Generative AI for business: where it pays off and where it bites

Generative AI for business is best understood as a fast, fluent and fallible assistant: it drafts, summarizes, classifies and codes at a scale no team can match, but it does not know when it is wrong. Companies that get value from it pick narrow, high-volume tasks, keep a human accountable for the output and measure results against a baseline. Those that treat it as magic tend to end up with impressive demos and no change in the numbers.

What generative AI actually does

Generative AI refers to models that create new content, such as text, images, code or audio, rather than simply sorting or scoring existing data. The best-known kind is the large language model (LLM), a system trained on huge amounts of text to predict the most likely next word. Scale that prediction up and you get software that can write an email, explain a contract clause or produce working code.

Two features matter for managers. First, the output is probabilistic: ask the same question twice and you may get different answers. Second, the model has no built-in sense of truth. It produces what sounds plausible, which is usually right on well-trodden topics and sometimes badly wrong on specific facts. Every sensible deployment is designed around those two traits.

Most business use comes in three flavors: a general chat assistant for employees, AI features built into software the company already uses, and custom applications that connect a model to the company’s own documents and systems. The last is often done with retrieval-augmented generation (RAG), where the system first looks up relevant internal documents and then asks the model to answer using only those sources.

Where it creates value

The strongest cases share a pattern: lots of language-heavy work, many repeated tasks, and a person who can quickly judge whether a draft is good.

Customer service

Support is the best-studied use so far. In a study of 5,179 customer support agents published by the National Bureau of Economic Research, access to an AI assistant raised productivity, measured as issues resolved per hour, by 14% on average and by 34% for novice and low-skilled workers. The finding is worth dwelling on: the tool appeared to spread the habits of the best agents to the newest ones. For businesses with high staff turnover, faster onboarding may be worth as much as the raw speed gain.

Marketing and sales

Drafting product descriptions, ad variations, email sequences and localized copy is classic generative work. The value is less about replacing writers than about producing more variants to test and freeing people for strategy and editing. The risk is sameness: if every competitor uses similar tools with similar prompts, generic copy loses its edge, so brand voice and human editing matter more, not less.

Software development

Coding assistants suggest lines and functions, explain unfamiliar code, write tests and translate between languages. Developers often report the biggest help on boilerplate and unfamiliar libraries. The catch is review: AI-suggested code can contain bugs or security flaws, so it has to pass the same testing and code review as anything a person writes.

Knowledge work

Summarizing long reports, drafting first versions of memos, searching internal policies and turning meeting notes into action lists are everyday wins. Here, RAG systems that cite their sources are far more trustworthy than a general chatbot answering from memory.

What it really costs

The license or usage fee is the visible part. The real bill has more lines, and a worked example makes that plain.

Say a company with 200 support agents is weighing an AI assistant. Hypothetically, it might pay:

Cost item	What it covers	Example annual figure (hypothetical)
Licenses or usage fees	Per-seat fees or per-token API charges	$120,000
Integration	Connecting to the help desk, knowledge base and CRM	$80,000
Data preparation	Cleaning and updating help articles the model draws on	$40,000
Review and quality control	Spot checks, escalation handling, monitoring	$50,000
Training and change management	Teaching staff when to trust or override the tool	$30,000
Total		$320,000

If the tool lets each agent handle, say, 10% more tickets, and a fully loaded agent costs $60,000 a year, the capacity gained is worth about $1.2 million, comfortably above the cost. If the real gain turns out to be 2%, it is worth $240,000 and the project loses money. That is why measurement, not enthusiasm, should decide.

The risks: hallucination, privacy and IP

Hallucination

A hallucination is a confident but false output, such as an invented statistic, a policy that does not exist or a citation to a case that was never decided. You cannot fully eliminate it, but you can contain it: ground answers in approved documents, require citations, limit the tool to tasks where errors are easy to catch, and keep a human sign-off on anything sent to customers, regulators or courts.

Data privacy and confidentiality

Employees pasting customer records, contracts or source code into an unapproved tool is the most common real-world failure. Companies should approve specific tools, check vendors’ terms on data retention and training use, classify which data is permitted, and log usage. Existing privacy and sector rules still apply; the new tool does not suspend them.

Intellectual property

IP questions run in two directions. On the output side, the U.S. Copyright Office has been studying how copyright applies to AI-generated material and has taken the position that protection depends on human authorship, so purely machine-generated content may not be protectable. The USPTO has issued its own guidance on AI and inventorship. On the input side, litigation over whether training on copyrighted works is lawful is still working through the courts. Practical steps: review vendor indemnities, keep records of human contribution to important creative work, and get legal advice before relying on AI output for core brand assets.

Other risks are worth naming too: biased outputs in hiring or lending, prompt injection (where hidden instructions in a document hijack the model), and over-reliance, where staff stop checking.

Governance with the NIST AI RMF

Companies do not have to invent governance from scratch. The NIST AI Risk Management Framework, released in 2023, is voluntary and organizes the work into four functions:

1. Govern: set policies, roles and accountability for AI across the organization.
2. Map: understand the context of each use, who it affects and what could go wrong.
3. Measure: test and track risks such as accuracy, bias, security and reliability.
4. Manage: prioritize risks, act on them and monitor after deployment.

NIST followed up with a Generative AI Profile (NIST AI 600-1) that lists risks that are unique to or made worse by generative AI, including confabulation (NIST’s term for hallucination), data privacy, information security and intellectual property, with suggested actions for each. A mid-sized company can use it as a checklist: a short acceptable-use policy, an inventory of AI tools in use, a named owner for each application and a review process proportional to the stakes.

AI governance also connects to broader strategy. Deciding where AI fits is part of setting corporate strategy, and it rarely works without the data and process foundations covered in our guide to digital transformation.

How to run a pilot that proves ROI

A pilot should answer one question: does this tool improve this workflow enough to justify its full cost? Here is a practical sequence.

1. Pick one workflow. Choose something frequent, measurable and low-risk, such as drafting replies to common support tickets, not “use AI across marketing.”
2. Measure the baseline. Record time per task, volume, quality scores, error rates and customer satisfaction for several weeks before the tool arrives.
3. Set success criteria in advance. For example: at least 10% more tickets handled per hour with no drop in satisfaction scores.
4. Use a control group. Give the tool to one team and not to a comparable one, so seasonal swings do not masquerade as AI gains.
5. Keep humans in the loop. Require review of outputs and track how often people edit or reject them. A high rejection rate is an early warning.
6. Count the full cost. Include integration, review time and training, not just licenses.
7. Decide. Scale, redesign or stop. A clear “no” after eight weeks is a good outcome compared with a vague program that drifts for two years.

Watch for the gap between time saved and value captured. If AI saves each employee 30 minutes a day but nothing about staffing, workload or output changes, the savings exist only on paper.

The bottom line

Generative AI for business is neither a fad nor a free lunch. The evidence so far points to real gains in customer service, content, coding and knowledge work, especially for less experienced staff, alongside real risks around accuracy, privacy and IP. The companies that benefit treat it like any other investment: start narrow, measure honestly, govern sensibly with a framework such as NIST’s, and scale only what works. For more on how technology is reshaping companies and jobs, explore our Tech & Innovation hub and our trend report on automation and the future of work.