Ask a business what it wants to automate and you usually get the answer that is most annoying, not the answer that is most expensive. Those are rarely the same process. The annoying one is visible; the expensive one is quietly spread across four people and never measured.
So this guide starts with the audit, not the technology.
What business process automation actually means
Automation is software doing repeatable work a person currently does by hand. Concretely, it looks like:
- An order arriving by e-mail becoming a record in your system, with no typing.
- A supplier invoice being read, matched to a purchase order and queued for approval.
- Quotes generated from a price matrix instead of a copied Word file.
- Follow-ups going out on day 2, 7 and 21 without anyone remembering to send them.
- The Monday report assembling itself before Monday.
What it is not: a strategy, a platform purchase, or a project that needs to touch everything at once. Automation is a series of small, unglamorous wins that compound.
How to find what to automate first
Spend two hours on this and you will save far more than two hours of engineering. List every recurring task your team does and score each one 1–5 on four axes:
| Axis | Ask | Score 5 when… |
|---|---|---|
| Volume | How often does it happen? | Many times a day |
| Repetitiveness | Is it the same steps every time? | Identical steps, no interpretation |
| Rule clarity | Could you write the rules on one page? | The rules already exist in writing |
| Error cost | What does a mistake cost? | Mistakes cost real money or a customer |
Add the scores. Anything at 16 or above is a strong candidate. Anything under 10 should probably stay manual. The automation will cost more to maintain than the work it removes.
Then apply a sanity filter: how many hours a week does this consume, times the loaded hourly cost of the person doing it? That is your payback numerator. An automation that costs $2,000 and saves six hours a week pays for itself in roughly two months.
The four types of automation
1. Connective: moving data
The plumbing between systems: website form to CRM, CRM to accounting, marketplace orders to inventory, payment gateway to ledger. Cheapest to build, fastest to pay back, and the type most businesses need the most of. If a person is currently the integration between two tools, this is your answer.
2. Document: reading data
Invoices, purchase orders, delivery notes, bank statements, ID documents, contracts. Modern extraction handles messy real-world scans far better than the template-based OCR of a few years ago, but it needs a validation step: extract, check against known values, flag anything uncertain for a human. Anyone promising 100% hands-off document processing is selling you a future clean-up job.
3. Conversational: talking to people
Chat and voice agents that answer questions, qualify enquiries, book appointments and create tickets. This is where AI has changed most in two years; an AI calling agent can now hold a useful phone conversation and write the outcome to your CRM. Best deployed on the front of a process, where volume is high and each decision is small.
4. Decision support: ranking and recommending
Reorder suggestions from sales velocity, lead prioritisation, credit-risk flags, anomaly alerts on spend. Note the word support: the software proposes, a human approves. Fully automated decisions belong only where the rules are unambiguous and the downside is small.
Automations that consistently pay off
| Automation | Typical saving | Type |
|---|---|---|
| Enquiry capture from every channel into one CRM with an owner assigned | Recovers the enquiries that used to be lost entirely | Connective |
| Quote generation from a price matrix, with approval rules | Hours per week, plus fewer pricing errors | Connective |
| Supplier invoice extraction and PO matching | Days a month in accounts | Document |
| Automatic follow-up sequences on quotes that go quiet | Measurable conversion lift, no extra headcount | Connective |
| Out-of-hours call answering and booking | Every missed call becomes a captured one | Conversational |
| Daily operations report assembled and sent automatically | An hour a day of someone senior | Connective |
| Reorder alerts from live stock and sales velocity | Fewer stockouts and less dead capital | Decision support |
| Payroll inputs from attendance data | A day or two a month, and fewer disputes | Connective |
Not sure which of these fits you?
Describe the task your team complains about most and roughly how often it happens. We will tell you whether it is worth automating, and what it would cost, within 24 hours.
What not to automate
- A process nobody agrees on. If two managers describe the approval chain differently, you have a management problem, not an automation project.
- Anything that happens twice a month. Low volume, high maintenance. Leave it.
- Judgement work. Hiring decisions, pricing exceptions, complaint resolution. Automate the paperwork around them instead.
- Relationships that are the product. If clients pay for access to your people, do not put a bot in front of them.
- A process you are about to change anyway. Wait for the new process, then automate that.
What it costs
| Automation | Typical build | Ongoing |
|---|---|---|
| Single connective automation (two systems, clear rules) | $800–$3,000 | Minimal; hosting only |
| Multi-step workflow with approvals and notifications | $3,000–$8,000 | Hosting plus occasional rule changes |
| Document extraction pipeline with validation | $3,000–$10,000 | Per-page or per-document processing cost |
| Conversational agent (chat or voice) | $3,000–$12,000 | Usage-based per conversation or minute |
The pattern that works: start with one automation under $3,000, measure the hours it actually returns, then reinvest that proven saving into the next one. Businesses that begin with a platform-wide automation programme usually spend more and adopt less.
A two-week rollout that works
- Days 1–2: pick the single highest-scoring process. Write the current steps down, including every exception the team mentions. The exceptions are the actual specification.
- Days 3–4: agree the rules and the failure behaviour. What happens when data is missing, when a value looks wrong, when an API is down?
- Days 5–8: build it, with logging from the first commit.
- Days 9–10: run it in shadow mode. The automation does the work, a human checks the output, nothing is trusted yet. This is the step people skip and regret.
- Days 11–14: switch on, keep the manual fallback documented, and set an alert to a named owner for every failure.
Then measure the same number you measured before you started: hours spent, errors, turnaround time. If it did not move, say so out loud and fix it before automating anything else.
Five mistakes that make automation backfire
- Automating a broken process. You get faster mistakes and less visibility. Fix the rules first.
- No owner. Every automation needs a name attached. Unowned automations rot until someone notices the numbers are wrong.
- Silent failure. If it can fail without telling anyone, it will, usually during your busiest week.
- No manual fallback. APIs go down. There must be a documented way to do it by hand for a day.
- Optimising the wrong end. Automating reporting while order entry is still manual is polishing the output of a bottleneck.
Where to start this week
Pick the one task your team complains about most, then check it against the scoring table. If it scores 16 or higher, it is almost certainly worth a small build. If it does not, find the quieter task that does, usually something in accounts or order entry that nobody has thought to complain about because it has always been done by hand.
Automation is the cheapest of the five systems most businesses need, and it is the one that makes the others worth more. We build these as part of the AI & automation practice. Tell us the task and you will get an honest read on whether it is worth automating at all.