AI Capabilities Across Business Functions

Good to see you again. In the previous lesson, you learned to assess a generative-AI proposal through its practical limitations: accuracy, grounding, variability, privacy, security, cost, and the consequences of error. That assessment remains essential here. A promising AI capability is not automatically a promising business opportunity.

This lesson moves from what AI can and cannot reliably do to where it can create value. By the end, you should be able to map a business function such as sales, operations, finance, or customer service to a specific workflow, the AI capability that fits it, a meaningful output, and a measurable business outcome.

---

Start with capabilities, not fashionable use-case labels

“AI for sales” or “an AI customer-service agent” is too vague for a useful commercial conversation. It hides the key questions:

• What work is currently difficult, slow, high-volume, or inconsistent?
• What type of output is required?
• Is the AI expected to understand, predict, recommend, generate, detect, or optimize?
• Who will use the output, and what will they do with it?
• Which metric should improve if it works?

A capability is the repeatable thing the system does. A function is the part of the organization where that capability is used.

For example, summarization is a capability. It can support:

• a sales representative preparing for a customer meeting;
• a service agent catching up on a long case history;
• an operations manager reviewing shift handovers;
• a finance analyst reading supplier contracts.

The same capability can create value in different functions, but its required data, acceptable error rate, and controls will differ.

The AI Opportunity Radar is a useful starting point for account conversations. It prevents a narrow focus on public chatbots by showing that AI opportunities can be:

• internal and everyday, such as administrative support in finance, HR, and IT;
• external and everyday, such as customer-service support;
• internal and potentially game-changing, such as supply-chain, R&D, or operational optimization;
• external and potentially game-changing, such as AI-enhanced products or personalized customer experiences.

Treat this as a conversation map, not a prioritization model. A customer-service assistant may be more valuable than an ambitious “game-changing” concept if it targets a painful, high-volume workflow with usable data and a clear success measure.

---

A practical vocabulary for mapping AI

Most enterprise AI opportunities can be described with six capability families. You do not need to begin by naming a model or vendor. Begin with the kind of work the business needs done.

AI capability	Typical output	Example business tasks	Often fits best with
Extract and classify	Fields, categories, tags, confidence levels	Read invoices, classify tickets, identify document types	Document AI, OCR, language models, rules
Predict and detect	Forecast, probability, risk score, anomaly alert	Forecast demand, identify churn risk, flag unusual transactions	Predictive machine learning
Recommend and prioritize	Ranked list, next-best action, allocation suggestion	Prioritize leads, recommend stock transfers, triage cases	ML, optimization, business rules
Generate and summarize	Draft, summary, explanation, response	Draft outreach, summarize calls, create case notes	Generative AI, often grounded in approved sources
Perceive speech and images	Transcript, defect flag, object detection	Transcribe calls, inspect products, monitor safety conditions	Speech AI, computer vision
Optimize and coordinate	Schedule, route, plan, resource allocation	Plan delivery routes, schedule staff, allocate inventory	Optimization methods, ML, rules-based workflow tools

The boundaries are not always rigid. An invoice workflow, for instance, may combine document extraction, classification, rules-based validation, exception routing, and generated explanations for a human reviewer. That is why enterprise AI is usually a workflow solution, not a standalone model.

A useful rule is:

Map the business problem to the required output first; then identify the AI capability, data, and workflow changes needed to produce it safely.

---

---

Mapping sales: help the seller decide and communicate

In enterprise selling, scarce seller time is usually the central constraint. Reps may spend hours researching accounts, deciding where to focus, preparing for meetings, updating CRM records, and drafting follow-ups. AI can support those activities, but “AI that sells for you” is usually an inaccurate and risky claim.

A more precise sales map looks like this:

Sales workflow	Suitable AI capability	Data typically needed	Useful output	Business measure
Choose where to spend seller time	Lead scoring, account prioritization, prediction	CRM history, engagement data, firmographic data, past conversion outcomes	Ranked account or lead list with reasons	Meetings created, conversion rate, pipeline quality
Prepare for an account meeting	Search, summarization, generation	Permitted CRM notes, call records, account plans, approved collateral	Account brief, meeting recap, suggested questions	Research time, meeting preparedness, follow-up quality
Create a first draft of outreach	Generation	Seller inputs, approved messaging, account context	Draft email or call outline	Time to first touch, reply rate, seller adoption
Identify pipeline risk	Prediction, anomaly detection, summarization	Opportunity stage history, activity data, forecast data	Stalled-deal alert or forecast risk summary	Forecast accuracy, slippage rate, deal-cycle length
Keep CRM records current	Extraction, classification, summarization, workflow automation	Call transcripts, emails, meeting notes	Suggested CRM updates and next steps	Administrative time, record completeness

There are two important distinctions here.

First, lead scoring is not the same as lead generation. A scoring model estimates which existing leads or accounts deserve attention based on available signals and historical patterns. It does not create demand or replace a seller’s judgment about relationship quality, strategic fit, or buying context.

Second, a generated email is not an approved commercial commitment. A language model can produce a draft quickly, but a human seller remains accountable for pricing, product claims, contractual language, and the relationship itself.

Salesforce’s example makes the workflow concrete: AI helps select a likely high-priority prospect, prepare a point of view, and draft outreach. It also emphasizes grounding: outputs should use the organization’s CRM information and defined sales practices rather than generic model knowledge.

A credible sales-oriented hypothesis might be:

“For account executives managing large portfolios, use permitted CRM activity and approved sales material to prioritize accounts and create editable meeting briefs. Measure reduced preparation time and improved CRM completeness, while keeping pricing, commitments, and outbound messages under seller approval.”

That statement has a user, workflow, capability, data boundary, outcome, metric, and control. It is much stronger than “We provide an AI sales agent.”

---

Mapping customer service: resolve routine work and strengthen human support

Customer service includes two distinct opportunity types:

1. Customer-facing assistance, where an AI system responds directly to a customer.
2. Agent assistance, where AI helps a human support employee serve the customer.

They may use similar language capabilities, but they carry different risk.

A customer-facing assistant can answer common questions, help users navigate a process, retrieve order status, or guide basic troubleshooting. It may improve availability and response time. But it must draw on approved, current information and have a clear handoff path for complex, sensitive, or ambiguous situations.

Agent assist is frequently a safer initial opportunity. It can summarize case history, retrieve relevant knowledge articles, propose a response draft, classify the issue, and complete post-call documentation. The human agent can inspect the output before it affects the customer.

Service workflow	Suitable AI capability	Example output	Potential value	Essential boundary
Triage incoming cases	Classification and prioritization	Issue category, urgency, recommended queue	Faster routing, reduced backlogs	Validate routing accuracy; enable manual reassignment
Find answers during a live case	Search and summarization	Relevant policy or troubleshooting guidance	Lower handling time, greater consistency	Use approved knowledge sources and show source references
Support the service agent after a call	Speech transcription and summarization	Call notes, actions, CRM update draft	Less administrative work	Agent checks facts before saving records
Answer frequent customer questions	Conversational generation plus retrieval	Guided answer or self-service response	Faster resolution and 24-hour availability	Escalate exceptions; do not invent policy or account facts
Identify recurring service problems	Sentiment analysis, clustering, analytics	Trend report on common complaint themes	Root-cause improvement, reduced repeat contacts	Treat patterns as evidence to investigate, not settled fact

The mapping should always follow the customer journey. A conversational interface is valuable only when it improves a meaningful point in that journey: finding information, completing a simple request, receiving a timely update, or reaching the right human agent with minimal repetition.

Avoid an overly broad proposal such as “replace the contact center with AI.” A better first scope might be:

“Use AI to classify incoming support requests, retrieve approved answers for agents, and draft post-call documentation. Measure average handling time, transfer rate, repeat-contact rate, and agent adoption. Keep account changes, refunds, complaints, and exceptions under defined human approval rules.”

---

Mapping operations: predict, detect, optimize, and act on the physical world

Operations is broad: supply chain, procurement, logistics, manufacturing, field service, warehousing, and internal service delivery can all sit within it. Its AI opportunities often look different from a text chatbot because the data may be transactional, sensor-based, image-based, or time-series data.

The central operations question is often:

Can the organization see what is happening early enough to make a better operational decision?

Consider a distributor that struggles with stockouts and excess inventory. Several AI capabilities could apply:

• Demand forecasting estimates future demand by product, location, or period.
• Anomaly detection identifies unexpected demand changes, late supplier performance, or unusual production behavior.
• Optimization recommends how to allocate inventory or schedule delivery routes under capacity and cost constraints.
• Generative AI lets planners ask questions in natural language and produces summaries of exceptions—but it does not replace the forecasting or optimization engine itself.

Operations workflow	Suitable AI capability	Typical inputs	Output	Business measures
Plan inventory and replenishment	Forecasting and optimization	Sales history, stock levels, supplier lead times, seasonality	Demand forecast or replenishment recommendation	Stockout rate, excess stock, inventory carrying cost
Maintain industrial equipment	Predictive modeling and anomaly detection	Sensor readings, maintenance records, failure history	Early warning of likely failure	Unplanned downtime, maintenance cost, asset availability
Inspect quality or safety	Computer vision	Images or video from production or facilities	Defect or hazard flag, visual evidence	Defect rate, inspection time, safety incidents
Coordinate field service	Optimization and workflow support	Work orders, technician skills, location, parts availability	Schedule or routing recommendation	First-time fix rate, travel time, SLA performance
Process operational documents	Extraction, classification, workflow automation	Purchase orders, bills of lading, forms, claims	Extracted fields, exception queue, routing suggestion	Cycle time, error rate, cost per document

Operations proposals must account for real constraints that a model may not know: safety requirements, labor rules, machine capacity, supplier agreements, weather, and site-specific judgment. Therefore, a recommendation should be explainable enough for an operations manager to validate or override it.

For example, a “predictive-maintenance AI” does not promise that a machine will fail on a particular date. It may flag a pattern that merits inspection. The operational value comes when the organization can act on that signal through maintenance planning and has a way to measure whether earlier intervention reduced downtime.

---

Mapping finance: automate information work, improve visibility, protect control

Finance teams often have high-volume, rules-governed, document-heavy work. That makes them strong candidates for AI support—but also makes controls non-negotiable. Finance is responsible for reliable records, controlled processes, and decisions that affect cash, customers, and regulatory obligations.

A useful finance map separates supporting analysis from authorizing financial outcomes.

Finance workflow	Suitable AI capability	Example output	Potential value	Control needed
Process invoices and expense documents	OCR, extraction, classification, rules	Extracted supplier, amount, line items, exception flag	Faster processing, lower manual effort	Match against approved records; retain approval controls
Reconcile accounts or transactions	Matching logic, anomaly detection	Suggested match or exception queue	Faster close, fewer unresolved exceptions	Deterministic rules and human review for ambiguous cases
Forecast cash flow, revenue, or cost	Predictive modeling and scenario analysis	Forecast range, variance alert, explanatory summary	Better planning and earlier intervention	Compare performance with actuals; expose assumptions
Detect fraud or unusual activity	Anomaly detection and classification	Risk alert with evidence	Faster investigation and risk reduction	Investigation and decision remain with authorized staff
Prepare management commentary	Summarization and generation	Draft explanation of budget variance or close results	Reduced reporting effort	Finance validates all figures and statements
Review contracts or policies	Extraction and summarization	Key clauses, obligations, missing fields	Faster initial review	Legal, procurement, or finance experts validate conclusions

A common mistake is to frame an AI-generated narrative as “financial analysis.” A generated explanation can help a finance professional communicate a variance, but the actual numbers should come from authoritative finance systems and validated calculations.

Similarly, a fraud model should flag suspicious activity for investigation; it should not be described as an automatic verdict of wrongdoing. The distinction matters commercially, operationally, and ethically.

---

The same opportunity can span multiple functions

Enterprise problems rarely respect departmental boundaries. A single use case may start in sales, create work in operations, and affect finance.

Take dynamic pricing and inventory management:

• Sales wants to protect revenue and conversion.
• Operations wants stock availability and feasible fulfillment.
• Finance wants healthy margins and controlled discounting.
• Customer service may handle questions and complaints caused by price changes.

The AI may forecast demand, recommend prices, and suggest replenishment actions. But the opportunity should not be sold as “an operations tool” if commercial leadership owns pricing decisions, or as “a sales tool” if supply constraints determine whether the offer can be fulfilled.

This is why an account-level AI conversation benefits from a simple map:

Business function	Current friction	AI-supported workflow	Shared dependency
Sales	Reps discount inconsistently	Margin-aware deal guidance	Current costs, stock availability, approval policy
Operations	Demand changes create stockouts	Demand forecast and allocation recommendations	Sales pipeline and order data
Finance	Margin erosion appears too late	Variance detection and scenario reporting	Price, cost, and transaction data
Customer service	Customers ask about availability or price changes	Approved answers and proactive updates	Accurate inventory and pricing data

The commercial lesson is important: identify the business owner, but also identify the functions that provide data, execute the change, or carry the risk if the system fails.

---

A reusable template for opportunity mapping

When a prospect says, “We need AI in our department,” turn the broad request into a testable business hypothesis:

For [specific users]
in [named workflow],
use [AI capability]
on [permitted, authoritative data]
to produce [specific output],
so they can [decision or action],
measured by [business metric],
with [human review, escalation, or control].

Here are four examples:

Function	Weak statement	Stronger opportunity hypothesis
Sales	“We need AI to improve prospecting.”	“For BDRs, rank inbound leads using CRM engagement and firmographic data, with reasons shown to the rep, to improve speed-to-follow-up and qualified-meeting conversion.”
Customer service	“We want an AI chatbot.”	“For support agents, retrieve approved troubleshooting guidance and draft editable responses for common technical issues, to reduce handling time while routing unresolved cases to specialists.”
Operations	“We need AI for the supply chain.”	“For inventory planners, forecast demand by SKU and region using sales and inventory history, then flag likely stockouts, to reduce lost sales and excess inventory.”
Finance	“We want AI for finance automation.”	“For accounts-payable analysts, extract invoice fields and identify exceptions against purchase-order and goods-receipt data, to reduce invoice cycle time while preserving approval controls.”

Notice what these statements avoid:

• unsupported claims about guaranteed revenue or savings;
• vague references to “automation” without a workflow;
• an assumption that generative AI is the right technology;
• a claim that people or controls are unnecessary;
• a promise to use data before confirming permission, quality, and access.

---

Key takeaways

An effective AI opportunity map connects five things:

1. Business function and workflow: where work happens and what currently causes friction.
2. AI capability: extract, classify, predict, recommend, generate, perceive, or optimize.
3. Data and output: what the system uses and what it actually produces.
4. Business action and metric: what changes after the output, and how value is measured.
5. Control and accountability: who reviews, approves, overrides, or owns the outcome.

Across functions, the patterns are consistent:

• Sales uses AI to prioritize, prepare, communicate, and maintain CRM quality.
• Customer service uses AI to route, retrieve knowledge, assist agents, and resolve suitable routine questions.
• Operations uses AI to forecast, detect, inspect, optimize, and coordinate.
• Finance uses AI to extract information, detect anomalies, forecast, reconcile, and draft analysis under strong controls.

Next, you will decide whether a business problem is genuinely best suited to AI, conventional automation, or process redesign. That decision prevents a common and expensive error: using AI to automate a workflow that should first be simplified, standardized, or redesigned.