Good AI business process automation removes handling work without removing ownership. AI agents may read, classify, draft or route a task. A person still controls the rules, exceptions and final actions. That operating design matters more than the platform name.
The four tests
A process needs to pass all four. Missing one does not always kill the idea, but it raises the cost and risk.
Repeated
The task happens often enough for saved handling time to repay the build.
Readable
The system can access the inputs as text, documents, records, images or another usable format.
Rule-led
A capable person can explain what a good result looks like and where judgement starts.
Checkable
A person or deterministic rule can catch a bad output before it causes harm.
A weekly task with one owner, a known source and a draft output. It is frequent enough to matter and safe enough to test without disrupting the business.
Processes that usually fit
| Process | What AI handles | What a person keeps |
|---|---|---|
| Document triage | Reads, classifies, extracts fields and routes the file. | Checks uncertain cases and approves sensitive decisions. |
| Lead qualification | Enriches an enquiry, checks fit rules and drafts a next step. | Owns the relationship and decides priority. |
| Recurring reports | Combines data, flags changes and drafts commentary. | Explains causes, makes recommendations and signs off. |
| Research briefs | Collects sources, extracts facts and organises a first brief. | Checks sources, resolves conflicts and decides the argument. |
| Internal knowledge | Finds approved material and answers staff questions with citations. | Maintains the source set and handles policy exceptions. |
| Content operations | Moves approved briefs through research, drafting and checks. | Owns facts, tone, judgement and final publication. |
Processes that need extra care
A useful workflow may still carry personal, confidential or commercially sensitive data. That changes the design.
- Customer support. Let AI classify and draft. Keep refunds, complaints and account changes behind approval.
- Invoices and finance. Extract and reconcile data. Keep payments and ledger changes behind deterministic checks.
- Recruitment. Organise applications and interview notes. Do not let a model make an unchecked employment decision.
- Legal documents. Summarise and compare clauses. Keep legal advice and acceptance with a qualified person.
- Health or safety work. Use AI for clerical support only unless the system has appropriate clinical or safety governance.
The ICO AI and data protection risk toolkit covers risks to people’s rights. The NCSC lifecycle security guidance covers security throughout the system lifecycle.
Processes I would not automate first
Rare work, political decisions, disputed facts and irreversible actions make poor first projects. So does a process nobody can explain or own.
- A task that happens once a quarter.
- A decision with no written standard.
- A process built on missing or disputed data.
- An action that sends money, deletes data or commits the business without review.
- A workflow where every case is genuinely different.
How to choose between two candidates
Score each process from one to five on frequency, clarity, data access, reversibility and owner commitment. Pick the highest total with the lowest consequence of failure.
Do not start with the task that sounds most impressive. Start with the one your team already understands. That makes testing faster and arguments about success much harder.
What to write down before speaking to a builder
- What starts the process.
- Which systems contain the source material.
- What decisions happen in the middle.
- What the finished output looks like.
- Who checks it.
- What happens when the system is unsure.
Those six answers are enough for an AI automation consultant to challenge the idea and scope the next step. The AI automation cost guide then shows how the number is built.
