What should businesses actually automate with AI in 2026?

AI can now read, write, research, interpret documents, use software and carry out increasingly long pieces of work. That doesn't mean your business should automate everything it can. Here's where I'd actually start.

If you run a business in 2026, you've probably heard some version of this: You should be automating more with AI. Fair enough.

But what? That's the useful question. Because almost every part of a business now contains something AI could theoretically help with.

You could automate emails. Sales. Support.

Research. Documents. Reports.

Marketing. Administration. Software.

Scheduling. Internal knowledge. But the fact that something can be automated doesn't mean it should be.

So rather than giving you another enormous list of AI use cases, I'd start somewhere more practical. Look for work that is: repeated,

time-consuming, information-heavy, reasonably predictable,

easy to check, and where mistakes don't immediately create serious consequences. That's usually where the good opportunities are hiding.

First: don't automate the job

Automate the work inside it. This distinction matters. Take somebody working in sales.

Their job isn't one thing. They might: research prospects,

read enquiries, qualify leads, update the CRM,

write follow-ups, prepare proposals, answer questions,

schedule calls, negotiate, build relationships,

and decide which opportunities deserve attention. AI may be useful across several of those tasks. That doesn't mean you should hand "sales" to an AI agent.

Break the work apart. Then decide what belongs where.

A useful rule for 2026

Before we go through the business, here's the principle I'd use: Automate the repetitive. Use AI for the interpretive.

Keep people around the consequential. It's deliberately not an absolute rule. But it's a very useful starting point.

If a task simply follows a rule, normal automation may be enough. If something needs reading, understanding, comparing, summarising or drafting, AI may add something. If a decision has serious financial, legal, safety, reputational or human consequences, think carefully about where a person should remain involved.

Now let's walk through the business.

1. Enquiries and lead handling

This is one of the first places I'd investigate. A new enquiry arrives. What happens?

Someone reads it. Works out what it's about. Checks whether it's relevant.

Perhaps looks the person up. Creates or updates a CRM record. Routes it.

Writes a response. Creates a follow-up. None of those steps is particularly enormous.

Together they create a lot of administration.

What AI can help with

AI can potentially: understand what an enquiry is about, categorise it,

identify urgency, extract useful information, summarise what the person wants,

research relevant context, prepare a response, and suggest what should happen next.

What ordinary automation can handle

Once the interpretation has happened, normal automation may: create the CRM record, assign the lead,

send notifications, create tasks, set follow-up dates,

and move the opportunity through the appropriate workflow.

Where I'd keep a person

High-value opportunities. Unusual requests. Negotiation.

Sensitive conversations. Anything where understanding the relationship matters more than processing the enquiry.

Don't automate blindly

Don't let AI send every sales response without first understanding the kinds of enquiries you receive and the consequences of getting one wrong. GOOD FIRST PROJECT: Have AI classify and prepare incoming enquiries while a person approves the response.

2. Sales follow-up

A surprising amount of sales is remembering. Who needs following up? What did they ask?

What was promised? When should we contact them? What has changed since the last conversation?

AI can make this much easier.

What AI can help with

Summarise previous conversations. Identify the important context. Prepare personalised follow-ups.

Suggest the next useful action. Highlight opportunities that appear to have gone quiet.

What automation can handle

Wait for a specified period. Check whether someone replied. Create the task.

Update the stage. Schedule the next reminder.

Where I'd keep a person

Anything involving: pricing discretion, negotiation,

important relationships, commitments, or judgement about whether pursuing an opportunity is appropriate.

GOOD FIRST PROJECT: Prepare follow-ups automatically, but let the salesperson review before sending.

3. Customer service triage

Customer service contains huge amounts of repeated interpretation. What does this person need? Is it urgent?

Which team should handle it? Can we answer from existing information? Does this need escalation?

AI can be extremely useful here.

What AI can help with

Read incoming messages. Understand intent. Identify sentiment or urgency signals.

Find relevant information. Summarise previous interactions. Prepare suggested responses.

Route unusual cases.

What automation can handle

Create tickets. Assign queues. Update statuses.

Send acknowledgements. Schedule follow-ups. Close completed workflows.

Where I'd keep a person

Complaints. Vulnerable customers. Unusual circumstances.

High-value accounts. Situations where the business may need to make an exception.

Don't automate blindly

A fast wrong answer isn't better customer service. Measure resolution and accuracy, not simply response speed. GOOD FIRST PROJECT: AI-assisted triage before autonomous customer communication.

4. Repetitive customer questions

This is slightly different from general customer service. Many businesses receive the same questions constantly. Where is my order?

What documents do I need? How does this work? Do you provide X?

What's included? Can I change Y? If the answers already exist somewhere reliable, AI can make them easier to access.

What AI can help with

Understand the customer's wording. Find the relevant information. Explain it appropriately.

Ask for missing information.

What ordinary software can handle

Account information. Order status. Known dates.

Tracking. Prices. Structured account data.

Where I'd be careful

Don't let the AI invent policies, pricing, contractual terms or promises because it couldn't find the answer. A useful system needs to know when to say: I need somebody to check this.

5. Documents

This is one of the areas I'd look at very seriously in 2026. Businesses are full of PDFs, forms, contracts, applications, reports, invoices and other documents. People spend enormous amounts of time opening them and finding the bits that matter.

What AI can help with

Read documents. Extract information. Summarise them.

Compare versions. Identify missing information. Classify documents.

Answer questions about their contents. Highlight unusual clauses or differences for review.

What automation can handle

Rename files. Move them. Create records.

Trigger workflows. Request missing documents. Notify people.

Archive completed cases.

Where I'd keep a person

Legal interpretation. High-consequence decisions. Ambiguous information.

Anything where an incorrect extraction could materially affect somebody. GOOD FIRST PROJECT: Extract and organise information for a person rather than immediately allowing AI to make the final decision from it.

6. Internal knowledge

Ask employees how much time they spend finding things. The latest price list. The correct policy.

A process. A previous proposal. A technical answer.

A client document. Something someone remembers seeing six months ago. This is an enormous hidden cost.

What AI can help with

Search across business information using normal questions. Find relevant documents. Summarise them.

Compare information. Answer questions using approved sources.

The difficult bit

The AI isn't necessarily the problem. Your information might be. If there are four versions of the same policy and nobody knows which is current, AI doesn't magically solve that.

It may simply find the wrong one faster. GOOD FIRST PROJECT: Choose one well-maintained body of internal knowledge rather than connecting AI to every file the company has ever created.

7. Reporting

Reporting is another excellent place to look. Especially if somebody currently spends hours assembling information that other people consume in minutes.

What automation can handle

Collect data. Run queries. Calculate metrics.

Update dashboards. Schedule reports.

What AI can help with

Explain what changed. Summarise patterns. Compare periods.

Identify anomalies worth investigating. Turn data into a readable briefing. Suggest questions somebody may want to explore.

Where I'd keep a person

Interpreting why something happened when the evidence isn't clear. Making major decisions from the analysis. Approving external or board-level conclusions where accuracy matters.

GOOD FIRST PROJECT: Automate report preparation and use AI to draft the narrative, with a person checking the interpretation.

8. Research

AI can dramatically reduce the effort involved in the first stages of research. Competitors. Markets.

Products. Suppliers. Regulations.

Technology. Potential customers. Industry developments.

What AI can help with

Find information. Read multiple sources. Summarise.

Compare. Extract relevant facts. Organise findings.

Identify areas requiring deeper investigation.

Where I'd be careful

AI-generated research can sound extremely confident while being: out of date, poorly sourced,

missing context, or simply wrong. For research that matters, you need sources.

And somebody needs to distinguish between what the evidence actually says and what the AI inferred from it. GOOD FIRST PROJECT: Use AI to accelerate research gathering and synthesis, not to remove verification.

9. Meeting administration

Meetings generate a ridiculous amount of secondary work. Notes. Actions.

Follow-ups. CRM updates. Documents.

Emails. Tasks. This is an obvious area for improvement.

What AI can help with

Summarise the discussion. Identify decisions. Extract actions.

Draft follow-up emails. Prepare information for the CRM.

What automation can handle

Create tasks. Set deadlines. Update records.

Send reminders. Schedule follow-ups.

Where I'd be careful

An AI summary is not necessarily an authoritative record. Important commitments, decisions or contractual points should be checked. GOOD FIRST PROJECT: Turn meeting notes into proposed actions and updates for quick human approval.

10. Inbox and email administration

I wouldn't hand somebody's entire inbox to an autonomous AI on day one. But there's a lot between "AI writes an email for me" and "AI controls my email".

AI can help with

Classifying messages. Summarising threads. Identifying actions.

Finding messages needing attention. Preparing replies. Extracting dates and commitments.

Automation can handle

Routing. Labels. Tasks.

Reminders. Follow-up dates. Moving information into other systems.

Where I'd keep a person

Sensitive communication. Complaints. Negotiations.

Commitments. Anything where tone or context matters substantially. GOOD FIRST PROJECT: AI prepares the inbox rather than owning it.

11. Finance administration

Finance contains plenty of repetitive work, but it also contains obvious consequences. That makes it a good example of where assistance and automation need boundaries.

AI can potentially help with

Reading invoices. Extracting fields. Categorising incoming finance documents.

Explaining variances. Preparing summaries. Matching supporting information.

Automation can handle

Known calculations. Moving records. Status changes.

Reminders. Reconciliation rules. Approval routing.

Where I'd keep a person

Payments. Changing bank details. Financial commitments.

Unusual transactions. Anything with significant fraud risk. GOOD FIRST PROJECT: Use AI to prepare and organise finance work, not give it unrestricted authority to move money.

12. Operations

Operations is where some of the least glamorous and most valuable opportunities live. Every company has repeated coordination. Orders.

Suppliers. Schedules. Stock.

Jobs. Approvals. Deliveries.

Exceptions. Status updates.

Look for

People repeatedly checking systems. Information being copied. Someone chasing another person.

Work waiting for approval. Status updates being manually communicated. The same exception being handled repeatedly.

AI may help with

Understanding exceptions. Reading supplier communications. Summarising operational information.

Prioritising issues. Preparing responses.

Automation may handle

Status movement. Notifications. Scheduling.

System updates. Escalation timers. Routine checks.

GOOD FIRST PROJECT: Find one operational process that crosses multiple systems and map where the human is currently acting as the connector.

13. Marketing production

This is probably one of the most obvious AI uses. And therefore one of the easiest to use badly. Yes, AI can produce:

articles, social posts, emails,

ad copy, product descriptions, ideas,

images, and campaign variations. The question isn't whether it can.

The question is whether producing more content is actually the business problem.

Better uses may include

Researching customer questions. Repurposing strong original material. Analysing existing content.

Preparing variants. Extracting insights from interviews. Creating first drafts.

Finding gaps in a content library.

Where I'd keep a person

Positioning. Original point of view. Claims.

Brand judgement. Anything published in the company's name that actually matters. GOOD FIRST PROJECT: Use AI to multiply strong thinking, not manufacture a mountain of generic content.

14. Sales proposals and documents

If your business repeatedly creates proposals from similar building blocks, there may be a substantial opportunity here.

AI can help with

Understanding the enquiry. Selecting relevant information. Creating the first draft.

Tailoring explanations. Summarising scope. Turning notes into structured content.

Automation can handle

Templates. Pricing calculations. File creation.

CRM updates. Approval routing. Sending after approval.

Keep people around

Pricing exceptions. Contractual commitments. Scope.

Promises about delivery. Anything bespoke enough that misunderstanding it creates work later. GOOD FIRST PROJECT: Generate a first proposal from approved information and structured sales notes, then review it.

15. Recruitment administration

There is useful administrative work AI can assist with here, but hiring decisions require considerably more care.

AI can help with

Preparing job descriptions. Organising interview notes. Scheduling communications.

Summarising information supplied by candidates. Preparing consistent interview questions.

Where I'd be very cautious

Automatically deciding who deserves an interview. Ranking candidates based on opaque AI judgement. Inferring personality or suitability from irrelevant characteristics.

Making consequential employment decisions without meaningful human oversight. This is a good example of an area where: AI can reduce administration without needing to make the decision.

16. Software development

AI-assisted development has moved quickly. It can now help: write code,

understand codebases, debug, create tests,

refactor, write documentation, build interfaces,

connect APIs, and work through larger technical tasks. For businesses, the important consequence isn't simply that developers type faster.

It's that more software becomes economically viable to build. Small internal tools. Integrations.

Customer portals. Process-specific applications. Things that previously weren't valuable enough to justify weeks of development may now be worth testing.

Where I'd keep strong human oversight

Architecture. Security. Authentication.

Permissions. Data. Production deployment.

Anything the business will depend on. GOOD FIRST PROJECT: Build a small internal tool around a known process rather than beginning with a huge replacement system.

17. Quality checking

AI can also sit on the other side of work. Instead of creating something, it can check it. Compare a document against requirements.

Look for missing fields. Check whether something follows a format. Review content against a set of criteria.

Identify inconsistencies. Flag unusual cases. This is particularly interesting because AI can sometimes become a second pair of eyes rather than the person doing the work.

Be careful

A check is only useful if the checker is reliable enough for the consequence involved. For important work, AI can flag issues without being the final authority.

18. Repetitive web and software tasks

This category is becoming much more interesting. Many business processes still require someone to: open a website,

log in, find something, copy information,

enter information elsewhere, download a document, upload it,

and repeat. AI systems are increasingly capable of operating software through the interface. That potentially opens up processes that were awkward to automate because the systems didn't have suitable integrations.

Good candidates

Repeated browser-based work. Known systems. Clear outcomes.

Easy-to-check actions. Low-consequence changes.

Bad candidates

Giving an AI broad access to important systems and telling it: Sort this out. Start narrow.

So where would I start?

Not with all 18. Please don't come away from this article with an enormous AI transformation programme. Pick one.

I'd look for something with these characteristics: It happens frequently. A five-minute task performed 200 times matters more than a two-hour task performed once a year.

The current process is understood. If nobody agrees how it works, fix that first. The inputs are available.

The system can actually get the information it needs. The output is easy to check. You can tell whether it worked.

Mistakes are recoverable. A failure doesn't immediately create a disaster. There is measurable value.

Time, capacity, speed, quality, consistency, revenue or customer experience can improve. That's a very good first AI automation project.

I'd avoid starting with the highest-risk process

The most valuable process in the company isn't necessarily the best first project. If it involves: large financial transactions,

legal commitments, sensitive personal data, safety-critical decisions,

irreversible actions, or major reputational consequences, I'd probably learn somewhere else first.

Build confidence on work where the consequences are manageable. Then increase autonomy as the evidence justifies it.

There's a useful progression

You don't have to go from: person does everything straight to:

AI does everything. There are several stages in between. AI observes

It reads or analyses but doesn't change anything. AI recommends It suggests what a person could do.

AI prepares It drafts the response or action. AI acts with approval

A person checks before execution. AI acts within limits Routine cases proceed automatically. Exceptions escalate.

AI runs the workflow The system handles most of the process and involves a person where required. That's a much safer and more useful way to think about adoption than simply asking whether a task is "automated".

Don't forget the cost of checking

This one catches businesses out. Suppose AI completes a task in 30 seconds that used to take somebody 15 minutes. Fantastic.

Except the employee now spends ten minutes checking whether the AI got it right. You haven't saved 14 minutes and 30 seconds. You've saved four and a half minutes.

Perhaps that's still worthwhile. Perhaps it isn't. Measure the whole process.

And don't measure success by how much AI you're using

A successful AI project isn't: 85% of our employees now use AI. It's something more like:

Enquiries are answered four hours faster. Preparing this report takes 20 minutes instead of three hours. Our team can process twice as many applications.

Fewer requests are dropped between systems. We can now offer a service that wasn't economically practical before. That's business value.

The best opportunities are often surprisingly boring

I keep coming back to this. The AI demonstrations that attract attention tend to be dramatic. The business implementations that create value often aren't.

An invoice gets processed without someone retyping it. A lead doesn't sit unnoticed for two days. A report appears on Monday morning without someone spending Friday afternoon building it.

An employee finds the answer without asking three colleagues. A customer gets routed to the right person first time. A piece of information moves between two awkward systems.

Nobody makes a viral video about it. But the business gets better.

What I'd do this week

Don't buy anything yet. Speak to the people actually doing the work. Ask:

What do you do repeatedly that feels like a waste of your time? Write down the answers. Then look for the patterns:

COPYING CHECKING SEARCHING

CHASING WAITING REWRITING

RE-ENTERING SORTING SUMMARISING

MOVING INFORMATION Those are your starting points. Then map the process.

Work out which steps can disappear. Which can use normal automation. Which genuinely benefit from AI.

And where a person still adds value. That's a much better AI strategy than starting with a list of tools.

What should your business automate with AI in 2026?

The work where AI creates a meaningful improvement. Not the work that makes the best demo. Not the work everybody on LinkedIn is talking about.

Not everything technically possible. Start with repeated work. Find the friction.

Measure it. Remove what shouldn't exist. Automate what is predictable.

Use AI where interpretation adds value. Keep people around the decisions where people matter. Then measure what changed.

The best AI automation isn't the one with the most AI in it. It's the one where, afterwards, the business simply works better.

Where to go next

• Don't start with AI. Start with the work. The Creative Sauce AI approach to finding the right problems before choosing the technology.
• Before you automate a process with AI, draw the process Map Trigger, Information, Decision, Action, Check and Next Action.
• AI vs automation: what's the difference, and when do you need each? Work out which parts need fixed rules, which need interpretation and where a person belongs.
• The important question isn't what AI can do. It's what you should let it do. Once you've found something worth automating, decide how much authority the AI should have.

Related: Don't start with AI. Start with the work..