: in the official statistics for 2025–2026, businesses using SI are not the majority. Use is concentrated in large organisations and knowledge-intensive services. An organisation saying it uses SI does not show in which process, at what maturity or with what result.

The Webtures view is that service organisations should add a second question alongside “how many SI tools do we have?”: “which customer or operational problem are we solving, at what cost and with what risk?”

Customer service: dividing work between the assistant and the human agent

Customer service is the most visible application area for SI in the services industry. Request classification, reply suggestions, conversation summaries, information search and the execution of simple transactions are different tasks. Success in one does not mean the others have been validated.

One of the strongest pieces of evidence in this area is a study published in The Quarterly Journal of Economics in 2025. The researchers examined a generative SI-based chat assistant rolled out in stages to 5,172 customer support agents at a company that sells business-process software. Access to the assistant increased the number of issues resolved per hour by 15% on average. Among less experienced and lower-skilled agents, the increase reached 30%.

Among the most experienced and highest-skilled agents, however, there was a small gain in speed and a small decline in quality. The study also reports that customers communicated more politely and asked to speak to a manager less often. These results belong to a model in which the assistant makes suggestions to a human agent; they cannot automatically be generalised to a system that talks to customers directly and independently.

Klarna’s February 2024 announcement shows a different model. The company announced that in its first month the customer assistant handled two-thirds of customer service chats with 2.3 million conversations, that this was equivalent to the workload of 700 full-time agents, and that it estimated a profit improvement of 40 million dollars for 2024. The same announcement also states that customers can speak to a live agent if they prefer.

This is a company statement. The profit figure is an estimate, not a realised result, and it was not measured in an independent controlled experiment. The phrase “agent equivalent” describes workload; it does not prove a change of the same size in headcount.

In a good design, the system should be able to say when it does not know and direct the customer to the appropriate channel. When handing over to a human agent, the conversation summary, the checks already made and the open issue should be passed on. A customer having to explain the same problem again is one of the hidden costs of automation.

Measurement should track first-contact resolution, repeat contacts, incorrect answers, incorrect transactions, human handover and customer satisfaction together. An assistant that closes conversations early may look efficient, but that does not mean it solved the customer’s problem.

The risk of incorrect information: what the assistant says, the organisation says

One of the most concrete SI risks in the services industry is giving customers incorrect conditions, prices or policy information. The Moffatt v. Air Canada decision of 14 February 2024 by the Civil Resolution Tribunal in the Canadian province of British Columbia is a well-documented example of this risk.

The chat assistant on the airline’s website told a customer that a bereavement fare could be applied for retroactively, after the ticket had been bought. The company’s rule, however, did not allow retroactive applications. The airline argued that the assistant should be treated as a separate entity responsible for its own actions.

The decision did not accept this defence. It stated that the assistant was part of the website, that the company was responsible for all the information on its site, and that it made no difference whether the information came from a static page or a chat assistant. The company was ordered to pay the customer a total of 812.02 Canadian dollars.

This is a single small claims decision and not binding precedent in every country. Its management principle is nevertheless clear: information given by an assistant in a customer channel can be read as information provided by the organisation.

For this reason, knowledge management comes before the model. A single approved source, a version and effective date, and an update process with a clear owner should be defined for policies, prices, campaigns and service conditions. The retrieval approach known as RAG can help ground answers in the organisation’s documents. However, finding the right document, the document being current and the answer being consistent with the document should each be tested separately.

Voice assistants and clear disclosure

Real-time voice assistants can be used for high-volume calls such as appointments, reservations, order status and fault reports. The voice channel is faster than the written channel but less tolerant of errors. The customer cannot reread the answer; a misheard date or amount can turn into a transaction without anyone noticing.

Disclosure obligations also matter in this channel. According to the European Commission’s explanation of Article 50, under the EU AI Act people must be informed that they are interacting with an SI system when systems interact with natural persons, unless this is obvious. These transparency obligations have applied since 2 August 2026.

In the US, the Federal Communications Commission (FCC) announced on 8 February 2024 its ruling that SI-generated voices count as “artificial” voices under the Telephone Consumer Protection Act (TCPA). As a result, the consent rules that apply to calls using artificial or prerecorded voices now also apply to SI-generated voices.

The design should include not using voice as the sole method of identity verification, additional verification for transactions such as payments and account changes, clear disclosure of recording and retention conditions, and an easy route to a human agent. Customers should also be told in plain terms which transactions the voice assistant cannot carry out.

Professional services: the verification burden stays with the expert

Professional services such as law, accounting and engineering are among the areas that adopt SI faster because of their knowledge-intensive nature. In Eurostat data, the 2025 use rate in professional, scientific and technical activities is 40.4%. Among these enterprises that use SI, 35.5% use it for business administration processes and 30.9% for accounting, controlling or finance management.

The core risk in this area is a fluent but incorrect output reaching a client or an official body under an expert’s signature. The High Court of England and Wales, in its Ayinde decision of 6 June 2025, stressed that freely available generative SI tools are not capable of conducting reliable legal research. According to the decision, these tools can produce coherent and plausible-looking answers that are entirely incorrect and can cite sources that do not exist.

The same decision states that SI is a tool that carries risks as well as opportunities, and that its use must take place with an appropriate degree of oversight that ensures compliance with professional and ethical standards. This principle is not limited to law. It applies to every service team that prepares reports, proposals, audit notes or technical assessments.

In practice, citing sources should be mandatory, references should be checked in the primary source and the tool in which client data is processed should be recorded. The responsibility of the expert who signs the output should be clear. The speed of preparing a draft should be assessed after the time spent on verification and correction has been deducted.

Operations, hospitality and field services

In service operations, SI can be used for background tasks such as workload and staff planning, appointment and route planning, document processing, and stock and demand forecasting. The value in these areas is often invisible to the customer: the right capacity being ready at the right time, a missing document being caught early or an unnecessary visit being avoided.

Eurostat data show the differences between sub-sectors. In transportation and storage, the 2025 use rate is low among the services sub-sectors at 11.2%, while 19.1% of the enterprises in this sector that use SI apply it for logistics. In accommodation, the use rate is 21.3%, and 58.8% of accommodation enterprises that use SI use it for marketing or sales.

This distribution suggests that in accommodation businesses, SI first entered content and sales processes that touch the customer, while operational use has remained more limited. For scenarios in which product, stock and catalogue data are decisive, our retail and e-commerce guide offers a detailed framework.

Using a general statement such as “it reduces route costs by this much” or “it improves demand forecasting by this much” is not correct. The result depends on the structure of the service network, the maturity of existing planning, data quality and the operating model. Moving from a prediction-based recommendation to automated action is also a separate authority decision.

Agentic SI: how much authority should a service agent have?

Agent-based SI can carry out multi-step tasks by using tools towards a goal. In the services industry this can mean actions such as rescheduling an appointment, starting a refund, requesting a missing document or opening a case. Open-source standards such as the Model Context Protocol make it easier for SI applications to connect to external systems. Easier connection, however, does not remove the question of authority.

OWASP’s 2025 risk list for LLM applications traces the risk of “excessive agency” to three sources: excessive functionality, excessive permissions and excessive autonomy. The recommended measures include minimising tools and permissions, requiring user approval for high-impact actions and leaving authorisation to downstream systems rather than to the model.

Prompt injection, first on the same list, is the unintended alteration of a model’s behaviour by user input or by content in external sources such as web pages and files. A message from a customer or a document the agent reads should not be able to change the agent’s access rules. Transaction rules should not be left solely to the model’s instructions; they should be enforced in the application layer.

When a transaction fails, the agent may retry; but a retry should not lead to the same refund being made twice. It should be possible to see where a partially completed workflow stopped. Human approval should also be a control in which the decision can genuinely be assessed; an approval click that becomes automatic under pressure does not provide effective oversight.

Customers researching, booking or buying services with their own agents is another new area of preparation. Presenting service scope, prices, availability and cancellation conditions in a form machines can read correctly is the foundation of agentic commerce readiness.

Technical architecture: how should services SI be built?

A reliable structure manages the model’s access to information and its ability to act in separate layers. Service conditions are a source of information; order, appointment and account status are current data that must come from the authoritative system.

When the model version changes, not only answer quality but also tool use, error behaviour and cost should be retested. In agent loops, a single customer request can turn into several model calls. Cost per request should therefore be tracked together with cost per correctly resolved request.

How the service channel will continue when the SI system is not working should also be decided in advance. The human work queue, capacity limits and customer communication should be tested. Even if the model responds, working incorrectly or excessively slowly should be treated as an operational problem.

In supplier assessment, data use, subcontractors, service continuity, change notification, access to records and the exit plan should be examined as closely as model quality. How the knowledge base, conversation records and workflow rules will be moved when the contract ends should be clear from the start.

Regulation: which distinctions matter when preparing for 2027?

Global service companies cannot work to a single SI timeline. The purpose of use, the country, the organisation’s role as provider or deployer, and existing consumer, data protection and sector rules should be assessed together. For organisations operating in the EU, the European Commission’s AI Act explanations show the following timeline:

The AI Omnibus, published as Regulation 2026/1744, entered into force on 27 July 2026 and postponed the application dates for high-risk systems. This does not mean that all obligations have been put on hold. The transparency rules have applied since 2 August 2026, and existing consumer and data protection rules remain in force. No service application should be classified against a single date.

Annex III lists uses such as recruitment and candidate evaluation, decisions concerning employees and performance monitoring, the creditworthiness of natural persons, and risk assessment and pricing in life and health insurance. A customer support assistant and a recruitment tool that screens applications are not assessed in the same risk class. For distinctions specific to credit assessment, our finance and banking guide offers a separate framework.

For a shared risk language in operations outside the EU, the NIST AI Risk Management Framework can be used. Published on 26 January 2023, the framework is intended for voluntary use; on 26 July 2024 NIST published a separate profile for risks specific to generative SI. Guidance of this kind should not be presented as law or as an automatic compliance certificate.

The recommended output for the implementation team is a live inventory that shows each system’s purpose, data types, permissions, target countries, owner and applicable rules. The legal assessment should be carried out on the basis of this concrete description of use.

Employees and the workforce: what does the evidence say?

The evidence on SI’s workforce impact in the services industry does not point in a single direction. Studies using different countries, periods and methods answer different questions. Combining their results into a single “job loss” or “job creation” rate is misleading.

In the customer support study, the largest gain was seen among agents with less experience. The researchers also reported evidence that the assistant supported employee learning. This suggests that in some roles SI can speed up how quickly a new employee becomes proficient; however, it is a result measured at a single company.

The working paper from the Stanford Digital Economy Lab, updated in August 2026, examines the period up to June 2026 using ADP payroll data in the US. The study reports that it finds no evidence of widespread, economy-wide job displacement. In contrast, the employment of workers aged 22–25 in SI-exposed occupations is 19% below the level it would have reached had it kept pace with their less-exposed peers.

According to the same study, this gap arises mainly through reduced hiring of young workers. The authors interpret the findings not as causal estimates but as early, descriptive indicators.

From Denmark, the working paper by Humlum and Vestergaard links the adoption of chat assistants to administrative labour market records. The study finds precise null effects on earnings and recorded working hours and rules out effects larger than 2% two years after the launch of ChatGPT. Neither study is a peer-reviewed journal article, and they rely on data from different countries.

The implication for service organisations is to protect the learning path of entry-level roles. If the tasks through which an agent or analyst develops expert judgement are fully automated, the organisation’s future pool of experts may weaken. Because the EU’s SI literacy provisions have applied since 2 February 2025, employee training is not only a development topic for organisations operating in Europe.

Training should not cover tool use alone. The skills to check sources, spot errors, protect customer data and intervene at the right moment should be developed together.

How should success be measured?

A shared measurement glossary should be prepared for the SI programme. Time savings, cost reduction, revenue impact, customer experience and risk reduction should be calculated separately.

The measurement design should use a control group or a suitable comparison method. All of an improvement should not be attributed to SI without separating out seasonality, request mix and simultaneous process changes.

Capacity gains in particular should be handled with care. A team saving 1,000 hours does not mean that 1,000 hours of pay automatically falls out of the budget. That time can be used to handle more requests, shorten waiting times or give more time to complex cases.

Sample investment calculation: separating capacity value from cash savings

Low, base and high realisation scenarios should be prepared for the investment assessment. If repeat contacts or complaints are rising, a speed gain alone cannot justify scaling.

Implementation plan for the first 90 days

Tasks at the information and preparation levels can be preferred for the first project. Applications such as reply suggestions for agents, conversation summaries or policy search also require privacy and accuracy controls; however, they offer a different starting point from a system that carries out transactions directly.

In shadow operation, system outputs can be assessed before they reach customers. Acceptance and stop criteria should be written before the pilot. It should be clear who will intervene if an incorrect customer record, a critical information error, an unauthorised transaction or customer harm is observed. Good average performance alone is not enough.

2027 roadmap and team structure

Final quarter of 2026: The SI inventory is compiled. Tools employees use on their own initiative, data sharing and supplier dependencies are made visible. Article 50 disclosure design is reviewed for assistants that interact with customers.

First quarter of 2027: Evidence is produced on the selected request types. Quality and commercial measures are fixed. The approved source of information, version tracking and content ownership are turned into an organisational process.

Second quarter of 2027: Successful workflows are expanded gradually. Human handover, transaction records and outage scenarios are tested under real operating conditions.

Third quarter of 2027: Uses that may fall within Annex III, such as recruitment or employee evaluation, are reassessed against the 2 December 2027 timeline with the organisation’s legal and compliance team.

Final quarter of 2027: The application portfolio is reviewed against proven value. The 2 August 2028 timeline is confirmed for services linked to products within the scope of Annex I, and the next year’s budget is built accordingly.

The business unit should be responsible for the outcome, the technology team for reliable operation, the customer experience team for information accuracy and handover quality, and the risk and compliance teams for the relevant controls. “The SI team is responsible” is not an adequate role description on its own.

SI visibility for service brands

When researching service providers, customers may also turn to SI-assisted search and assistants. Even if the organisation’s name appears, visibility can create commercial and reputational problems if service scope, prices, opening hours or cancellation conditions are conveyed incorrectly.

Google’s official guidance states that there are no additional requirements or special optimisations for appearing in AI Overviews and AI Mode; what matters is that the page is indexed and eligible to be shown with a snippet. The same guidance recommends that structured data match the visible text on the page and explains that these features may use a “query fan-out” technique that issues multiple related searches across subtopics.

Service content should make the service scope, region, pricing conditions, update date and official service name clear. Old campaigns and current offers should be kept separate. As the Air Canada decision shows, contradictions between pieces of information across the organisation’s own channels are a source of risk for both customers and machines.

Measurement can be based on representative questions for different countries, languages and service needs. Brand mentions, source citations and information accuracy should be tracked separately. The aim of this work is not to mislead models or to guarantee a definite recommendation; it is to make the organisation’s verifiable information accessible, consistent and current.

Frequently asked questions

Where does SI create the most value in the services industry?

There is no single universal ranking. Customer support, employee access to information, document-heavy processes and operations planning are strong candidates. Priorities should be set according to the organisation’s data quality, request volume and measurable problem.

How many businesses use SI?

According to Eurostat, 20.0% of EU enterprises with 10 or more employees used at least one SI technology in 2025. In the US Census Bureau’s data for December 2025 to May 2026, the rate is in the 17–20% range. Because the two data sets rely on different questions and populations, they should not be combined.

Will SI replace human agents in customer service?

Some requests can be automated and the way agents work can change. The strong evidence comes from models in which the assistant supports the agent. For complex, sensitive or exceptional requests, human handover should remain part of the design.

Who is responsible if a chat assistant gives incorrect information?

The legal outcome varies by country and case. In the Moffatt v. Air Canada decision, the company tried to present the assistant as a separate entity; the decision stated that it made no difference whether the information came from a static page or the assistant. Organisations should manage assistant information as their own information.

What is the EU AI Act date for high-risk systems?

According to the European Commission, the rules for high-risk systems within the scope of Annex III will apply from 2 December 2027, and those for Annex I systems embedded in products from 2 August 2028. The Article 50 transparency obligations have applied since 2 August 2026.

Does using RAG completely prevent incorrect answers?

No. The wrong document may be retrieved, the source may be out of date or the model may misinterpret the source. The retrieval and answer generation stages should be assessed separately.

What is the most important outcome of preparing for 2027?

Building an operable structure that shows which system accesses which information, which transactions it can carry out, who monitors it and what value it produces.

The Webtures approach: accurate information, controlled agents, measurable service

The SI transformation in the services industry requires addressing customer trust and operational efficiency together. Representing the brand accurately in SI interfaces, giving customers consistent information across every channel and improving internal processes in a measurable way require a shared strategy.

The approach Webtures recommends is to assess current visibility and the customer journey, identify inconsistencies in information and content, and define priority transformation areas with concrete goals. Technical implementation, legal assessment and regulatory compliance should be carried out with the organisation’s relevant teams and experts, with responsibilities kept clear.

For service organisations preparing for 2027, a strong start means producing evidence within a limited scope and growing the structure that works in a controlled way.

Let’s assess your service brand’s SI visibility and transformation priorities together.

Get in touch with Webtures →