Which Business Process Should You Automate With AI First?

A job title bundles many different activities: gathering information, making routine decisions, handling exceptions, building trust, negotiating, approving, documenting, and recovering from errors. Saying automate the office manager or replace the receptionist prevents useful analysis. Decompose the role into trigger, input, decision, action, handoff, and outcome. You may find three good assistance tasks and several responsibilities that should remain human.

For example, quote follow-up includes detecting an open quote, checking permission, choosing timing, drafting a message, interpreting the reply, answering questions, negotiating price, confirming availability, updating status, and attributing payment. AI may classify replies and draft from approved facts. Rules may control eligibility and timing. A salesperson should own objections and commitments. The useful design distributes work rather than pretending one agent should own the entire relationship.

Task decomposition also reveals simpler fixes. If employees retype fields because two systems are not connected, ordinary integration may solve the problem. If nobody knows the next action, a status model and queue may be enough. If the policy is unclear, write it. AI should address variable language or judgment support where it adds value, not decorate every operational gap.

Ask employees to log repetitive work for two weeks using task, trigger, frequency, minutes, systems, common errors, waiting time, and downstream effect. Review calls, inboxes, spreadsheets, forms, and handoffs. Customer questions are particularly valuable because the same question may expose missing content, sales training, routing, and product design. Do not rely only on what feels annoying; interruptions and waiting are often undercounted.

Look for queues: unread voicemails, unassigned leads, pending quotes, documents waiting for fields, inbox messages awaiting classification, or reports assembled repeatedly. A queue provides volume, aging, ownership, and outcome data. It also makes exceptions visible. By contrast, vague aspirations such as better marketing or smarter operations are too broad to test.

Include existing automation. The best opportunity may be repairing a brittle rule, consolidating duplicate tools, or adding an exception classifier to a deterministic workflow. Replacing a working system with AI carries switching cost and new risk. Rank net improvement over the current process, not excitement relative to doing nothing.

Frequency creates learning and potential leverage, but volume alone is not value. Multiply eligible runs by avoidable minutes, then add measurable conversion, quality, and risk effects. Use loaded labor cost rather than salary alone, and distinguish work removed from work shifted to review. If the saved time is fragmented into thirty-second pieces that employees cannot redeploy, financial value may be smaller than the spreadsheet suggests.

Revenue opportunities need a causal chain. Faster lead classification matters only if qualified leads reach people sooner and that changes bookings or collected revenue. Better summaries matter only if staff use them and reduce handling time or errors. A generated blog post is not valuable because it exists; it must attract qualified traffic, support a buyer, or reduce a real content cost without creating review debt.

Assign a value score from one to five using evidence. Five means frequent, materially expensive or revenue-linked, and likely to change an outcome. One means rare, trivial, or impossible to connect downstream. Record the calculation and uncertainty. A transparent rough estimate is more useful than a precise-looking vendor ROI calculator built on hidden assumptions.

A strong candidate has inputs with recognizable structure and enough representative examples to test. The output has an acceptance standard: fields match the document, the summary contains required facts, the category agrees with a human label, or the draft uses only approved information. If experts cannot agree on what good looks like, the model cannot be evaluated meaningfully.

Review data rights and quality before access. Customer messages may contain personal or sensitive information. Vendor contracts, employee records, payments, and recordings deserve strict treatment. Use sanitized examples for early design. Decide which system is authoritative, how updates occur, and whether the vendor stores or trains on inputs. A task that requires exporting the whole customer database to an unreviewed tool is a poor first experiment.

Fast verification is powerful. A human can compare extracted fields with one page, accept or reject a classification, or edit a short draft. It is harder to verify a strategic plan, nuanced legal conclusion, or autonomous conversation whose effects emerge weeks later. Favor tasks where review cost is lower than creation cost and errors are visible before consequences occur.

Risk-adjust the upside. Errors involving money, safety, legal rights, privacy, employment, eligibility, vulnerable people, or public claims can outweigh routine savings. Consequence and detectability both matter. An internal draft reviewed by an employee has lower exposure than an outbound message sent instantly. A wrong internal category that is corrected the same day differs from a false booking confirmation that changes a customer's plans.

Estimate exception rate and complexity. If half of cases require expert judgment, an autonomous workflow may create a review queue without removing much work. Assistance can still help by gathering context, but the business case must count human handling. Design the exception path before the happy path and confirm someone has capacity to own it.

Integration burden includes authentication, API limits, field mapping, duplicate handling, retries, monitoring, vendor changes, and security review. A task that looks like a one-hour prompt may become a multi-system software project. Give reversibility a high score: shadow mode, human review, limited cohort, and easy rollback reduce the cost of being wrong.

The first maturity level is observation: classify historical cases, summarize without acting, or compare model recommendations with what people did. This creates a test set and shows disagreement. The second level drafts or recommends while a person decides. The third automates low-risk, high-confidence cases and escalates the rest. Higher autonomy should follow measured reliability and operational capacity, not a vendor plan level.

Research results are consistent with an augmentation-first posture. In the NBER support study, agents could use, change, or ignore suggestions, and gains were concentrated among less experienced workers. In the professional writing experiment, participants completed bounded writing tasks with AI assistance; the work did not demand the precise proprietary context many real decisions require. These studies are evidence that task fit matters, not that every task should be delegated.

Design the human role deliberately. Reviewers need the source, proposed output, reason, confidence or rule state, and ability to correct the system. Capture edits as feedback, but do not assume every edit is ground truth. Periodically analyze whether reviewers rubber-stamp, disagree, or create new bottlenecks. Human-in-the-loop must improve control, not supply a decorative approval click.

Before integration, assemble a sanitized test set with normal cases, edge cases, failures, and adversarial inputs. Define accuracy and guardrails by category. Run the candidate model or process manually, record outputs, and have qualified reviewers score them without knowing which version produced each output where feasible. Estimate review time and exception rate. This may disprove the idea before a contract or build.

If offline results are promising, shadow the live workflow without affecting customers. Compare classifications, summaries, or recommended actions with actual outcomes. Then pilot one low-risk cohort with human approval. Use a baseline or holdout and count full costs. Set a decision date and kill criteria. A pilot that continues indefinitely is an ungoverned production system.

The best first process often looks ordinary: triaging a shared inbox, extracting fields from a common document, summarizing calls into approved CRM fields, drafting routine responses for review, or prioritizing a callback queue. These tasks can improve operating leverage while generating proprietary labeled data. Start there, learn how the organization operates AI safely, and earn the right to consider higher-consequence workflows.