What To Automate And What Still Needs A Human Touch In Your Business

There is a moment in every automation conversation where someone says, "But we can't automate that. It requires a human."

They are usually right. And they are usually wrong about which "that" they are referring to.

The hardest part of automation is not the technology. It is drawing the line. Automate too little and you are leaving money on the table. Automate too much and you break things, client relationships, compliance obligations, the judgment calls that make your business actually work.

I have spent years helping businesses in regulated industries figure out where that line is. Insurance agencies, mortgage brokerages, financial advisory firms, title companies. The pattern is remarkably consistent. They over-rely on humans for work that machines should handle and under-invest in the human moments that no machine can replicate.

Here is how to think about it clearly.

The Automation Spectrum#

Most people think about automation in binary terms. Automated or not automated. That is the wrong frame.

Every task in your business sits somewhere on a spectrum from fully deterministic to fully judgment-dependent. The trick is mapping each task accurately, then matching the right level of automation to the right level of human involvement.

Fully deterministic tasks have clear inputs, clear rules, and clear outputs. If condition A, then action B. No ambiguity. No exceptions. These should be fully automated. No human in the loop.

Rules-based tasks with known exceptions follow a pattern most of the time but occasionally hit a case that does not fit. These should be automated with escalation paths. The machine handles the 90 percent. The human handles the 10 percent.

Judgment-intensive tasks with data support require a human to weigh options, consider tradeoffs, and make a call. But the human makes a better call when they have organized, accurate data in front of them. These should be human-led with AI-assisted preparation.

Relationship-sensitive tasks are about trust, empathy, and nuance. The value is in the human connection itself. These should stay human, full stop.

McKinsey's 2024 State of AI report found that 65 percent of organizations are now regularly using generative AI in at least one business function, but the most successful deployments are those that clearly define which tasks are automated and which remain human-led.[1] The failures almost always come from blurring that line.

What To Automate: The Deterministic Work#

Let me be specific. Here are the categories of work that should be automated in nearly every regulated business.

Data Movement Between Systems#

Your CRM says one thing. Your commission system says another. Your accounting software says a third thing. Someone on your team spends hours making sure they all match.

This is the definition of deterministic work. Data exists in system A. It needs to exist in system B. The rules for how it maps are known. There is no judgment involved.

Gartner predicts that by 2026, 75 percent of organizations will have adopted some form of automated data integration, up from 35 percent in 2022.[2] The businesses that still move data manually by that point will be at a serious competitive disadvantage.

Automate data movement. All of it. Between your CRM and your management system. Between your commission platform and your accounting software. Between your document management system and your compliance tracker. If data is moving from point A to point B based on known rules, a human should not be touching it.

Reminders and Follow-Ups#

How much of your team's day is spent sending reminders? "Hey, we still need that document." "Just checking in on the status of this application." "Reminder that this deadline is in three days."

These are not conversations. They are notifications. They follow a pattern: a condition is met (X days since last activity, document still missing, deadline approaching), and a message is sent. The content of the message is templated. The timing is rules-based.

The World Economic Forum's Future of Jobs Report 2023 identifies administrative and secretarial tasks, including routine correspondence, as among the roles most likely to be augmented by automation in the next five years.[3] Reminders and follow-ups are the lowest-hanging fruit.

Automate every reminder and follow-up that follows a predictable pattern. Use templates with dynamic fields. Trigger them based on system events, not human memory. Let your team spend their communication time on conversations that actually require thought.

Status Updates#

"What's the status of the Smith application?"

Someone on your team looks it up in one system, checks a second system for additional context, and types a response. This happens dozens of times a day in most businesses I work with.

Status inquiries are pure information retrieval. The data exists. It needs to be delivered to the person asking. No judgment required.

Automate status visibility. Build dashboards. Send proactive notifications when statuses change. Give clients and partners self-service access to status information where appropriate. Every status inquiry your team handles manually is a signal that your systems are not communicating well enough.

Document Routing and Classification#

When a document arrives, an email attachment, a fax, an uploaded file, someone has to look at it, figure out what it is, and put it in the right place. Application? Route to processing. Amendment? Route to servicing. Claim? Route to claims.

The NIST AI Risk Management Framework provides guidance on deploying AI for classification tasks, noting that document classification is among the highest-confidence use cases for current AI technology.[4] The accuracy rates for document classification AI now exceed human accuracy for well-defined document categories.

Automate document routing. Train classification models on your document types. Set confidence thresholds, if the AI is 95 percent confident it is an application, route it automatically. If confidence is below the threshold, escalate to a human for classification.

Scheduled Reports and Compliance Filings#

If you generate the same report every week, month, or quarter, and the report follows the same format and pulls from the same data sources, that report should generate itself.

The SBA Office of Advocacy reports that regulatory compliance consumes significant resources for small businesses, with routine filings being a major component.[5] Automate report generation. Automate compliance filing preparation. Have a human review the output before it is submitted, but do not have a human build the output from scratch every time.

What Needs A Human: The Judgment Work#

Now here is the other side. These are the tasks where automation should support but not replace human involvement.

Exception Handling#

Exceptions are, by definition, the cases that do not fit the rules. A commission payment that does not match any expected pattern. A client request that falls outside normal parameters. A compliance situation that requires interpretation of ambiguous regulations.

Gartner's customer service research found that while AI can resolve up to 40 percent of customer service interactions without human involvement, the remaining 60 percent, the exceptions, the complex cases, the emotionally charged situations, require human judgment and often determine whether a customer stays or leaves.[6]

In your business, exceptions are where your team's expertise creates the most value. Do not automate exception handling. Instead, automate the work around exception handling:

• Automatically identify and flag exceptions.
• Automatically gather the context needed to resolve them.
• Automatically route them to the person with the right expertise.
• Let the human focus on the judgment call, not the research.

Tradeoff Decisions#

Many business decisions involve tradeoffs where there is no objectively correct answer. Should you push back on a carrier's commission schedule or accept it to maintain the relationship? Should you prioritize speed or thoroughness on a complex application? Should you take on a client whose profile is borderline for your risk appetite?

These decisions require weighing multiple factors, considering long-term implications, and applying business judgment that reflects your company's values and strategy. AI can provide data to inform these decisions. AI should not make them.

McKinsey's research notes that the most effective AI deployments augment human decision-making rather than replacing it, particularly in complex domains where context and relationships matter.[1] Give your team better data. Do not give them less authority.

Relationship-Sensitive Communication#

There are conversations where the human element is the entire point. A client who just suffered a loss needs empathy, not efficiency. A producer who is underperforming needs a nuanced conversation, not an automated performance review. A strategic partner considering a deeper relationship needs to feel your commitment, not your technology.

The Salesforce SMB Trends report found that 79 percent of small business customers say the experience a company provides is as important as its products or services.[7] For regulated industries where trust is the foundation of every relationship, this number is probably higher.

Keep relationship-sensitive communication human. Always. This includes:

• Breaking bad news to a client (claim denied, application declined, rate increase).
• Negotiating with carriers or partners on complex deals.
• Handling complaints or disputes that involve emotional stakes.
• Onboarding high-value clients where first impressions set the tone.
• Any communication where the client's trust in your organization is at stake.

Regulatory Interpretation#

Regulations are written by lawyers for lawyers. Interpreting how a regulation applies to a specific situation often requires judgment about intent, context, and risk tolerance.

The NIST framework specifically calls out regulatory interpretation as a domain where AI should assist but not decide, noting the risk of AI systems that appear confident in regulatory conclusions without adequate understanding of context.[4] Your compliance team should use AI to research regulations, find precedents, and organize relevant guidance. But the interpretation, the decision about what the regulation means for this specific situation, should be human.

Designing The Handoff#

The hardest part of the automation-versus-human question is not deciding which bucket each task falls into. It is designing the handoff between the two.

Most automation failures happen at the boundary. The machine does its part, but the handoff to the human is clumsy, incomplete, or confusing. The human gets a notification without context. Or they get too much context and have to sort through irrelevant information. Or the handoff happens at the wrong time in the wrong channel.

Here is how to design handoffs that work.

Rule 1: The Machine Summarizes, The Human Decides#

When work escalates from automated to human, the handoff should include a concise summary of what happened, what the exception is, and what options are available. Do not dump raw data on the human. Do not force them to re-research the situation. Give them what they need to make a decision, and nothing more.

Rule 2: Escalation Criteria Are Explicit#

Write down the specific conditions under which work moves from automated to human. "When the confidence score is below 85 percent." "When the transaction amount exceeds $10,000." "When the client has had a complaint in the last 90 days."

Vague escalation criteria like "when it seems complicated" are useless. Your rules should be specific enough that you could explain them to a new employee on their first day.

Rule 3: Humans Can Override But Cannot Be Bypassed#

Give humans the ability to override automated decisions. If the system routes something automatically but the human disagrees, the human wins. But do not allow processes to bypass automation entirely. If someone starts manually handling work that should be automated, you lose the consistency and efficiency gains.

Deloitte's Tech Trends research emphasizes that the most successful automation implementations maintain clear governance over which processes are automated and ensure that manual workarounds do not gradually re-emerge.[8] This is a management discipline, not just a technology decision.

Rule 4: Track Handoff Volume and Quality#

Measure how many items escalate from automated to human. Measure how long humans take to resolve escalated items. Measure how often humans override automated decisions. These metrics tell you whether your automation boundary is in the right place.

If too many items escalate, your automation rules are too conservative. If humans frequently override automated decisions, your rules need updating. If escalated items take a long time to resolve, the handoff context is not sufficient.

The Framework In Practice#

Let me walk through how this applies to a real process. Take commission reconciliation in an insurance agency.

Fully automated (no human):

• Download commission statements from carrier portals.
• Parse statement data into a standardized format.
• Match statement line items to expected commissions in your management system.
• Flag matches as reconciled. Update records.

Automated with human approval:

• Identify line items that match within a tolerance (e.g., amount is within 2 percent of expected). Present these to a human for batch approval.
• Identify line items where the commission rate differs from the contracted rate. Flag these with the contract terms attached. Human reviews and either approves the variance or initiates a dispute.

Human-led with data support:

• Commission payments that cannot be matched to any expected record. The system gathers available context (carrier, policy type, date range, amount) and presents it to the human for research.
• Patterns of short-pays or missing payments from specific carriers. The system identifies the pattern. The human decides whether to escalate.

Fully human:

• Calling a carrier to dispute a systematic commission error. This is a relationship and negotiation task.
• Deciding whether to terminate a carrier relationship over persistent payment issues. This is a strategic business decision.

This is not complicated. But it requires deliberate design. Most businesses skip the design step and either automate everything (and break things) or automate nothing (and waste time).

Common Mistakes#

Let me flag the patterns I see most often.

Automating the visible stuff and ignoring the important stuff. Businesses often automate the processes that are most annoying rather than the ones that cost the most. Annoying and expensive are not the same thing. Prioritize by business impact.

Keeping humans in the loop "just in case." If a process is truly deterministic and you have validated the automation, remove the human review step. Every unnecessary human touchpoint adds latency and cost. If you do not trust the automation, fix the automation. Do not add a human as a safety blanket.

Automating the decision but not the preparation. Some businesses automate the easy part (moving data) but leave the hard part (finding the right data for a decision) manual. Flip that. Automate the preparation, the research, the context gathering, the data assembly, and let the human focus on the actual decision.

Failing to update the boundary. The right automation boundary today is not the right boundary in six months. As your data improves, your AI learns, and your processes mature, tasks that used to require human judgment become more deterministic. Review and adjust the boundary quarterly.

The World Economic Forum predicts that 44 percent of workers' core skills will need to change in the coming five years, with the shift being toward analytical thinking, creative thinking, and technology literacy.[3] The boundary between what machines do and what humans do is moving. Your business needs to move with it.

Building Your Automation Map#

Here is a practical exercise. Take 30 minutes and do this for your business.

1. List your top 10 processes by volume. How many times per week or month does each one occur?
2. For each process, list the steps. Be granular. Not "process the application" but "receive application, check for completeness, enter data into system, verify data, route for underwriting review."
3. Classify each step. Deterministic, rules-with-exceptions, judgment-intensive, or relationship-sensitive.
4. Draw the line. For each process, identify where automation should handle the work and where humans should.
5. Design the handoffs. For each transition from automated to human, define the escalation criteria and the context that should be provided.

This map becomes your automation roadmap. It tells you what to build first, what to build next, and what to leave alone.

The Goal Is Not Less Human#

I want to be clear about something. The goal of drawing this line is not to remove humans from your business. It is to put humans where they create the most value.

Your best people should not spend their days copying data between systems, sending reminder emails, and looking up statuses. They should spend their days handling the complex situations, making the judgment calls, and building the relationships that drive your business forward.

Gartner's research forecasts that by 2027, AI-augmented workers will outperform their non-augmented peers by 35 percent on tasks that combine analytical and creative skills.[2] The future is not humans versus machines. It is humans with machines versus humans without them.

The businesses that draw the line well, that automate ruthlessly where automation fits and invest deeply in human capability where it matters, are the ones that will win in the next decade. Not because they have better technology. But because they use their people better.

That is the whole point. Automation is not the goal. Better use of human talent is the goal. Automation is just how you get there.

If you want help drawing the automation line in your business, we work with regulated companies to design processes that put machines and humans exactly where they belong. Book a discovery call at thelobbi.io/discovery.#