The Benefits and Drawbacks of Artificial Intelligence: When Does It Deliver Real Business Value?
The same AI tool can free up hours for one team, while for another it generates more review work, errors and cost than it saves. It’s no accident that research findings range so widely, from a 40% time saving to a 19% slowdown. The benefits and drawbacks of artificial intelligence depend heavily on the task it’s used for, the data involved and the degree of human oversight in place.
But how can a leader know whether AI is genuinely creating value, rather than simply bringing new costs and risks into the organisation? In this article, we set out the benefits confirmed by research, the practical limitations, and the data protection and information security dangers, before finally looking at which processes are worth introducing AI into.
What Are the Benefits and Drawbacks of Artificial Intelligence, in Brief?
Artificial intelligence can speed up work and take the load off a team, but it also brings new review, data protection and information security tasks. The question for leaders is therefore not whether the company should use AI, but in which processes it delivers more benefit than the cost and risk involved in adopting and overseeing it.
| Benefit | The related drawback or risk |
|---|---|
| Faster drafting, translation and summarising | Factual errors; mandatory review |
| Greater capacity with the same team | Loss of competence and extra review work |
| More consistent first drafts for routine tasks | Homogenisation and a weaker brand voice |
| Faster data analysis and decision preparation | Biased suggestions; a decision trail that’s hard to audit |
| Quick to trial, low barrier to entry | Shadow AI, data leakage and scaling costs |
| New services and improved customer experience | Compliance, supplier and information security risks |
The weight of the benefits and risks differs by sector and by process. The benefits and drawbacks of AI can only be assessed realistically once the current level of effort, the cost of errors, the sensitivity of the data involved and the time required for human review are known.
What Are the Benefits of Artificial Intelligence in Business Operations?
The benefits of artificial intelligence show up mainly in tasks that involve processing repetitive, high-volume text or structured data, where any errors can be spotted and corrected in time. Examples include writing summaries, translation or organising information.
In the K&H Innovation Index survey for the first half of 2026, 58% of the Hungarian companies surveyed, all with annual revenue of at least HUF 300 million, used some form of AI solution. The most common application was translation at 39%, followed by text processing at 33% and marketing communication support at 22%. The result is based on a survey of 360 medium-sized and large companies, so it can’t automatically be projected onto every Hungarian business.
In business practice, the following benefits appear most often:
- Shorter turnaround time: A draft or summary can be produced faster, leaving more time for editing and decision-making.
- Freeing up capacity: Time saved on routine tasks can be redirected to higher value-added work.
- Faster onboarding: Less experienced colleagues can get support in producing first drafts and templates.
- Faster processing of large documents: AI can speed up the first-pass processing and organising of long materials, though the conclusions still need human review.
- Reducing language barriers: Communication with foreign-language customers and processing of supplier materials can speed up.
- Flexible scaling: The solution can be extended to more staff and processes, but licensing, integration, training and oversight costs typically grow alongside usage.
According to Eurostat’s 2025 survey using the same methodology, 20.0% of EU enterprises with at least 10 employees, 10.4% in Hungary and 42.0% in Denmark, used at least one AI technology. This does point to Hungary lagging behind the EU average, but the survey uses a different company base and methodology from the K&H research, so the two sets of percentages can’t be compared directly.
How Much Does Artificial Intelligence Increase Productivity?
There’s no single general percentage that shows how much artificial intelligence increases productivity. The actual result depends on the task, the tool used, the employee’s experience, how much time review takes, and whether speed, the number of tasks completed or the quality of the work is what’s being measured. As a result, no single figure that applies to every company can be derived from research that uses different methodologies.
Some frequently cited results:
| Study and task | Sample and metric | Result | Important caveat |
|---|---|---|---|
| Software development (Cui et al.) | 4,867 developers; completed tasks | +26.08% | Three field experiments; results varied considerably between them, with less experienced developers seeing the larger gains. |
| Customer service (Brynjolfsson et al.) | 5,000+ customer service agents; issues resolved per hour | close to +14% | Conducted in a single company setting; less experienced and lower-performing agents improved the most. |
| Professional writing (Noy and Zhang) | 453 graduate professionals; time and external quality assessment | 40% shorter completion time, +18% quality | The study was limited to mid-level professional writing tasks, so the results can’t be generalised to all types of work. |
| Open-source development (METR) | 16 experienced developers, 246 tasks; completion time | 19% longer completion time | Conducted on well-known codebases using tools available in early 2025; despite the measured slowdown, participants felt they had worked 20% faster. |
The most important lesson for leaders is that work that feels faster doesn’t necessarily mean an actual improvement in performance. The return on AI can only be established if the process’s time requirement, quality, error rate and total cost are recorded before it’s introduced. Employees’ impressions alone don’t provide a reliable basis for the assessment.
What Are the Drawbacks of Artificial Intelligence?
The drawbacks can arise partly from how generative models work and partly from how they’re adopted within an organisation. Large language models produce answers on a probabilistic basis, so fluent phrasing is no proof that the output is factual, complete or appropriate for the decision at hand.
- Hallucination: The model can confidently state an untrue claim, figure or even a non-existent reference.
- Extra review work: Part of the gain from a faster first draft can be offset by the professional and factual review it needs.
- Bias: Biases and imbalances present in the data used to train the model can influence its answers. This can put certain groups at an unjustified advantage or disadvantage, which is especially risky in recruitment and performance evaluation.
- Loss of competence: If staff routinely hand professional judgement and problem-solving over to AI, the organisation’s expertise, independent thinking and judgement can weaken over time.
- Homogenisation: Without a clear brand voice and consistent editing, AI-produced content can easily become generic and lose the company’s distinctive tone.
- Vendor dependency: A change in a provider’s pricing, terms, data handling or model can force a costly switch.
- Environmental and cost burden: Energy demand, usage fees, integration, training and review can all increase total cost.
- Compliance burden: Depending on the area of use, the company must meet data protection, transparency and documentation requirements, and ensure adequate human oversight.
Among the drawbacks of AI, a lack of measurement is especially dangerous in practice. If an organisation doesn’t record its starting position, it later has no way of telling whether faster work actually produced real savings, or simply shifted the effort onto other stages of the process.
What Are the Dangers of Artificial Intelligence from an Information Security Perspective?
The dangers of AI in a corporate environment come from two main directions: unapproved tool use within the organisation, and attackers using AI to speed up and make more convincing their fraud or intrusion attempts. The first can be significantly reduced with the right rules and choice of tools.
The most important risk categories:
- Shadow AI: Staff use unapproved accounts or applications and upload confidential documents, customer data or source code.
- Data leakage through input: A provider’s data retention, model development and transfer terms may not meet the company’s requirements. There can be differences between free and business plans, but the specific contract and settings always need to be checked.
- Hijacking through malicious instructions (prompt injection): an instruction hidden in a document, webpage or email can change how the AI system behaves. The risk is especially high if the tool has access to company data or can carry out actions on its own.
- Model and vendor risk: An incident at the provider, a faulty update or a change in terms can also affect company processes and data.
- Incomplete auditability: Without proper logging, it’s difficult to reconstruct who entered what data, what output they received, and how it was used.
How Do Attackers Use Artificial Intelligence?
Attackers mainly use AI to speed up social engineering and make it more convincing. They use it to create personalised phishing messages, cloned voices, fake profiles and deepfake videos. AI mainly lowers the cost of well-known fraud methods and increases how easily they can be scaled.
A few figures illustrate the scale of the risk:
- AI-related complaints: In 2025, the FBI Internet Crime Complaint Center received 22,364 complaints that referenced AI-related information; the adjusted reported losses exceeded $893 million. This doesn’t mean every case is proven to be fraud committed solely using AI.
- Voice phishing (vishing): According to Mandiant’s M-Trends 2026, vishing was the initial attack vector in 11% of the intrusions investigated in 2025. According to Google Cloud’s summary, it was the initial infection vector in 23% of cloud-related incidents. These statistics measure vishing as such, not specifically the use of AI-generated voice.
- Deepfake video call: In 2024, a finance employee at Arup’s Hong Kong office was persuaded to transfer around $25.6 million during a video conference in which the other participants were deceptive, manipulated representations.
The key lesson for leaders is that approval can no longer be based on the apparent authenticity of video or voice alone. Every unusual financial instruction should be confirmed by calling back on a known contact number and requiring approval from at least two employees.
What Data Protection Risks Come with Using Artificial Intelligence?
Data protection risk arises when personal data enters an AI system for which the organisation hasn’t clarified the legal basis, purpose limitation, data minimisation, notification to data subjects, data processor terms or international data transfers. Trade secrets aren’t a GDPR category, but uploading them can still have serious contractual and information security consequences.
Three common sources of error:
- An unapproved account used with business data. The data handling, retention and contractual terms of personal and free plans often differ from business solutions. A business subscription alone is no guarantee, though: the specific contract, settings, sub-processors and data transfers all need to be checked.
- No data classification. If it isn’t clearly set out what personal, confidential or trade-secret data may be entered into the AI tool, an employee can make a mistake in good faith.
- Inadequate notice. Under Article 50 of the AI Act, providers of AI systems that communicate directly with people must ensure that those affected know they are communicating with an AI system, unless this is obvious from the circumstances.
According to the European Data Protection Board, whether an AI model can be considered anonymous requires a case-by-case assessment. The data controller must also demonstrate the appropriate legal basis and that the GDPR principles are met.
The regulatory environment changed in 2026: the Digital Omnibus was published in the EU Official Journal on 24 July 2026 and entered into force on 27 July. Application of the obligations for high-risk AI systems under Annex III has been pushed back to 2 December 2027, and for high-risk systems embedded in regulated products, to 2 August 2028.
The postponement doesn’t mean every AI Act obligation has been delayed. The AI literacy requirement has already applied since 2 February 2025, and the Article 50 transparency rules since 2 August 2026. A general AI record-keeping obligation doesn’t automatically apply to every organisation and every use of AI, though; the documentation requirement always depends on the specific role and risk category involved.
Is Artificial Intelligence Taking Away Jobs?
Based on the current evidence, generative AI is expected to transform more roles than it eliminates outright, because most occupations consist of tasks, some of which still require human input. That doesn’t mean every group of workers is affected equally, however, or that some entry-level tasks couldn’t disappear.
According to the International Labour Organization’s 2025 analysis, one in four jobs worldwide may be exposed to some degree to the effects of generative AI, but the most likely outcome is that roles will be transformed rather than fully replaced.
In several productivity studies, less experienced employees gained the most, while some senior groups showed a slight or negative effect. This could reduce the value of certain entry-level tasks in the short term, and it raises a question: where will juniors learn the trade if AI is already doing a significant share of the practice tasks?
How Should a Decision-Maker Weigh the Benefits and Drawbacks of Artificial Intelligence?
Weighing things up leads to a good decision when it’s done at the level of the individual process, and takes into account not just the expected time saved but also the cost of review, errors, data protection, licences and integration.
The following four questions provide a good first filter:
- How much time and cost does the process take up today? If there’s no baseline data, measure this first; without it, the return on investment can’t be demonstrated later.
- How high is the cost of an error? Usually lower for an internal memo, higher for a customer contract or a regulatory submission. Where the cost of an error is high, AI can provide a draft, but it can’t take over the decision or the responsibility for it.
- What data enters the system? Personal data or trade secrets require an approved provider, an appropriate contract, a legal basis, data minimisation and security controls. A business subscription alone isn’t enough.
- Who reviews the output, and who is responsible for it? How the review is carried out, how thorough it is, and who is responsible for it need to be set out before the pilot begins.
It’s worth starting a pilot with 2-3 clearly defined, measurable processes, and then deciding on expansion based on the actual results. This way, the benefits and drawbacks of artificial intelligence appear not as theoretical arguments but as comparable business data.
How Can the Risks of Artificial Intelligence Be Reduced?
A significant share of the risks can already be reduced if the company uses approved tools, restricts the input of sensitive data, sets out clear lines of human review and responsibility, and regularly measures quality and incidents. The depth of the controls should always be matched to the purpose of use, the sensitivity of the data and the possible consequences of an error.
Our guide, Using Artificial Intelligence, sets out the detailed process for adoption. You can read about the specific technical controls, such as access management, logging, endpoint protection and incident handling, in our article Information Security for Hungarian SMEs in 2026.
The Benefits and Drawbacks of Artificial Intelligence - Frequently Asked Questions
Which workflows are not a good place to start introducing AI?
A process that rarely repeats, has no clear quality benchmark, handles sensitive data, or where a single error could have significant financial, legal or security consequences, isn’t a good first choice. It’s worth starting with a task that’s clearly scoped, reversible and easy for a person to review.
What are the signs that AI isn't actually speeding up the work?
It’s a warning sign if review and correction take as much time as, or more than, AI saves in producing the first draft. That’s why it’s important to measure not just the initial speed, but also the total turnaround time, the follow-up work, the error rate and the number of tasks that had to be reopened.
How can staff's professional expertise be preserved alongside using AI?
AI should remain a supporting tool and shouldn’t automatically take over professional judgement or the final decision. It’s worth occasionally carrying out certain tasks without AI, jointly analysing flawed outputs, and documenting the professional considerations on which staff accept or reject a suggestion.
What should be done if a flawed AI output has reached a customer?
The first step is to stop any further use of the flawed content or decision, then assess its impact and correct the information if necessary. After that, it’s worth preserving the related inputs, outputs and logs, identifying the root cause, and adjusting the review process so the same problem doesn’t recur.
How can two enterprise AI tools be compared?
The two tools should be tested on the same tasks, representative of real work, using the same evaluation criteria. Alongside the quality of the answers, review time, total cost, data-handling terms, auditability, integration options and the difficulty of switching providers should all be taken into account.
How often should enterprise AI use be reviewed?
A review is warranted after any significant change to the model, provider, settings or workflow, when a new use case is introduced, and following a security incident. Even without these triggers, it’s advisable to check tools, permissions, data handling and measured business results at predetermined, risk-proportionate intervals.
When is it worth bringing in an external expert for AI adoption?
External support is especially warranted if the tool handles personal or commercially sensitive data, connects to company systems, can carry out actions on its own, or touches a regulated process. Bringing in an expert is also worthwhile if the right data protection, information security or integration expertise isn’t available in-house.
In Summary
The benefits and drawbacks of artificial intelligence don’t automatically balance each other out. For certain tasks, AI can bring significant time and capacity gains, while in other settings it doesn’t speed things up at all, or it stops being worthwhile once review and integration costs are added in. The best candidates are repetitive, clearly scoped and reviewable processes; the greatest risks come from uncontrolled tool use, the input of sensitive data, convincingly wrong output, and fraud attempts that appear authentic.
A leader’s task, then, isn’t to decide for or against AI in general. The expected benefit, the total cost, the cost of an error, the data protection and security conditions, and who reviews the output and is responsible for the result all need to be determined process by process.
Would you like to see clearly where AI creates real value and where it’s just a risk in your organisation? Our AI maturity organisational assessment can help with that. Take a look at our AI consulting services, and get in touch with us. In a free consultation, we’ll discuss which processes are worth starting with and what controls you’ll need.
