Back Office Automation: How to Decide What to Automate First

Your finance team closes the month three days late because two people spend those days re-keying figures from one system into a spreadsheet and then into another system. Your operations coordinator forwards the same approval request twice a week because nobody knows whose desk it is sitting on. Nobody in the company describes this as a technology problem — it is simply how the work has always been done. It is also the cheapest thing in the business to fix.

Where the white-collar week actually goes

Before you buy anything, spend one week measuring. Ask five back-office people to keep a rough tally of their hours against these five buckets. The point is not precision. The point is to find out which bucket is largest, because that is the bucket you automate first.

• Copy-paste between systems. The same order number, customer name or invoice figure typed into two or three places. This is the most common and the most automatable line on the list.
• Filling forms. Purchase requests, leave requests, expense claims, customer onboarding packs — usually paper or a spreadsheet template emailed around.
• Chasing approvals. Time spent finding out who has to sign, whether they have signed, and reminding them that they have not. This work produces no output at all.
• Filing and retrieving documents. Renaming PDFs, dropping them into folders, then searching for them again three weeks later during an audit.
• Assembling reports. Exporting to CSV, pivoting, formatting, pasting into a deck, emailing it every Monday morning.

To convert those hours into a currency figure before you decide anything, work through the cost of manual work using your own payroll numbers.

Workflow engine or AI agent: which task belongs where

These are two different tools, and buying the wrong one is the usual reason automation programmes stall. A workflow engine executes rules you have written down. An AI agent handles input that has no fixed shape. Most real back-office processes need both, in sequence: the agent reads the messy input, the engine executes the deterministic part.

If the middle column reads "AI agent" for most of your list, read how we integrate AI into existing systems before scoping the build. If it reads "workflow engine", start with process automation.

The 5-question test: is this task automatable?

Run every candidate task through these five questions before it enters a backlog. A task that scores badly on questions 3 and 4 will consume more engineering than it returns, however painful it feels day to day.

1. 1. How often does it repeat?

   Count occurrences per month, not minutes per occurrence. A two-minute task performed 400 times a month is a better candidate than a two-hour task performed once a quarter. Below roughly 20 occurrences a month, an automation rarely pays for its own maintenance.

2. 2. How clear are the rules?

   Can someone write the decision logic on one page without using the phrase "it depends on the situation"? If the rule lives only in one person's head, your first task is to write it down. That document is a deliverable in its own right, and it is the part buyers most often skip.

3. 3. What is the exception rate?

   Out of 100 items, how many break the standard path? A 5% exception rate means straight-through processing for 95 items and a short handover queue for the rest. A 40% exception rate means you are automating chaos, and the exception queue becomes the new bottleneck. Redesign the process first.

4. 4. Is the input format standard?

   A structured API response, a fixed CSV, a database table — automatable with a workflow engine today. A photographed delivery note, a free-text email, a scanned PDF arriving from 40 different suppliers — also automatable, but with an AI extraction layer and a human confirmation step in front of it. Both are viable; they carry different costs and timelines.

5. 5. What does an error cost?

   Quality management describes error cost with the 1-10-100 rule: an error caught at data entry costs one unit to fix, caught downstream ten, and caught by the customer a hundred. That is a general industry heuristic, not a KodDelta measurement, but it is the right way to decide how much validation to build. Where the error cost is high you build approval gates and audit trails, and the automation costs more. Where it is low, ship the simple version.

What good looks like in practice

Automated does not mean unattended. Every process we build has three visible states: items that completed straight through, items sitting in an exception queue with a named owner, and items that failed with a logged reason. A system that cannot show those three numbers on one screen loses its users by month two.

Design decisions that matter more than the technology choice:

• One process first. Not a department, not a platform. The first module ships in 2–4 weeks and starts returning hours while the second one is being specified.
• Named exception owners. An exception queue with no owner is a slower inbox.
• An audit trail per item. Who or what changed the record, when, and from what value. Auditors will ask, and so will the first person who disbelieves a number.
• Integration over replacement. The systems that already hold your data keep holding it. See how we connect existing systems.
• A measured baseline. Record current cycle time and error rate before go-live, or you will never prove the change afterwards. The ROI guide covers that measurement discipline.

Systems built this way include Elevatora SAHA for lift maintenance and field service management, Elevatora İMALAT for lift-component manufacturing tracking, and PAP for pharmaceutical import and export process tracking. Each started as one process and grew module by module.

Budget and timeline you can plan against

• Single focused module: $3,000–6,000, 2–4 weeks. One process, straight through, with an exception queue.
• Multi-department system: $8,000–15,000, 4–8 weeks. Several processes sharing one data model and one permission structure.
• Platform with AI on top: $20,000+, 3–6 months.
• Annual maintenance: optional, 12–25% of build cost.
• Licence: perpetual, unlimited users, source code delivered.

The first clickable prototype arrives within 2–4 weeks of kickoff, so the decision to continue is made against something your team has used rather than a specification document. If the choice in front of you is a product licence versus a build, the package-versus-custom comparison covers that separately.

Bring the largest bucket from your one-week tally and we will scope the first module against it. Request a quote, or compare the delivery bands on the pricing page.

Frequently asked questions

What counts as back office automation?

Any software that performs an administrative task a person currently does by hand: moving data between two systems, filling a form, routing an approval, filing a document, or assembling a recurring report. It covers finance, HR, procurement, order administration and shared services.

Should I use a workflow engine or an AI agent?

Use a workflow engine when the rules are written down and the input format is fixed — approvals, routing, scheduled reports, system-to-system syncs. Use an AI agent when the input is unstructured text or a scanned document and the task needs classification, extraction or drafting before a rule can be applied.

How much back-office work can realistically run without a person?

That depends on your exception rate, not on the technology. A process where nine out of ten items follow the same path can run straight through, with the remaining items routed to a named person. A process where every second item is an exception should be redesigned before it is automated.

What does a first back-office automation module cost at KodDelta?

A single focused module is $3,000–6,000 and ships in 2–4 weeks. A multi-department system is $8,000–15,000 over 4–8 weeks. Annual maintenance is optional at 12–25% of build cost. The licence is perpetual with unlimited users and the source code is handed over.

Do we have to replace our existing ERP or accounting system?

No. Most back-office automation sits between systems you already run and talks to them through their APIs, database views or file exports. Replacing a working system of record is a separate and much larger decision.