- >
- Workflow>
- What Processes Should Not Be Automated with AI?
What Processes Should Not Be Automated with AI?
The processes that should not be automated with AI depend on human empathy, high-stakes judgment, or original strategy. Fixed if/then rules belong in deterministic automation, and a broken workflow needs repair before anything touches it.
TL;DR
- Emotionally sensitive conversations with customers, patients, and employees stay with people who offer real empathy.
- For legal, medical, financial, and ethical calls, a named human decides and AI supports.
- Strategy, vision, and original creative work stay human-led, since AI remixes existing patterns.
- Fixed if/then logic belongs in deterministic rules, which cost less and behave predictably every time.
- Test each step: does it have a rule, an owner, an audit trail, and a reason to exist?
Automating with AI means handing a task to a probabilistic model. The model predicts each output from statistical probabilities, so one input can return a different result on each run. Deterministic automation is different: it executes a written rule and gives the same answer every time. Confusing the two is where most bad automation decisions start. Knowing when not to use AI workflow automation begins with that distinction.
A quarterly review ends, and leadership asks the operations head which workflows AI will take over next. The honest answer is never a list of whole processes. This guide covers seven categories of work to keep with people or plain rules. It adds a reference table and a per-step test. The test sorts each part of a process to a model, a rule, or a named human.
Which high-stakes decisions should stay with people?
Legal, medical, financial, and ethical decisions need a person who answers for the outcome. A model has no license to lose, no court to face, and no patient to explain itself to. It cannot carry human accountability for a wrong call.
The risk is specific. A probabilistic model produces a fluent, confident answer whether it is right or wrong. In medical diagnosis, legal review, credit decisions, or a fairness-versus-cost policy call, a confident error does real harm. Choices about who benefits and who bears the cost are value judgments, and a model only matches patterns.
What to do: Use AI as a decision-support tool. Let it summarize the case file, flag inconsistencies, and surface similar past matters. A qualified person then reviews, decides, and signs, and the record carries their name.
Why should emotionally sensitive conversations stay human?
Any conversation that depends on empathy and emotional intelligence belongs to a person. That covers a clinician delivering hard news and a manager handling a layoff. It also covers an account lead calming an upset client and an HR partner hearing a harassment complaint.
AI produces sympathetic phrasing without understanding the person in front of it. People notice. Once someone realizes a script is responding to their distress, their trust in the organization drops. Relationship-building work, the kind that keeps a key client or a valued employee, rests on that trust.
What to do: Let AI triage the inbox, route the case to the right owner, and draft background notes. The conversation itself, and every promise made in it, stays with someone who will still be there next week.
Should AI handle strategy, vision, and creative work?
Strategy, vision, thought leadership, and brand voice stay human-led, because AI remixes existing patterns. A model generates the most statistically likely next idea. That helps with variations and fails at a breakthrough no dataset contains yet.
Corporate direction is a bet on a future nobody has written down. Original creative work, such as the campaign concept or product idea competitors haven’t considered, follows the same logic. A model fed typical inputs returns predictable ideas.
What to do: Use AI for research, first drafts, and quick variants of an idea you already own. A person sets the direction, chooses which concept goes forward, and protects the voice that makes the brand recognizable.
Can AI make hiring and promotion decisions fairly?
Hiring, promotion, and employee monitoring should not run on AI alone. A model trained on past decisions reproduces their bias. If yesterday’s shortlists skewed toward one profile, the model learns to score it as “good.”
Hiring bias of this kind is hard to spot from inside the tool. The output looks objective, the scoring seems consistent, and nobody owns the reasoning. Surveillance tools that score employee activity carry the same flaw and add a privacy cost.
What to do: Define structured criteria before screening, and have a human panel make every decision. Let AI handle scheduling, interview summaries, and reminders. Audit outcomes by group on a fixed schedule, and keep the name of whoever approved each hire or promotion.
Do fixed-rule tasks need AI at all?
If you can write the rule as if/then, you need deterministic automation, not AI. Rule-based automation costs less to run and behaves identically every time. A reviewer can read its audit trail without a data scientist in the room.
Putting a model on a fixed rule is over-engineering. You pay for inference and accept variance you didn’t need. You also lose the ability to explain why a record went where it did. All of these are rule work:
- Routing a request by amount or department.
- Sending a reminder after 48 hours.
- Syncing a field between two systems.
- Opening a ticket when a form arrives.
What to do: Build these with workflow rules, scripts, APIs, or webhooks. Save AI for steps where the input is unstructured.
Should you automate a broken or poorly mapped workflow?
No. Automating a broken process only accelerates mistakes. Suppose three people handle the same request in different orders. A handoff lives in someone’s head, or a step exists because nobody remembers why. Automation makes those faults run faster and at greater volume.
Teams that automate chaos end up debugging the automation when the real defect is the process underneath. Redundant approvals simply run faster. Unmapped exceptions fall through at machine speed.
What to do: Map the workflow as it actually runs. Agree on the order of steps and who owns each one, then cut anything that adds no value. Some processes should simply be eliminated. Once the steps are stable and written down, decide which ones a rule or a model should take.
What about sensitive and confidential data?
Confidential customer, employee, financial, and health data should not go into public AI tools. Once pasted into a consumer chatbot, the information sits outside your access controls and retention policies. You cannot pull it back.
Privacy is only half the issue. Data that encodes protected characteristics carries bias into whatever the model produces next. A privacy risk becomes a fairness problem.
What to do: Approve named AI tools for each data class, and redact identifiers before anything reaches a model. Keep regulated data in your systems of record. Give everyone who touches the data a role with defined permissions. Log access so you can show an auditor who saw what.
Should you automate, augment, or keep the work human?
Most work lands in one of three buckets: automate with rules, augment with AI, or keep it human. Harvard Business School Online sorts the first two by frequency and value. High-frequency, low-value tasks suit full automation. Lower-frequency or higher-value tasks involving risk, ethics, or customer relationships benefit from augmentation. In that model, AI assists a person who still owns the result.
| Kind of work | Approach | Reason |
|---|---|---|
| Fixed routing, reminders, field sync | Automate with deterministic rules | Same input gives the same output, with a full audit trail |
| Pulling fields from unstructured email or PDFs | Augment with AI, person checks flagged fields | Pattern work suits models, and errors stay visible |
| Drafting, summarizing, background research | Augment with AI | Faster first drafts, and a person edits |
| Legal, medical, financial, ethical calls | Keep human, with AI as decision support | Someone must answer for the outcome |
| Sensitive conversations | Keep human | Empathy is not a script |
| Strategy, brand voice, original ideas | Keep human-led | Models remix what already exists |
| Hiring and promotion | Keep human, with structured criteria | Training data carries past bias |
| Broken or unmapped workflows | Fix or eliminate first | Speed multiplies the error |
Use the table as a first sort. Then run any process that spans several rows through a per-step test.
How do you decide which steps of a process to automate?
Ask the automation question one step at a time. Most processes mix three kinds of work: rules, patterns, and judgment. A deterministic system runs the rules, AI handles the patterns, and a named person owns the judgment. Category lists stop at whole processes, which is why teams either over-automate or freeze.
What questions sort each step?
- Can you write this step as an if/then rule? If yes, a deterministic rule runs it.
- Who is accountable if this step goes wrong? If the answer is a named person, AI supports their decision.
- Can you audit what happened afterward, including who acted and when? If not, the step isn’t ready for automation of any kind.
- Is the step worth keeping at all? Eliminate it before you automate it.
- The test sorts steps but does not prove a model is accurate enough. Only a trial on your own data shows that.
What does the split look like in an invoice approval?
Take a vendor invoice arriving by email.
- Reading the invoice is pattern work. AI pulls the vendor name, amount, PO number, and due date from an unstructured email or PDF. It flags low-confidence fields for a person to check.
- Routing is rule work. An invoice over US$5,000 goes to the department approver. A match against the PO moves to payment, and a mismatch becomes an exception. A deterministic rule does this identically every time.
- The exception decision is judgment. A disputed amount, a missing PO, or a vendor under review goes to a named person. That person approves, rejects, or escalates. The record keeps who decided, what they chose, and when.
AI saves the typing and rules cut the waiting. A person stays in the lead on the one decision that carries risk. In practice, that person makes the call rather than clicking through a queue of model outputs.
Where does Kissflow fit in a split process?
Most tools automate one workflow at a time, and chatbots answer single prompts. Kissflow works with large enterprises to redesign how work gets done, builds it on its own agentic platform, and stays with the change until the customer’s team owns it.
Kissflow starts from the Procure-to-Pay outcome your finance team owns and redesigns every approval in it. On the platform, agents read each invoice, check it against the PO, and recommend a route. A named approver decides every exception. Role-based access, data controls, and an audit trail keep every decision traceable. The platform connects to your ERP, so each result posts back to the system of record. In year one, Kissflow helps you set up the capability. From year two, your team runs it.
Frequently asked questions
What’s the difference between AI and automation?
Deterministic automation runs a written rule and gives the same result every time. AI hands the task to a probabilistic model, so one input can produce different results. Use rules for fixed logic and AI for unstructured input.
Is automation possible without AI?
Yes. Deterministic automation works from fixed if/then rules. Examples include routing a request by amount, sending a reminder after 48 hours, or syncing a field between two systems. It costs less to run than AI and gives the same result each time. A reviewer can follow its audit trail. AI earns its cost only on unstructured input, such as emails, PDFs, or scanned forms.
What things cannot be done by AI?
AI cannot carry accountability for a legal, medical, financial, or ethical decision. It has no license to lose and no court to face. It also cannot offer real empathy, only sympathetic phrasing. AI cannot originate strategy or ideas that no dataset contains yet, since it remixes existing patterns.
Is automation better than AI?
For fixed-rule tasks, yes. Deterministic automation costs less, behaves predictably, and is easier to audit than a model. AI is the better fit for unstructured input such as emails and PDFs. Most real processes combine the two, and risky calls still go to someone who answers for them.
What are the limitations of using artificial intelligence?
AI output is probabilistic, so errors arrive sounding as confident as correct answers. Models trained on past decisions repeat the bias in them. Data pasted into public AI tools leaves your access controls. Each limit means a person checks AI work wherever an error is costly.
Who is accountable when an AI-assisted decision goes wrong?
A named human should be. For legal, medical, financial, and ethical calls, AI works as decision support. A qualified person reviews, decides, and signs, with their name on the record. The system should also log when they acted, so an auditor can trace each call.
Tell us which process you want to redesign, and we will map who decides at each step.