AI in IT Service Management
Artificial intelligence can speed up detection, triage, and knowledge work in IT service management, and it is safe to use only where a person remains accountable, answers are grounded in your own material, and the risks of error, bias, and data exposure are managed.
Why this matters
Service management involves a great deal of language and pattern recognition: reading tickets, summarizing incidents, searching knowledge, drafting replies, spotting unusual signals in monitoring data. These are the tasks at which recent AI systems appear most capable, and vendors have rushed to add them to service desks and operations tools. Teams will be asked, if they have not been already, whether to use them.
The opportunity is real, and so are the dangers. AI systems can state falsehoods with confidence, reflect biases in their training and in the data they are given, expose confidential information if fed carelessly, and lead people to trust output they have not checked. Several widely reported cases show the cost of misplaced trust. A service organization that adopts AI without thought can damage its customers and its own reputation.
By 2026, AI has also entered the framework. PeopleCert's public certification page describes the new ITIL as "Practical and AI-native by design" and lists an ITIL AI Governance (Version 5) module that covers governing AI "responsibly and at scale" (see Topic 2.2 — ITIL). This guide could not verify from the owner's publications what ITIL 5 says about AI in detail, so this topic does not describe the framework's position beyond those statements. It draws on the Wikipedia articles on AIOps and on AI hallucination, on NIST's AI Risk Management Framework page, and on practical reasoning, and it labels practice as such.
Core concepts
Where AI helps in service management
Most uses fall into a few groups. The table shows examples and the kind of risk each carries, as this guide's reading.
| Use | What the AI does | Main risks |
|---|---|---|
| Detection and correlation (AIOps) | Finds anomalies, groups related events, suggests likely causes | False alarms, missed events, opaque reasoning |
| Service desk assistance | Suggests answers, summarizes tickets, drafts replies, routes requests | Wrong or invented answers, tone, privacy |
| Knowledge work | Drafts and improves articles, searches knowledge | Fabricated content, stale sources |
| Incident and problem analysis | Summarizes timelines, suggests hypotheses, drafts reviews | Plausible but unfounded causes |
| Change and release support | Reviews changes, drafts notes, suggests risk | Overconfidence, missed context |
| Code and automation | Writes scripts and configuration | Insecure or incorrect code, hidden dependencies |
| Analysis and reporting | Summarizes data, drafts reports | Wrong calculations, invented figures |
AIOps
The Wikipedia article describes AIOps as using AI, machine learning, and big data analytics to automate and improve data center management, aiming to detect, diagnose, and resolve IT issues more efficiently than traditional methods. It says Gartner first defined AIOps in 2016. The article lists capabilities including event correlation, anomaly detection, root cause analysis, causality determination, predictive analytics, incident management, automated remediation, performance prediction, resource allocation, and capacity management. It says AIOps draws on observational data, such as logs and monitoring, and interaction data, such as tickets, events, and incident records.
The article claims benefits, such as mean time to detect improving by 15 to 20 percent and critical incidents dropping by over 50 percent, but notes these come from a single consultancy article and are not independently verified in the text. It states no criticisms, failure cases, or limitations. Treat vendor claims accordingly, and test any tool against your own incidents (see Topic 3.10 — Monitoring and Event Management).
Generative AI and hallucination
The Wikipedia article on hallucination in AI defines it as a response that contains false or misleading information presented as fact, and notes that the term is criticized for anthropomorphizing software. It gives causes: pre-training on next-word prediction rewards guessing when the model lacks information, datasets may be incomplete or unrepresentative, and OpenAI researchers argue that training and evaluation reward guessing over acknowledging uncertainty. Examples it records include:
- Mata v. Avianca (May 2023): a lawyer submitted six fake cases generated by ChatGPT; the judge imposed a fine of US$5,000 on 23 June 2023.
- Moffatt v. Air Canada (February 2024): a tribunal held an airline to a bereavement policy that its chatbot had invented.
- Reports by a consultancy that contained false citations, and a study that found 69 of 178 GPT-3 references had incorrect or nonexistent identifiers.
These are the consequences of trusting unverified output: legal liability, financial exposure, and damaged trust. In service management, a chatbot that invents a policy or a procedure creates the same class of risk.
Mitigation: grounding and checks
The article describes retrieval-augmented generation (RAG) as letting the model access external sources before answering. It also lists data-related methods, model and inference methods, and proposals such as guardrail software and confidence scoring, and notes that no source it gives quantifies how much any of these reduces hallucination. For a service team, practical mitigations are:
- Ground answers in your own approved material, such as the knowledge base, and show the source.
- Keep a person in the loop for anything that affects a customer, a commitment, or production.
- Ask for what the AI is unsure of, and treat unsupported statements as unverified.
- Test regularly on real cases, including cases where the right answer is "I don't know."
Risk management frameworks
The National Institute of Standards and Technology (NIST) published the AI Risk Management Framework (AI RMF 1.0) on 26 January 2023, which it describes as intended for voluntary use, to help manage risks from AI to individuals, organizations, and society and to improve the ability to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI products, services, and systems. NIST's page names four functions, Govern, Map, Measure, and Manage, but does not describe them. It also released a Generative Artificial Intelligence Profile (NIST-AI-600-1) on 26 July 2024 to help organizations identify risks unique to generative AI, and says the framework is being revised. Read the framework itself for detail. Other rules about AI, such as laws and standards, differ by country and sector; ask your legal and compliance functions, and obtain the text (see Topic 5.7 — Governance, Risk, and Compliance).
Data protection and privacy
Everything given to an AI service may be stored, used for training, or exposed, depending on the service and the terms. Tickets, logs, and chat transcripts often contain personal data, credentials, and confidential business information. Before using any AI tool with real data, find out where data goes, who can see it, how long it is kept, and whether it is used to train models, and apply your information security policy (see Topic 3.13 — Information Security Management). Remove or replace personal and secret data before pasting it into a tool you have not approved.
Accountability and transparency
A person or a team must remain accountable for decisions an AI contributes to. When AI drafts a reply, the sender is responsible. When AI suggests a root cause, the investigator is responsible for testing it. Tell customers when they are talking to an automated system where the law, a contract, or fairness requires it, and give a route to a person. Keep records of what the AI was asked and what it produced for significant decisions, as you would for other tools (see Topic 5.1 — Agile Service Management for evidence).
Bias and fairness
AI systems learn from data that reflects past decisions and gaps. A triage assistant trained on historic tickets may treat some groups of users or some kinds of problems as lower priority. Check outputs across groups, involve people with different perspectives in testing, and watch for patterns. The sources read for this guide do not detail bias, so this is a general caution, not a finding.
Measuring AI
Measure the quality and the effect of any AI step, not just its volume: how often humans override it, how often it is wrong, how long tasks take with and without it, and what users say (see Topic 5.6 — Metrics and Measurement). Be cautious about vendor figures; the AIOps article's own benefit claims rest on one source.
Skills and prompts
The practical skill of working with AI includes writing clear instructions, supplying the right material, asking for uncertainty, and verifying output. Topic 6.3 — AI Skills and Prompts covers this in detail.
Limits of the sources
The AIOps article contains no criticisms and its benefit figures come from one source; the hallucination article is flagged as needing an update and was read in part; NIST's page does not describe the framework's functions; and PeopleCert's page does not explain what ITIL 5 says about AI. This topic therefore gives a cautious account and leaves specifics to primary sources.
Best practices
- Start with a low-risk use and a clear question. Pick a task where errors are cheap and easy to catch, such as drafting a summary for internal use, and define what success means.
- Keep a person accountable. Name who is responsible for what the AI contributes to, and require review for anything that reaches a customer, changes production, or makes a commitment.
- Ground answers in approved material. Where you use AI for answers, connect it to your own knowledge base and show the source. Make it say "I don't know" when it lacks one.
- Protect data. Decide what data may go to which tool, remove personal and secret data, and use approved tools with terms you have read. Involve security and legal staff early.
- Test with real cases. Collect a set of real tickets and incidents with known answers, run the AI on them, and record where it fails. Repeat after any change of tool or settings.
- Verify claims and references. Check names, numbers, citations, and procedures against sources, as the cases of invented court citations show. Never accept references you have not seen.
- Be honest with customers. Say when an automated system is answering where it matters, and give an easy route to a person.
- Watch for bias. Review outputs across user groups and kinds of problems, and involve people who can spot gaps. Adjust or stop use where harm appears.
- Measure quality and effect. Track override rates, error rates, time saved, and user views. Compare against the process without AI, and be skeptical of vendor figures.
- Keep records. For significant decisions, keep what was asked and produced, who reviewed it, and what was decided. This supports accountability and audit.
- Use a risk framework. Adopt a recognized approach, such as NIST's AI RMF or an organizational AI policy, to structure decisions about where AI is allowed. Treat it as a guide to thinking.
- Train people. Teach staff how the tools work, where they fail, and what is allowed. Update the training as tools and rules change.
Questions to discuss with your team
Where are we already using AI, officially or not?
Ask honestly. Staff often use public tools with real tickets. The aim is not to punish but to understand. List the uses and the data involved. Discuss which are acceptable and which need a safer route. Agree on a short policy.
What task would we most like help with, and what happens if the AI gets it wrong?
Pick a task and walk through the failure case. If the answer is an apology email, the risk is small; if it is a wrong instruction to a customer, the risk is higher. This helps rank uses by risk. Choose a first pilot with limited harm.
How would we know that an AI answer is wrong?
Ask who checks answers and how. Common answers: nobody, or the person who asked. Discuss sources, tests, and second reviews. Agree on a minimum check for each kind of use.
What data are we comfortable putting into an AI tool, and what are we not?
Collect the types: tickets, logs, customer names, credentials, contracts. Mark each as acceptable, acceptable after removal of sensitive parts, or never. Check with security and legal. Write it as a one-page guide.
Who is accountable when an AI-assisted decision goes wrong?
Use a past example of a tool-assisted decision. Ask who answered. Agree that a named person remains accountable and what evidence they should keep. Note any gap where nobody would.
What would convince us that a tool is worth keeping?
Ask the team to describe the evidence they would accept: time saved with equal or better quality, fewer errors, user satisfaction. Discuss how to gather it in a trial. Agree on a review date.
In practice: an agile ITIL example
A national tax authority's IT service desk handles about 1,200 contacts a week from staff, many on routine questions. The desk manager proposes an AI assistant to suggest answers from the knowledge base and draft replies. The authority's security officer is concerned about personal data in tickets and about wrong answers reaching staff.
They agree on a limited pilot. The assistant works only from approved knowledge articles, shows which article each suggestion comes from, and says when it finds nothing relevant. Analysts review every suggested reply before sending, and the tool cannot send on its own. Tickets are scanned for personal data and credentials, which are removed before the assistant sees them. A test set of 200 past tickets with known answers is used before the pilot, and the team records where the assistant fails. The desk manager and the security officer agree on measures: how often analysts edit or reject suggestions, time to first response, and a short satisfaction question for staff. The pilot runs for three months, and the manager reports to a joint review that includes an analyst and a staff representative.
At the end, analysts accept most suggestions for common questions and response time for these falls. The test set shows that the assistant gives a confident but wrong answer for questions on a recently changed policy, because the knowledge article was out of date. Some problems remain. Two analysts started to send suggestions with little checking when busy, a staff member with a non-standard question received a bland reply that missed the point, and the vendor's terms about data retention were unclear. The manager adds a rule that policy changes must update the knowledge article before they take effect, uses a sample check of sent replies by a senior analyst, and asks procurement and the security officer to clarify the terms before extending the pilot.
Four perspectives
Product and delivery teams
Delivery teams may add AI features to products and use AI to write code and tests. Check generated code for correctness and security, keep secrets out of prompts, and be clear to users when AI is involved. Build monitoring for AI behavior into your service.
Service and operations teams
Operations teams will meet AI in monitoring, desk tools, and analysis. Treat AI suggestions as hypotheses, test them, and keep records. Help define what data may be used, and give feedback to improve grounding material.
Leadership and governance
Leaders set policy, risk appetite, and accountability for AI. Ask for a use inventory, a risk framework, and evidence from pilots, not vendor claims. Make sure legal, security, and privacy staff are involved from the start.
Customers and users
Customers and users should be told when they are dealing with an automated system where it matters, and have a route to a person. Report wrong or strange answers, since they are evidence. Ask providers how they protect your data and check AI output.
Common failure modes
- Trusting unverified output. Fabricated facts reach customers. Instead, ground answers in approved material and verify before sending.
- Shadow AI. Staff paste real data into public tools. Instead, provide approved tools and a clear data policy.
- No accountable person. Nobody answers for AI-assisted decisions. Instead, name an owner and require review.
- Believing vendor figures. Benefit claims go unchecked. Instead, test on your own cases and measure.
- Stale grounding. The knowledge base is wrong and the AI repeats it. Instead, keep articles current and owned.
- Hidden bias. Some users or issues are treated worse. Instead, test across groups and review outputs.
- Automating before understanding. AI acts on production without checks. Instead, begin with suggestion and review, and move to action only with strong safeguards.
- No records. Decisions cannot be audited. Instead, keep logs of what was asked, produced, and decided.
Maturity model
| Dimension | Initiate | Develop | Standardize | Manage | Orchestrate |
|---|---|---|---|---|---|
| Policy and accountability | Use is ad hoc and unrecorded | Some teams have rules | A written policy, inventory of uses, and named owners | Policy compliance and incidents are reviewed | Policy is aligned across teams, suppliers, and regulators |
| Quality and grounding | Output is trusted by default | Some checks on important uses | Answers are grounded in approved sources and tested on real cases | Error and override rates are measured and acted on | Grounding and tests are shared and improved across teams |
| Data protection and fairness | Data are pasted freely | Some data rules exist | Data handling rules and bias checks are routine | Data incidents and fairness findings are tracked and fixed | Data and fairness practices are coordinated across the organization and suppliers |
| Value and learning | Value is assumed | Some trials are run | Pilots have agreed measures and decisions | Benefits and costs are measured against baselines | Learning from AI use spreads across services continuously |
Checklist
- We have an inventory of where AI is used in or around our services, including unofficial use.
- A named person is accountable for each AI-assisted use.
- A written policy says what data may go into which tools.
- Personal data and secrets are removed or masked before use with tools not approved for them.
- Answers shown to customers or used in decisions are grounded in approved material and reviewed.
- We have a test set of real cases and rerun it after changes.
- Customers are told when an automated system answers, where it matters, and have a route to a person.
- We measure override rate, error rate, time saved, and user views.
- Records are kept for significant AI-assisted decisions.
- Legal, security, and privacy staff have reviewed our use and the tool's terms.
AI prompts
Purpose: draft an AI use inventory from what people tell you.
Help me structure an inventory of AI use in our service organization.
What people have told us they use AI for (paste notes; remove names): [paste].
Tools mentioned: [list].
Types of data involved, as far as we know: [list].
Produce a table with: use, tool, data involved, who is accountable (use "unknown" if not given), risk if wrong (low, medium, high with a one-sentence reason from my notes), and one question to ask. Do not guess about tools or terms I did not mention. Ask clarifying questions first, and list what you are unsure of.
Purpose: design a test set for an AI assistant.
Design a test set for an AI assistant that suggests answers to service desk questions.
The knowledge sources it will use: [describe].
Types of question we get: [list with rough share if you know].
Cases where the right answer is "not in the knowledge base" or "escalate": [describe].
Propose a structure for 100-200 test cases (categories, how to choose real examples, how to record the correct answer and its source), how to score answers (correct, partly correct, wrong, confidently wrong, correctly declined), and how to repeat after changes. Do not invent test content; tell me what to collect. Ask clarifying questions first, and list what you are unsure of.
Purpose: write a one-page AI use policy for a team.
Draft a one-page AI use policy for [team].
Tools approved: [list, or write "none yet"].
Data we handle (types and sensitivity): [list].
Rules we must follow, with source text from legal, security, or contracts: [paste, or write "none known"].
Roles: [list].
Write plain-language sections: what AI may be used for, what data may and may not be put into tools, what must be checked by a person before use, who is accountable, how to record significant uses, how to report problems. Mark every item that is an assumption. Do not state that any law requires something unless I pasted its text. Ask clarifying questions first, and list what you are unsure of.
Purpose: review an AI-generated answer before it goes to a user.
Here is a draft answer that an AI tool produced for a user question.
The user's question: [paste, with personal data removed].
The knowledge article(s) the answer should be based on: [paste].
The draft answer: [paste].
Check the draft against the article only. List statements that are supported by the article (quote the part), statements that are not supported, anything missing that the user would need, and the tone. Suggest a corrected version that uses only the article. Do not add facts. List what you are unsure of.
AI skills
- prompt-engineering: use to write clear prompts that supply the right material and ask for uncertainty.
- machine-learning-model-card: use to document what a model is for, how it was tested, and its limits.
- data-governance: use to decide what data may be used with which tools.
- legal-data-privacy-compliance: use to prepare questions for legal and privacy staff about AI use.
- risk-register: use to record AI-related risks, owners, and treatments.
Key sources
- AIOps (Wikipedia): a short description of AI for IT operations, its capabilities, and its claimed benefits, with no criticisms.
- Hallucination (artificial intelligence) (Wikipedia): causes, examples, and mitigation ideas for false output.
- AI Risk Management Framework (NIST): the voluntary framework and its generative AI profile; read the framework itself for the details.
- ITIL certifications (PeopleCert): the owner's public statements about the new ITIL and its AI governance module.
References
- AIOps — Wikipedia — https://en.wikipedia.org/wiki/AIOps
- Hallucination (artificial intelligence) — Wikipedia — https://en.wikipedia.org/wiki/Hallucination_(artificial_intelligence)
- AI Risk Management Framework — NIST — https://www.nist.gov/itl/ai-risk-management-framework
- ITIL certifications — PeopleCert — https://www.peoplecert.org/browse-certifications/it-governance-and-service-management/ITIL-1
- Retrieval-augmented generation — Wikipedia — https://en.wikipedia.org/wiki/Retrieval-augmented_generation
- Artificial intelligence — Wikipedia — https://en.wikipedia.org/wiki/Artificial_intelligence
- Large language model — Wikipedia — https://en.wikipedia.org/wiki/Large_language_model
- Algorithmic bias — Wikipedia — https://en.wikipedia.org/wiki/Algorithmic_bias
- IT service management — Wikipedia — https://en.wikipedia.org/wiki/IT_service_management