Where Enterprise AI Delivers Real ROI: Six Domains That Work

Most companies can show AI usage. Far fewer can show a change in profit, cost, throughput, or risk.

That gap matters. A company can buy thousands of copilot seats and generate millions of tokens without changing how work gets done. A 2026 NBER survey of nearly 6,000 executives found that nine in ten reported no impact from AI on employment or productivity during the prior three years. The same executives still expected gains ahead.

The gains will come from specific workflows. Each one needs an owner and a baseline. The team must define work a reviewer can accept without correction. It also needs a way to capture the value. Saving an employee 20 minutes has little financial effect if the process and output stay the same.

At OpenNash, we group the enterprise opportunity into six domains. They can appear as customer products, internal applications or background systems. The interface matters less than the business result.

Start With the ROI Equation

Every AI business case should name the unit of work. Examples include a resolved support case, reviewed contract, accepted code change, qualified opportunity, investigated security alert or processed invoice.

The basic calculation is:

Annual value = work volume × improvement per accepted outcome − full operating cost

Full operating cost includes the model and infrastructure. It also includes software, integration work and human review. Add corrections, monitoring and support. The calculation should account for new failure modes too.

Track business measures that finance and operating leaders already understand:

• Cost per accepted case
• Cycle time
• Error and rework rate
• Revenue won or retained
• Gross margin
• Fraud or security loss avoided
• Downtime and scrap avoided
• Inventory turns and working capital
• Hiring or outsourcing avoided

Do not present prompts, generated documents, active seats, or token consumption as ROI. Those numbers describe activity. They do not prove value.

This test quickly separates serious projects from AI theater. It also explains why the following six domains attract sustained investment.

1. Customer Experience

Customer support remains one of the clearest enterprise AI use cases. The work has high volume. The company already tracks resolution time, contact rate, satisfaction, escalation, and cost per case.

Support AI creates more value when it can complete an approved action. It identifies the customer and retrieves account context. It applies policy, takes the action and records the result. Exceptions go to an employee. Useful workflows include order changes, account recovery and return processing. Appointment scheduling, claims intake and technical support also fit.

Agent assistance is often the best first step. AI can search knowledge, draft a response, summarize the case, and recommend the next action. A human remains responsible for the outcome. This model improves new-agent performance and creates production data before the company grants direct customer access.

A field study of 5,179 support agents found a 14 percent increase in issues resolved per hour. Less experienced agents saw the largest gains.

Measure successful resolution rather than ticket deflection. A bot can reduce visible volume by making support harder to reach. Track repeat contacts and reopened cases. Add refunds, escalation and satisfaction. Count human cleanup too. Those measures expose false savings.

A good first project handles one common request from start to finish. Pick a request with clear policy and a reversible action. Keep sensitive exceptions with employees.

2. Knowledge Work

Enterprise copilots can search internal knowledge, prepare analysis, draft documents, summarize meetings, and help subject-matter experts build their own workflows. This domain has broad demand because it touches almost every employee.

The business case needs discipline. Faster writing does not create value on its own. The organization must convert saved time into more output, better decisions, shorter cycle times, or lower staffing pressure.

A six-month randomized field experiment involving 6,000 workers found that access to a GenAI tool reduced time spent on email and modestly accelerated document work. It did not produce a meaningful change in meeting time. The result captures both the potential and the limitation of horizontal copilots. They help with individual tasks. They rarely redesign a cross-functional process by themselves.

SME-built operational automation turns an expert's procedure into a repeatable workflow. A finance lead can define how the company prepares a margin review. A compliance manager can define the evidence required for an audit package. A sales operations lead can define how account research enters the CRM. The AI system then runs that procedure against company data and tools.

Start with one recurring deliverable that consumes several hours each week. Define the source systems and required evidence. Specify the final format. Measure preparation time, review time, error rate, and delivery speed.

Knowledge systems also need source ownership. AI search cannot fix stale policies, duplicate documents, or conflicting records. Give each important source an owner and a freshness rule before employees depend on the answers.

3. Software and IT

Software teams adopted AI early because code can be tested. Repositories also contain years of structured context. That makes engineering a strong fit for generation and review. Migration, testing and debugging also benefit.

The highest-value use cases often sit outside greenfield code generation:

• Generate tests before changing legacy behavior
• Explain unfamiliar code and dependencies
• Reproduce bugs from logs and issue reports
• Upgrade frameworks and dependencies
• Apply repetitive refactors
• Review pull requests against local standards
• Draft runbooks and technical documentation
• Investigate incidents across logs and monitoring tools

Three enterprise field experiments with 4,867 developers found a 26 percent increase in completed tasks. The result was noisy across individual companies. Less experienced developers adopted the tool more often and saw larger gains.

Other evidence gives a useful warning. METR found that experienced open-source developers working in mature repositories took 19 percent longer with early-2025 AI tools. The developers believed they were faster. Repository knowledge, task shape, and review burden changed the result.

Measure lead time from issue to accepted production change. Count failed builds and review rounds. Track escaped defects, incidents and maintenance cost. Lines of generated code reward volume rather than quality.

Legacy modernization can create large returns. It also carries serious risk. Begin with codebase mapping, behavioral tests, narrow modules, and parallel validation. Models can generate a C-to-Rust translation. Tests and engineering controls must prove that it behaves correctly.

4. Sales and Go-to-Market

AI can improve sales when it helps teams choose the right account, understand the customer, and act at the right time. It performs poorly as a machine for producing unlimited outreach.

The valuable workflows include:

• Account and lead prioritization
• CRM enrichment and cleanup
• Churn and expansion signals
• Call analysis and coaching
• Proposal and RFP preparation
• Pricing and discount guidance
• Renewal briefs
• Territory planning
• Next-best action recommendations

These systems work best when they use first-party signals. Product usage and support history are useful. So are billing events, prior conversations and contract dates. They reveal more than generic web research. The agent should cite the signal behind each recommendation. A seller needs to know why an account moved to the top of the list.

Autonomous outreach is easy to build and hard to defend. Cheap personalization increases message volume across the whole market. Buyers receive more plausible spam. Brand risk rises while response quality falls.

Track qualified pipeline, conversion and sales-cycle time. Track retention, expansion revenue and CRM completeness. Avoid emails sent or records enriched. Do not count meetings without qualification. Those measures can rise while revenue quality declines.

A strong first project prepares a complete renewal or opportunity brief from trusted systems. It removes research and data entry. The account owner still decides what to say and when to say it.

5. Security, Risk, and Compliance

Security is a distinct enterprise AI domain because the value often comes from loss avoided. One prevented incident or fraud event can fund the system. Security teams also face large alert volumes and expensive analyst time.

Useful workflows include:

• Security alert triage and investigation
• Identity and access review
• Vulnerability prioritization
• Fraud and claims anomaly detection
• AML and sanctions case preparation
• Vendor risk review
• Contract obligation monitoring
• Audit evidence collection
• Policy checks against proposed actions

AI can gather evidence across tools and summarize the event. It can map the event to policy and recommend a response. Deterministic rules should enforce permissions and blocked actions. They should also enforce approval thresholds and audit requirements.

Deloitte's enterprise GenAI research found that cybersecurity and IT led successful scaling and reported ROI. Analysts spend much of their time assembling context from logs and tickets. They also need identity data, asset records and prior incidents.

Security AI creates its own attack surface. Treat retrieved content as untrusted. Restrict tools and credentials. Log every action. Require approval for destructive changes. Test prompt injection and poisoned documents. Test data leakage, privilege escalation and unsafe tool use before production.

Measure mean time to investigate and false-positive rate. Count analyst hours per case and losses avoided. Track audit preparation time and incident rate. Start with read-only investigation. First run the system beside the current process without taking action. Add narrow response actions only after it passes real cases.

6. Core Operations

Core operations belongs in the framework because it moves money, goods, documents and decisions through the company.

This domain covers:

• Procurement and spend management
• Accounts payable and receivable
• Financial close and reconciliation
• Supply-chain planning
• Demand forecasting and inventory
• Production scheduling
• Claims and underwriting
• Loan and application processing
• Logistics and route planning
• Predictive maintenance
• Visual quality inspection
• Field-service dispatch

Many of these workflows predate generative AI. They combine forecasting and optimization with computer vision. Rules, document extraction and language models handle other steps.

Document-heavy operations belong here when they support an operational process. An invoice workflow reads the document and matches it to a purchase order. It checks policy and updates the ERP. Any exception goes to a reviewer. The document is an input. The business outcome is a correct payment.

Physical operations can produce especially clear value. Siemens reports that an automotive predictive-maintenance deployment generated tens of millions in avoided downtime with payback in under three months. That is a vendor case study, so buyers should validate the baseline and attribution. The system creates value by detecting failures early enough to avoid lost production.

Procurement systems can reduce price variance and off-contract spend. Supply-chain systems can reduce stockouts, excess inventory and expedite fees. They can also cut planning time. Finance systems can shorten close cycles and reduce manual reconciliation. Claims systems can lower handling cost and surface suspicious cases sooner.

Start with an exception queue. Let existing systems handle clean transactions. Use AI to assemble context for the cases that consume human time. Measure cost per transaction and exception rate. Add cycle time, working capital and downtime where relevant. Manufacturing teams should also track scrap and forecast error.

Choose the Right Execution Mode

The six domains describe where value appears. The execution mode describes how the system fits into work.

Most companies should begin with assistance or read-only automation. Run the system on 50 to 100 real cases. Compare its proposed work with the human outcome. Record corrections and exceptions. Then add one permission at a time.

McKinsey's 2025 global survey found the highest reported GenAI use in marketing and sales. Product development, service operations and software engineering also ranked high. IT was another common target. Reported cost benefits were strongest in software engineering, manufacturing and IT. The hottest applications do not always have the clearest financial return.

For a first project, favor work with five traits:

1. Enough volume to matter
2. A measurable baseline
3. A clear end state
4. Cheap verification
5. A safe exception path

If the workflow lacks those traits, fix the process before adding AI.

How OpenNash Can Help

OpenNash connects AI to the systems where teams handle support cases, documents, code, sales records and operational exceptions. We start with work where cost, delay, revenue or risk can be measured before development starts.

Our process covers four stages:

1. Audit the operation and rank candidate workflows by value, feasibility, and risk.
2. Define which data and tools the system may access. Set approval rules and tests for expected failures.
3. Run the workflow beside the current process without letting it take action. Compare both results on real cases.
4. Deploy with monitoring, documentation, CI/CD, and a full ownership handoff.

We support all six domains covered here. That includes customer service, knowledge workflows and software delivery. It also includes GTM, security and core operations. The client owns the code and operating assets after deployment.

If you have a workflow with meaningful volume or cost, book a call with its monthly volume, current process and error rate. We will identify a small pilot with a measurable result.