BUSINESS OPERATIONS
Which business process is actually worth automating with AI?
Examine the work before choosing the tool, then separate process repair, ordinary automation, AI assistance and decisions that should remain human.
The best first process for AI automation is usually not the most impressive process in the business. It is a repeated, measurable piece of work with reliable information, a clear owner and mistakes that can be detected before they cause serious harm.
That may be preparing a routine customer reply, checking an invoice against a purchase order, finding deliveries that have not been invoiced or combining the same weekly reports.
It is less likely to be “run customer service,” “manage procurement” or “advise the chief executive.” Those are collections of decisions, relationships and exceptions—not single processes ready to hand to a tool.
Before choosing software, map the work. You may discover that the right answer is AI assistance, ordinary automation, a simpler process or no automation at all.
The executive answer
A process is a promising candidate when:
- it happens often enough for the improvement to matter;
- employees spend substantial time finding, copying, checking or reformatting information;
- the beginning, outcome and responsible owner are clear;
- the required information exists and can be accessed appropriately;
- normal cases can be separated from exceptions;
- a person can review consequential decisions;
- errors can be detected and corrected; and
- success can be measured against the present process.
Pause when the work changes every time, the rules are disputed, the source records are unreliable, the decision has a major effect on a person or nobody owns the result.
The OECD's 2025 review of AI adoption by small and medium-sized enterprises identifies connectivity, data, skills and finance as important enablers. Technology availability alone is not enough. Read the OECD discussion paper.
Start with the work people complain about
Ask department leaders and employees three questions:
- What do we repeatedly wait for?
- What information do we copy, search for or check more than once?
- Which avoidable mistake or omission keeps returning?
You are looking for a concrete description such as:
- “Sales waits for stock confirmation before replying to large enquiries.”
- “Finance compares delivery records with invoices every Friday.”
- “Customer service searches three folders before answering warranty questions.”
- “Branch managers copy the same numbers into a Monday spreadsheet.”
Avoid starting with “We need an AI agent.” That is a proposed solution, not a description of the work.
Map one process from trigger to result
A useful process map does not need specialist notation. Put seven items on one page:
- Trigger: What starts the work?
- Inputs: Which messages, documents or records are needed?
- Steps: What do people and systems do now?
- Decisions: Where is judgement or approval required?
- Exceptions: What causes delay, rework or escalation?
- Output: What useful result should exist at the end?
- Owner: Who remains responsible for that result?
Follow a real case rather than asking only how the process is supposed to work. The policy may say orders enter through the sales system. The Tuesday afternoon reality may include WhatsApp screenshots, voice notes, a salesperson's notebook and an urgent call from the managing director.
Both matter. Automation built around the official process will fail if the real inputs continue outside it.
Automate the preparation while keeping the decision visible.
Australia's National AI Centre recommends mapping processes before deciding where AI may help, including the people responsible, the pain points and whether the work is sufficiently information-rich. Use the National AI Centre's process-mapping guidance.
Separate four different types of improvement
Not every slow process needs AI.
| Improvement | Best fit | Example |
|---|---|---|
| Remove or simplify the step | Work exists because of duplication, unclear ownership or an outdated rule | Stop asking two departments to enter the same customer address |
| Ordinary automation | Inputs and decisions follow stable, explicit rules | Send a reminder three days before a known due date |
| AI-assisted work | The step involves interpreting language, documents, images or varied requests | Classify an incoming enquiry and prepare a draft reply |
| Human decision | The work requires accountability, negotiation, empathy or a high-consequence judgement | Approve a refund exception or decide employee discipline |
Most useful processes combine all four. Removing duplication may create more value than automating it. Ordinary rules may handle predictable routing. AI may prepare information from untidy inputs. A person may decide the exception.
The question is not “Can AI do the whole process?” It is “What is the most sensible division of work?”
Score the candidate process
Use a score from 1 to 5 for each factor. The number is not a scientific truth; it makes assumptions visible so managers can compare candidates consistently.
1. Business value
What would improve if this process worked better?
Look for customer waiting, lost revenue, staff capacity, avoidable cost, errors, risk or management delay. “Employees dislike it” is useful evidence, but estimate how often it happens and what the frustration causes.
High score: The process affects an important outcome, occurs often and creates visible cost or delay.
Low score: It is annoying but rare, or the improvement would not change a meaningful outcome.
2. Repetition and volume
Automation becomes more valuable when similar work recurs. Count cases per day, week or month. Also count the time between arrival and completion.
High volume is not essential if each case consumes substantial time, but a five-minute task performed twice a year is unlikely to justify a custom system.
3. Process stability
Can employees describe the normal path? Do different teams agree on the rules? Does the same input normally lead to the same next step?
Do not automate disagreement. If sales, operations and finance use different definitions of a completed order, connecting their systems will expose the conflict faster; it will not resolve it.
4. Information readiness
List every source the work requires. For each one, ask:
- Is it available in a consistent place?
- Who owns it?
- How current is it?
- Which source wins when records conflict?
- May the proposed system access it?
- Can a result link back to the source?
AI can help interpret messy information. It cannot make an outdated price list current or decide which of two conflicting policies management intended.
5. Detectability of error
How would someone know the result is wrong?
A draft email can be checked before sending. An extracted invoice total can be compared with the source. An invisible classification error that quietly denies service to a customer is much harder to detect.
Favour early candidates where a reviewer has evidence, errors are visible and the action can be reversed.
6. Consequence and risk
Consider money, customer promises, confidential information, employment, safety, legal obligations and reputation.
High risk does not mean AI can never help. It usually means the system should prepare or recommend rather than decide or act, and the controls must be stronger.
7. Ownership and adoption
Who owns the process today? Who will review exceptions? Who keeps the information current? Who will answer employees when something fails?
A process with no owner is not ready for automation. The project may create a new queue that everyone assumes somebody else is watching.
8. Measurability
Can you record the present time, waiting, corrections, cost or missed outcomes? Can the same measures be collected during a pilot?
If success is defined only as “people used the tool,” management will struggle to decide whether to keep paying for it.
A practical prioritisation table
Score value, repetition, stability, information, error detection, ownership and measurability positively. Treat consequence as a reason to narrow authority or postpone the candidate.
| Candidate | Value | Repetition | Stable process | Reliable information | Errors easy to detect | Clear owner | Overall reading |
|---|---|---|---|---|---|---|---|
| Prepare routine order-status replies | 4 | 5 | 4 | 4 | 5 | 5 | Strong first pilot; keep sending under human review |
| Recommend employee dismissal decisions | 5 | 2 | 2 | 2 | 1 | 4 | Poor automation candidate; high consequence and difficult error detection |
| Combine weekly branch reports | 4 | 5 | 3 | 3 | 4 | 4 | Promising after agreeing definitions and sources |
| Categorise an annual folder of old documents | 2 | 1 | 4 | 3 | 4 | 3 | Technically possible, but limited continuing value |
Do not simply add the numbers. Discuss why each score was given and what would change it. A low information score may become acceptable after one source is cleaned and assigned an owner.
Worked example: a WhatsApp order enters four systems
Consider an illustrative distributor. Customers send orders through WhatsApp. A salesperson confirms the product and quantity, checks stock, prepares a quotation, waits for customer confirmation, records the order in the ERP and tells dispatch.
The managing director initially asks for “an AI WhatsApp agent that handles orders.” Mapping the process reveals several different jobs:
| Step | Current problem | Sensible first treatment |
|---|---|---|
| Read the customer's message | Product names and quantities are expressed inconsistently | AI may prepare a structured order request |
| Match the product | Customers use informal names | Use AI to suggest likely products; salesperson confirms ambiguous matches |
| Check stock and price | Information lives in the ERP | Read the current ERP record; do not let AI invent either value |
| Prepare the quotation | Repeated formatting and copying | Ordinary template plus approved fields |
| Approve a special discount | Requires commercial judgement and authority | Named manager decides |
| Confirm the order | Customer must agree to product, quantity, price and delivery | Explicit customer confirmation |
| Create the ERP order | Repeated entry | Prepare the record first; allow write-back only after controls and testing |
The resulting pilot is much narrower than “automate WhatsApp orders”:
For enquiries about 20 stable products, prepare a structured order request, retrieve current stock and approved price, and draft a quotation for the salesperson to check. Do not promise delivery, apply a discount or create the ERP order without the required confirmations.
That scope is easier to test, measure and explain to employees and customers.
Choose a first pilot with a two-way door
Prefer a process where the business can reverse course without disrupting customers or losing essential records.
Useful first-pilot characteristics include:
- read-only access to source systems;
- drafts rather than automatic external messages;
- a limited group of trained users;
- a defined set of suitable cases;
- representative test examples;
- a visible exception queue; and
- a simple way to return to the present process.
The goal is not to make the pilot meaningless. It is to learn while the cost of being wrong remains contained.
Define a pilot sentence before discussing vendors
Use this format:
When [trigger] happens, the system may use [approved information] to prepare [specific output]. [role] remains responsible for [decision or approval]. We will compare [business measure] and [quality measure] across [representative sample and period].
For example:
When an existing customer asks where an order is, the assistant may read the order and delivery records and prepare a source-linked reply. A customer-service employee checks and sends it. We will compare preparation time and factual corrections across 100 suitable enquiries over four weeks.
This sentence gives technology, operations and management a shared definition of the job.
Warning signs that the process is not ready
Pause when:
- nobody can agree where the process starts or ends;
- the proposed benefit is based only on a vendor estimate;
- source records are routinely bypassed;
- employees maintain essential facts in private messages or notebooks;
- every case is described as an exception;
- the project needs broad access before a narrow test can begin;
- the output creates a commitment nobody has authority to approve;
- errors would be difficult to notice or reverse; or
- management cannot name the person who will own live performance.
Some of these problems can be fixed. Fixing them may be the most valuable part of the project.
Improve the process before increasing the autonomy
AI is useful when it can interpret information, find relevant context or prepare a next step that people currently assemble by hand. It is not a substitute for a clear source, an accountable owner or an agreed decision.
Start with work your team understands. Map a real case. Remove the unnecessary steps. Automate the predictable ones. Use AI where interpretation genuinely helps. Keep human responsibility where the consequence demands it.
Once you have a strong candidate, continue with which AI agent your business should start with and how to measure the pilot's ROI.
Research and helpful links
- OECD: AI adoption by small and medium-sized enterprises
- Australia's National AI Centre: map your processes
- UK Government: assessing whether AI is the right solution
- NIST AI Risk Management Framework Core
Research checked on 20 September 2026. Apply the operational, legal, professional and data-protection requirements relevant to your organisation and market.