Where AI Actually Creates Value Inside a Business

A simple way to make artificial intelligence look transformational: start with the technology, then list everything it can do. Writing. Summarizing. Coding. Forecasting. Classification. Conversation. Document extraction. Image analysis. Research. Scheduling. Personalization.

The list becomes impressive quickly.

There is also a better way to evaluate AI. Start with the business. Where is money being lost? Where is employee time being consumed? Where are customers waiting? Where are decisions being made with incomplete information? Where are employees repeatedly performing predictable work? Where is growth constrained because adding volume requires adding people at approximately the same rate?

Those questions lead to measurable AI opportunities.

Productivity only matters when something economically useful happens afterward

One of the strongest field studies of generative AI examined 5,179 customer-service agents. Researchers Erik Brynjolfsson, Danielle Li, and Lindsey Raymond found that access to an AI conversational assistant increased issues resolved per hour by 14% on average. Less-experienced and lower-skilled workers experienced a 34% improvement, while the impact on highly experienced workers was small. The researchers also reported improvements in customer sentiment and employee retention.

That is meaningful evidence. But imagine applying the result carelessly. A company hears "AI improves customer-service productivity by 14%" and assumes it will reduce support cost by 14%. That does not follow.

Perhaps employees handle more customers, but payroll remains identical. That might still be valuable if the company avoids future hiring or improves response time. Perhaps the saved capacity allows the company to provide better service. Perhaps employees use the additional time on other work. Perhaps their particular support environment experiences no measurable improvement at all.

The operational gain must be translated into an economic outcome. Hours saved are not automatically dollars saved.

Customer service: strong potential, but not everywhere

Customer service is one of the most obvious AI applications because many organizations process large volumes of language-based interactions. Strong candidates include order status, basic troubleshooting, account navigation, standard policies, product information, appointment changes, and routine triage.

The economics improve when volume is high, and the distribution of questions is predictable.

But customer service is also a useful example of where simplistic automation fails. A company should measure containment rate, accuracy, first-contact resolution, customer satisfaction, average handling time, repeat contacts, escalation, and cost per resolution.

If an AI system resolves 60% of conversations but causes 20% of those customers to contact the company again, the headline automation percentage is misleading. The correct metric is not "How many chats did AI touch?" It is "What happened to the economics and customer outcome?"

Sales: make salespeople better before trying to replace them

Sales organizations contain substantial administrative work. Representatives research prospects, update CRM records, draft follow-ups, prepare meeting notes, find product information, review past interactions, and prioritize accounts. AI can potentially reduce that workload.

The higher-value question is what the salesperson does with the recovered time. If a representative saves five hours per week and spends those hours on additional high-quality customer conversations, capacity may increase. If the five hours disappear into additional internal meetings, the company has bought productivity without capturing the value.

Sales leaders should connect AI deployment to revenue-process metrics such as selling time, qualified opportunities created, conversion rate, cycle length, average contract value, retention, and revenue per representative.

Marketing: the marginal cost of content is collapsing

Generative AI can produce drafts, variations, summaries, research, personalization, images, and campaign concepts rapidly. This reduces the cost of producing content.

But that creates a second-order problem. When every competitor can produce more content, producing content itself becomes less differentiating. Companies may publish ten times as much while customers pay ten times less attention.

The economic opportunity therefore lies less in "making more content" and more in reducing production bottlenecks, accelerating experimentation, personalizing communications where personalization affects behavior, improving analysis, and allowing skilled employees to focus on positioning, customer understanding, creative direction, and distribution.

Volume is an output. Business impact is the objective.

Document-intensive processes may be particularly attractive

Many organizations still employ people to read forms, extract information, compare contracts, classify documents, summarize long records, route requests, enter data, or search internal files.

These processes can be promising AI candidates because the current labor cost can usually be measured. The workflow also lends itself to staged automation.

An AI system might extract information. A deterministic rule checks required fields. A human reviews unusual cases. The system records confidence levels and errors.

Often the best AI architecture is not full automation. It is a narrower workflow in which AI performs the portion it handles well.

Internal knowledge is valuable when employees can actually find it

Many companies already possess the answer employees need. The problem is retrieval.

Policies live in SharePoint. Product information lives in another repository. Customer information lives in a CRM. Operating procedures sit in PDFs. Important decisions are buried in email or collaboration platforms. Employees ask colleagues because locating authoritative information takes too long.

AI-powered knowledge retrieval can improve this situation, but only if the underlying knowledge is credible. If policies conflict, files are obsolete, permissions are poorly configured, and nobody owns the source information, AI can make the problem more convenient without making the information correct.

The value case therefore includes information governance. AI does not eliminate the need to maintain institutional knowledge. It increases the importance of doing so.

Finance and forecasting

AI can support finance teams through variance explanations, document review, transaction categorization, anomaly detection, scenario generation, narrative reporting, and data analysis.

But financial functions also illustrate why accuracy thresholds matter. An AI-generated first draft of a management commentary can be reviewed cheaply. An autonomous system issuing payments based on uncertain interpretation creates a different risk profile.

The same technology can therefore be appropriate at one layer and inappropriate at another.

Supply chain and operations

Traditional machine learning has been creating value in forecasting, routing, predictive maintenance, fraud detection, and optimization long before the generative-AI boom.

This distinction matters. Executives should not assume the newest generative model is the best solution to every AI opportunity. Some problems require prediction. Some require optimization. Some require computer vision. Some require statistical models. Some require deterministic automation. Some require a language model. Some require no AI whatsoever.

The correct technology comes after the business problem has been defined.

The jagged frontier

An HBS/BCG experiment involving 758 consultants illustrates why selecting tasks matters so much. For tasks within the capabilities of GPT-4 at the time, consultants using the model completed more tasks, worked more than 25% faster, and received more than 40% higher human-rated performance scores. But researchers also found that AI could reduce performance on tasks outside its capability frontier.

This is one of the most important findings for executives. The useful unit of analysis is often not the job. It is the task.

"Can AI replace an analyst?" is too broad. Ask instead: Which analyst activities can AI perform reliably? Which can it assist? Which require human interpretation? Which require accountability? Which become more valuable when AI handles lower-value work?

The same logic applies to accountants, marketers, recruiters, developers, lawyers, customer-service representatives, and managers.

A practical opportunity screen

Executives can begin with four questions.

Is there enough volume? Automating a task performed ten times per year may never repay the implementation cost. A task performed 50,000 times per month is different.

Is the work sufficiently predictable? Repetition and standardization improve automation economics. Highly variable processes may require more supervision and exception handling.

Can success be measured? A company should know what metric AI is expected to change. Without a baseline, ROI becomes storytelling.

Does improving the task matter financially? Saving 80% of the time on a process costing $20,000 annually creates at most $16,000 of theoretical labor capacity before implementation cost. Saving 20% on a $10 million process may be much more interesting. Percentage improvement alone is not enough. The size of the economic base matters.

The opportunity portfolio

Rather than ask every department for "AI ideas," management should create an opportunity portfolio.

For each use case, estimate current annual cost, transaction volume, employee hours, current error rate, revenue impact, customer impact, data readiness, risk, implementation difficulty, expected benefit, and confidence in the estimate. Then prioritize.

A boring automation with measurable value may deserve funding before an impressive autonomous agent. That is not a lack of ambition. It is capital discipline.

AI creates value when capability meets an economically important problem under conditions where the business can capture the improvement. The opportunity exists at that intersection. Not everywhere.

Comdatus Consulting perspective: An AI opportunity assessment should produce a prioritized portfolio of business cases rather than a catalog of technologies. The objective is to identify where AI can improve revenue, cost, capacity, quality, or risk enough to justify implementation — and where it cannot.