Which business processes should you automate with AI first?
Published: Oct 7, 2026 · Last updated: Oct 7, 2026
Automate first the processes where a person reads messy input, makes a routine judgment and types the result into another system. Lead triage, support ticket routing, invoice extraction and first-pass contract review are good examples. Pick one that runs at high volume, has a clear right answer most of the time and does little damage when the AI is wrong, because a person checks the output.
Don't start with a process where one mistake can reach a customer, a regulator or a bank account without review. You can automate those later, once you have logs, accuracy numbers from your own data and a team that trusts the system.
Below are 12 processes across six departments. For each one you get what the AI does, what a person keeps, the data you need and a readiness test you can run this week. The last two sections cover when plain rule-based automation is the better choice and how to turn a pilot into a working system.
How to read each example
Each example answers the same four questions. These are the questions that decide whether a project ships or gets stuck as a pilot.
What the AI does: the narrow task the model performs. This is usually reading, classifying, extracting or drafting.
What stays human: the decisions or actions a person keeps, at least for the first few months.
Data needed: what must already exist and be accessible before you build.
Readiness test: a quick check that shows whether the process is ready now or needs groundwork first.
One rule covers every example: if you can't write down what a good output looks like for 20 real past cases, the process isn't ready for AI. Document it first.
Most of these are single AI steps inside a normal workflow, not autonomous agents. If you are unsure which one you need, our explainer on what an AI agent is covers the difference.
Sales and marketing: 3 examples
1. Inbound lead qualification and routing
What the AI does: reads form submissions and inbound emails and pulls out the company, the person's role, what they need and their timeline. It scores fit against your ideal customer rules and sends the lead to the right owner with a one-paragraph summary.
What stays human: the first real conversation, and any decision to disqualify a lead that looks large or unusual.
Data needed: a written ideal customer profile, a few hundred past leads labelled won, lost or junk, and API access to your CRM.
Readiness test: have two salespeople score the same 30 old leads. If they disagree on more than a handful, fix the definition before you build anything.
2. Turning call notes into CRM updates
What the AI does: turns call transcripts into CRM fields: next step, decision makers mentioned, objections and competitor names. It also drafts the follow-up email.
What stays human: sending the email, and changing the deal stage or forecast.
Data needed: calls recorded with consent, a transcription tool and a CRM with clearly defined fields.
Readiness test: check whether your team actually uses the CRM fields. If half of them are empty today, the AI will just fill fields that nobody reads.
3. Repurposing content
What the AI does: turns a webinar, a project write-up or a long article into draft social posts, email snippets and FAQ answers in your brand voice.
What stays human: fact checks, any claim about results, and the decision to publish.
Data needed: a style guide, a list of claims you are allowed to make and approved source material.
Readiness test: give your style guide to a new freelancer. If their work isn't usable, the model's won't be either.
Customer support: 2 examples
4. Ticket triage, tagging and priority
What the AI does: reads incoming tickets in any language and assigns category, product area, urgency and sentiment. It also spots duplicates and sends each ticket to the right queue.
What stays human: escalations, refunds and anything involving account access, legal threats or safety.
Data needed: several months of past tickets with the tags your agents applied, plus a list of categories you actually use.
Readiness test: pull 100 old tickets and check how consistently your agents tagged them. If past tags are inconsistent, the AI will copy that inconsistency. Clean up the categories first.
5. Drafting replies from your help docs
What the AI does: finds the help articles and resolved tickets that match the question and drafts a reply that cites them. When no source answers the question, it says so.
What stays human: an agent reviews and sends every reply. Later, auto-send can be allowed for a short list of low-risk categories such as order status or password reset links.
Data needed: an up-to-date knowledge base. This is mostly a retrieval problem, and our RAG pipeline guide explains how to set it up.
Readiness test: search your help centre for your 10 most common ticket reasons. If fewer than 7 have a current article, write the articles before you automate.
Finance and operations: 3 examples
6. Extracting data from invoices and receipts
What the AI does: reads PDFs, scans and photos in different layouts. It extracts the vendor, dates, line items and tax fields, then matches them to purchase orders.
What stays human: approving payment, resolving mismatches and setting up new vendors.
Data needed: sample documents from your main vendors, your chart of accounts and API access to your accounting system.
Readiness test: pull 50 invoices from last quarter. If most come from a few vendors with fixed layouts, simple template extraction may be enough and you may not need AI. The section below on plain automation covers this.
7. Checking expenses against policy
What the AI does: compares expense claims and receipts with your written travel and expense policy. It flags items that break the policy and quotes the reason and the clause.
What stays human: approvals, and any conversation with the employee.
Data needed: a written policy with specific limits, and your claim history.
Readiness test: give the policy and 10 past claims to someone outside finance. If they don't reach the same verdicts as your finance team, the policy is too vague for a model to apply.
8. Handling order and delivery exceptions
What the AI does: reads carrier updates, customer emails and order notes, and classifies the problem: wrong address, damaged item, delay or missing part. It drafts the customer message and suggests the next action.
What stays human: refunds, reshipments above a set threshold and disputes with carriers.
Data needed: access to your order system, carrier status feeds and a list of exception types with standard responses.
Readiness test: ask your operations lead to list every exception type and its standard fix on one page. If they can't, you don't have a process to automate yet.
HR, legal and internal knowledge: 3 examples
9. Summarising candidate applications
What the AI does: summarises each application against the must-have criteria in the job description. It shows the evidence for each criterion, notes gaps and drafts screening questions.
What stays human: every rejection and every shortlist decision. Never let a model reject candidates automatically. Hiring decisions carry legal and bias risk in many places.
Data needed: clearly written job criteria and access to your applicant tracking system.
Readiness test: check whether your must-haves are facts someone can verify, like "has shipped a mobile app to production", rather than adjectives like "strong communicator".
10. First-pass contract review
What the AI does: compares incoming contracts with your playbook. It flags clauses that differ from your standard positions on liability caps, termination, IP ownership and payment terms, and suggests fallback wording.
What stays human: all negotiation and sign-off, done by a qualified person.
Data needed: a clause playbook listing your preferred, acceptable and unacceptable positions, plus past signed contracts.
Readiness test: if the playbook exists only in your lawyer's head, write it down first. That document is worth having even if you never automate.
11. Internal knowledge assistant
What the AI does: answers staff questions about policies, processes and product specs from your internal documents and links to the source. When the documents don't cover something, it says it doesn't know.
What stays human: each document's owner keeps it current, and HR or legal answer sensitive questions directly.
Data needed: documents stored in a small number of accessible systems, with working permissions.
Readiness test: check permissions. If the assistant could show salary sheets or board papers to anyone who asks, fix access controls before you connect it.
Product and engineering: 1 example
12. Triage for bug reports and feedback
What the AI does: groups support feedback, app store reviews and bug reports into themes and merges duplicates. It pulls in reproduction details such as device and app version, and drafts a tracker issue with steps and a suggested severity.
What stays human: priority, roadmap decisions and closing issues.
Data needed: access to your issue tracker and feedback sources, plus written severity definitions with examples.
Readiness test: ask three people on the team what "critical" means. If you get three different answers, write one definition before you let a model apply it.
This example often pays off fastest for product teams. The work is high volume and tedious, and a wrong grouping costs a few minutes of a product manager's time, not a customer.
When plain automation beats AI
Many processes pitched as AI projects don't need AI. A rule-based workflow costs less to run, is faster, gives the same output every time and is easy to audit. Use plain automation when:
• The input is already structured: form fields, database rows or API events.
• The logic fits in if-then rules you could write on a whiteboard.
• The output must be exactly right every time, as with tax calculations, payroll or stock counts.
• Documents come from a few fixed templates that rarely change.
• The volume is so low that a person handles it in a few minutes a week.
Typical examples: syncing paid orders into your accounting tool, sending reminders when a task is overdue, and creating accounts when HR marks a new hire as started. Workflow tools handle these well. Our n8n automation guide shows where that kind of tool fits.
The best setups combine both. A rule-based workflow moves the data and enforces the rules, and AI is used only at the step where unstructured text or judgment appears. Before anything is written to a system of record, the workflow checks the AI output: required fields are present, values are within allowed ranges and IDs exist.
Decision rule: if you can write the rules, write the rules. Add AI only where a person currently has to read something before deciding.
How to go from first pilot to a working system
Follow these steps in order, and don't skip the middle ones:
1. Pick one process from the list above that passed its readiness test.
2. Build a test set of 50 to 100 real past cases with known correct answers.
3. Run the AI in shadow mode: it produces outputs while people keep working as usual, and you compare the two.
4. Switch to human review: people approve or edit each AI draft, and you log every edit.
5. Allow automatic actions only for categories where approval rates stay high and mistakes are cheap.
6. Keep monitoring accuracy, cost per task, response time and the share of cases sent to people.
Several factors drive build effort and running cost: how many systems you integrate, how varied your documents are, how much accuracy you need, how many model calls each task makes, whether data must stay on your own servers, and your audit logging requirements. Most projects stall between steps 3 and 5. Our guide on moving AI pilots to production covers the usual causes.
If you want help choosing and building the first one, Geminate Solutions' AI automation service starts by reviewing your process, not by picking a model. We have shipped 50+ products. You own 100% of the code and IP, and we sign an NDA before we talk.
