What Business Processes Should Be Automated First?

The best process to automate first is rarely the most impressive one.

It is usually the process your team performs repeatedly, understands well, can describe with reasonably clear rules, and wishes it did not have to execute manually every day.

That distinction matters because automation projects often begin backward. A company discovers an AI agent, workflow platform, chatbot, or automation tool and then starts searching for somewhere to use it. Research on robotic process automation has found that process selection is central to successful automation, with suitable candidates tending to be rule-based, mature, digitally accessible, repetitive, relatively low in variation, and frequent enough to justify the effort. Selecting the wrong process can undermine the value of the automation itself.

The more useful question is therefore not:

What can we automate?

It is:

Which process creates enough repetitive work, delay, error, or lost opportunity that automating it would produce a measurable business improvement?

For a small or midsize business, the answer often points toward lead handling, scheduling, routine follow-up, CRM administration, information transfer, document processing, recurring reporting, and other structured operational work—not toward immediately automating the hardest decisions in the company.

The Best First Automation Has a Specific Profile

Research on business-process and robotic-process automation is surprisingly consistent about what makes a strong candidate. A 2023 peer-reviewed study on process selection identified common characteristics including clear rules, process maturity, structured and digital data, low reliance on error-prone human input, relatively low complexity, interaction across multiple systems, high frequency, and meaningful process duration. A later implementation framework similarly found that practitioners commonly look for frequent, standardized processes with relatively few exceptions and a favorable economic case when choosing early automation projects.

More recent research published in 2026 reinforces the warning on the other side: poor-quality data, weakly defined rules, unstandardized workflows, and processes that have not been adequately modeled or understood can make automation unstable or unreliable.

In practical terms, a strong first candidate generally has six characteristics:

This leads to an important principle:

The process does not need to be completely mindless. Modern AI can work with less-structured information in ways that conventional rule-based automation cannot. But the greater the ambiguity and consequence of the decision, the more important human review, testing, and governance become. NIST's AI Risk Management Framework explicitly recommends evaluating AI according to its context of use, expected impact, measurement requirements, and associated risks rather than treating every application as equally appropriate for automation.

Do Not Confuse Automation with AI

This is one of the most important distinctions for a small business.

Some processes need ordinary automation, not artificial intelligence.

If the instruction is: “When a qualified lead submits this form, create the contact, assign the salesperson, create a follow-up task, and send an acknowledgment.” — that is predominantly a deterministic workflow.

If the instruction is: “Read this prospect's free-form message, determine which service they appear to need, summarize the request, extract important details, and prepare a personalized follow-up draft.” — AI may add useful capabilities because the input is less structured.

And if the task is: “Decide whether this applicant should be hired, determine whether this patient's symptoms require a particular medical intervention, or make a high-stakes decision from incomplete information” — the question is no longer merely whether the technology can automate the process. Risk, legal obligations, fairness, accuracy, and appropriate human responsibility become central considerations. NIST's AI guidance emphasizes risk management and evaluation throughout the AI lifecycle, while the EEOC has specifically warned that automated employment tools remain subject to federal anti-discrimination law.

The goal is not to maximize the amount of AI in the company. The goal is to assign the right kind of technology—and the right amount of human involvement—to each step.

The Business Processes Usually Worth Evaluating First

There is no universal list that fits every company. A dental practice, law firm, contractor, consulting firm, and e-commerce company have very different operating models. But when the process-selection criteria above are applied to a typical service-oriented small business, several categories repeatedly emerge as strong candidates.

Lead Capture, Routing, and First Response

A prospect fills out a website form. Someone receives an email. Someone else copies the information into the CRM. A manager decides who should receive the lead. The employee sends an acknowledgment. A reminder is created. The owner eventually checks whether anyone followed up.

That is not one task. It is a chain of connected tasks, and much of that chain can often be standardized.

The most valuable automation is not necessarily sending an email. It is ensuring that the lead reliably enters a process and does not disappear between systems. If lead-response delays are part of the problem, you can estimate potential revenue leakage using the Blkfriars calculator.

This is particularly relevant because current U.S. business AI adoption remains concentrated in a relatively limited number of functions. U.S. Census Bureau research based on the 2026 Business Trends and Outlook Survey found that 18% of firms were using AI in a business function during its November 2025–January 2026 reference period; among AI-using firms, 57% used it in three or fewer business functions, with sales and marketing the most commonly reported function at 52%.

That does not prove that every sales workflow should be automated. It does suggest that targeted, function-specific adoption is much more representative of current business practice than trying to automate an entire company at once.

Appointment Scheduling and Routine Reminders

Scheduling is another useful candidate when the underlying rules are clear. A business may already know which employees perform which services, how long each appointment requires, what hours are available, which locations or service areas apply, how much preparation time is required, and when reminders should be sent.

When those conditions are defined, much of the workflow can be deterministic:

Human involvement can remain available for exceptions. A caller who simply wants the next available appointment may not require an employee to manually inspect a calendar. Someone describing an unusual, sensitive, or ambiguous situation may.

Automate the predictable path. Escalate the exception.

CRM Data Entry and System-to-System Handoffs

Manual copying is one of the least glamorous—and often most appropriate—automation targets. Suppose an employee receives information through a form, email, phone conversation, spreadsheet, or booking application and then enters substantially the same information into a CRM, project-management system, accounting tool, or internal spreadsheet.

If the information already exists digitally and the destination is known, forcing a person to move the same data from one system to another is often worth examining. Academic research on RPA suitability specifically identifies digital, structured data and interactions across multiple systems as relevant automation characteristics. More recent research likewise emphasizes that data quality and well-defined process rules are prerequisites: automation becomes less reliable when information is inconsistent or the process itself is poorly understood.

That means the right sequence may be:

rather than: automate the mess exactly as it exists today.

Repetitive Document and Administrative Work

This is where AI and traditional workflow automation often work well together. Current Census data show that writing, document analysis, and information search are among the leading generative-AI tasks being used by businesses. OECD research similarly finds that smaller businesses often begin with accessible, off-the-shelf AI applications rather than highly customized autonomous systems, while strategic integration into operations remains uneven.

Other examples include preparing meeting summaries, extracting action items, categorizing inquiries, generating first drafts, turning structured notes into standardized documents, or locating information in an approved knowledge base.

There is credible evidence that AI assistance can improve productivity for certain knowledge tasks, although the effects vary substantially by task and worker. In a field study involving 5,179 customer-support agents, access to a generative-AI assistant increased issues resolved per hour by about 14% on average; gains were substantially larger among less experienced and lower-performing workers, while the most experienced workers saw much smaller effects.

A separate randomized experiment involving 453 professionals performing occupation-specific writing tasks found that access to ChatGPT reduced task completion time by 40% and increased evaluator-rated output quality by 18%. The researchers also cautioned that the experimental tasks did not fully reproduce the contextual and factual demands of real work.

Those studies do not establish that AI will save every business 14%, 18%, or 40%. They establish something more useful: some well-defined information and communication tasks are legitimate candidates for measured AI-assisted workflow experiments, and the results should be evaluated in the actual operating environment rather than assumed in advance.

Recurring Reporting and Information Consolidation

Many companies have recurring reporting processes that look something like this: employee exports data → updates spreadsheet → combines information from another system → formats totals → writes summary → emails report → repeats next week.

The value of automating this type of workflow comes from repetition. A task that takes only 20 minutes may seem unimportant until it occurs across several employees, systems, branches, clients, or days.

Automation research consistently treats frequency and duration as important selection criteria because automation investments become more valuable when a repeatable activity consumes meaningful aggregate effort. AI may then be added selectively—for example, to explain anomalies, prepare a first-draft management summary, or categorize free-form information—while conventional software performs the predictable collection and calculation.

Routine Customer-Service Triage

Customer service is a particularly useful example of why “automation” should not mean “remove the human.” A system can often handle a predictable first layer:

Research on AI-assisted customer support provides evidence that AI can make human support work more productive in some environments rather than simply replacing the worker. For a small business, that may mean automating routine questions about hours, scheduling, service areas, appointment preparation, order status, or basic intake while preserving human intervention for complaints, unusual circumstances, emotionally sensitive conversations, negotiation, or cases in which the system is uncertain.

That model fits NIST's broader human-centered approach: the objective should be the outcome of the human-AI system, not automation for its own sake. For businesses considering an AI receptionist or revenue recovery system, the same principle applies.

What Should Not Be Automated First

The easiest automation mistake is choosing a process because it is frustrating rather than because it is ready. Some workflows should be fixed, standardized, measured, or understood before automation is introduced.

A Process Nobody Can Consistently Explain

Ask three employees to describe what happens. If you receive three substantially different answers, you may not have an automation opportunity yet. You may have a process-design problem. Research on RPA implementation identifies inadequate process understanding, inconsistent data, weak business rules, and unstandardized workflows as barriers to reliable automation. Automation is unlikely to create clarity that the organization itself has never established. Map the process first.

A Process That Changes Constantly

Automation has maintenance costs. When rules, screens, data formats, approval requirements, or business logic constantly change, the automation must change with them. Academic process-selection research therefore identifies process maturity and stability as desirable characteristics. A rapidly evolving process may still eventually be automatable, but it is rarely the safest first project.

A Low-Frequency Process with Little Business Value

Being technically automatable does not make something economically worthwhile. Suppose an employee spends two hours performing a task—but only once per year. Compare that with a six-minute process performed 30 times each business day. The second process may create much more annual opportunity despite appearing trivial in isolation. Frequency, duration, workload reduction, and a positive cost-benefit relationship repeatedly appear in automation-selection research.

A Process Dominated by Exceptions

Consider a workflow where the employee follows a standard procedure only 40% of the time and improvises the remaining 60%. That is not necessarily impossible to automate, particularly as AI capabilities improve. But it is a poor candidate for an uncomplicated first automation because the exception-handling system may become more complicated than the original task. Early projects generally benefit from standardized processes, low complexity, and relatively few exceptions.

High-Consequence Decisions That Require Judgment

There is an important difference between automating “Send the customer a reminder 24 hours before the appointment” and automating “Decide whether this person qualifies for a consequential opportunity or service.”

Generative AI can produce inaccurate or fabricated outputs, and NIST recommends risk assessment, testing, monitoring, and governance appropriate to the use case. Employment provides a concrete example. The EEOC has stated that federal anti-discrimination laws apply to AI and algorithmic systems used to make or assist employment decisions and has specifically identified technology-related employment discrimination as an enforcement priority for fiscal years 2024–2028.

That does not mean a business cannot use technology in a high-consequence process. It means those should generally not be treated like ordinary workflow automations. Appropriate legal, compliance, subject-matter, security, and human oversight may be required.

Blkfriars Practical Automation Priority Scorecard

A small business does not need a sophisticated process-mining program to create an initial shortlist. Start by identifying five to ten recurring workflows and score each one.

Do not treat the total as an automatic decision. Use it to expose the tradeoffs. A process that scores extremely well on frequency and labor burden but poorly on risk may need an AI-assisted rather than autonomous design. A process that scores well on everything except data quality may need a data-cleanup project first. A process with perfect rules but negligible frequency may not justify the implementation cost.

The most useful next question is:

What is the smallest version of this process we could automate safely and measure?

Choose the Automation Type After Choosing the Problem

A practical decision model looks like this:

That distinction is becoming increasingly important because AI adoption does not automatically mean full process autonomy. The Census Bureau's 2026 firm-level research found that 66% of AI-using businesses reported using AI solely to augment tasks, while AI-related employment decreases were reported by only about 2% of firms.

OECD research points in a similar direction. Its 2026 survey of more than 2,000 SMEs across 12 OECD countries found that off-the-shelf AI use was increasing, but strategic, targeted, secure integration into business operations remained uneven; time constraints, maintenance costs, and skills gaps continued to impede implementation.

The goal is not maximizing AI use. The goal is choosing the correct operating model.

Pilot, Measure, Monitor, and Expand

A workflow should have a baseline before it has an ROI claim. Suppose a company wants to automate lead intake. Before changing anything, measure the current process:

Then run the automation on a controlled portion of the workflow. Measure the same variables again. NIST's AI RMF centers risk management around Govern, Map, Measure, and Manage, emphasizing that systems should be evaluated against their intended context and continuously managed rather than considered finished once they are deployed. That same discipline is useful even when the workflow does not contain AI.

The relevant question is not:

Did the automation run?

It is:

Did the process improve?

A successful automation might reduce response time, remove manual touches, decrease rework, produce more complete CRM records, improve booking rates, reduce administrative hours, or make performance more consistent. A failed automation can also produce useful information. It may reveal that rules were unclear, customer data were inconsistent, exceptions were more common than expected, employees needed a different interface, or the process should have been redesigned before it was automated.

This is why starting with one contained process can be more valuable than announcing an organization-wide “AI transformation.” Current business adoption itself remains incremental: Census data collected between December 2025 and May 2026 found overall U.S. business AI use hovering around 17%–20%, while fewer than 20% of firms with four or fewer employees reported using AI. The opportunity for smaller firms is therefore not merely adopting technology faster. It is adopting it selectively enough that the first project actually works.

Human-in-the-Loop Design

The strongest automation systems do not remove humans—they put humans in the right place. The operating principle is straightforward:

NIST's AI guidance emphasizes risk management and evaluation throughout the AI lifecycle, while the EEOC has specifically warned that automated employment tools remain subject to federal anti-discrimination law. Human review does not automatically eliminate compliance risk, but it ensures that consequential decisions are made with appropriate accountability. For businesses exploring AI strategy and advisory support, governance and acceptable-use planning should be part of the roadmap from the start.

The Bottom Line

The first business process to automate should not necessarily be the process with the most advanced AI potential. Start with the process that is:

For many small and midsize businesses, that means evaluating processes such as:

before attempting to automate complex negotiations, consequential decisions, unusual customer situations, or poorly understood internal processes.

And do not assume every automation needs artificial intelligence. A good operating system may combine:

• Rules-based automation for predictable actions
• AI for appropriate unstructured information work
• People for judgment, responsibility, exceptions, and relationships

That is the real objective. Not:

How much of the business can we automate?

But:

Where is repetitive work creating enough delay, cost, error, or lost opportunity that automation would produce a measurable improvement—and where should a human remain in control?

The strongest first automation solves a real problem, produces a result that can be measured, and teaches the company enough to make the second automation better than the first.

Sources & References