The test: must it come out the same every time?
Traditional software follows rules a person wrote. Give it the same input on Monday and on Friday and it does the same thing, which makes it predictable, easy to audit and inexpensive to run for each action. An AI agent works toward a goal and chooses its own steps. That flexibility is useful, and it also means two runs can differ.
So the first question about any job is not how clever the tool is. It is whether the job tolerates variation. A reply to a customer can be worded a hundred ways. A pay run cannot. Where the answer must be identical and explainable, rules win, and no amount of intelligence changes that.
- The same input must always give the same output
- You must be able to show exactly why it did what it did
- A wrong answer costs money or breaks a rule
- The data is already tidy: amounts, dates, quantities, codes
Payroll and other money that must be exact
Say a landscaping company in Abbotsford hires seasonal crews every spring. Hours, rates, deductions and remittances follow rules set by law and by your agreements. Payroll software applies them the same way for every person, and it keeps a record you can hand to your accountant. An agent that reasons its way to a number is the wrong tool for that job.
That does not shut agents out of the office. One can read a messy timesheet email, notice a missing entry and ask a supervisor to confirm it. But the calculation, the approval and the filing stay in software with fixed rules, and your accountant stays the authority on what those rules should be.
Invoices and inventory counts
Imagine a farm supply store reconciling deliveries against what is on the shelf. A stock figure is arithmetic: what came in, minus what went out, adjusted for counted losses. A database records each movement and always gives the same total. If a language model estimated that figure, you would not know it was wrong until a shelf ran empty.
Invoices work the same way. Pulling the numbers out of a crumpled supplier PDF is a fair job for AI, with a person checking the doubtful ones. Posting the amount, applying the tax setting your accountant chose and matching it to the purchase order is a job for rules. Let the agent read, and let traditional software decide and record.
Compliance records and audit trails
Some records exist so that someone else can inspect them: a safety log, a cold-storage temperature sheet, a signed work order, a history of who changed a customer's details and when. An inspector, an insurer or an auditor wants a record to mean the same thing next year. Rule-based systems with fixed fields, permissions and change history give you that.
Privacy belongs here too. Personal information in your care falls under PIPEDA and BC's Personal Information Protection Act, and who may see what is easiest to control in a system built with clear permissions. This is general information, not legal advice, so involve your lawyer when records hold personal details. Our AI governance and security work starts from the same point.
Where agents still earn a place
None of this makes agents a poor idea. The jobs suited to them sit around the rule-based core: reading a supplier's odd email, drafting a customer reply, sorting a shared inbox, summarizing a long file for a manager, or flagging an invoice that looks wrong. The system of record stays traditional, and the agent works through narrow permissions with a log of each step.
If you are unsure where a job falls, we can sort your list with you. Our traditional software development and web app development teams build the rule-based side, and we can add supervised agents beside it. Book a free consultation by phone, by video or at the Abbotsford studio, and bring the three jobs that worry you most.
Drawing the line as you plan for 2027
A useful exercise for next year's planning is a page with two columns. In one, list every job that moves money, counts stock or creates a record someone may inspect. In the other, list jobs that involve reading, writing and judging. Jobs in the first column usually stay with rules, and the second column is where a supervised agent trial makes sense. The page also becomes the start of your access rules.
- 01Mark which jobs touch money, stock or inspected records
- 02Choose one language-heavy job for a supervised agent trial
- 03Write down who approves what an agent is allowed to do
Questions & answers
No. Rule-based systems still run payroll, accounting, inventory and booking because they behave identically every time and leave a clear record. What is changing is that language-heavy work around them can now be assisted by agents. Most businesses end up with both, each on the jobs it suits.
It can help around payroll, such as chasing a missing timesheet or flagging an entry that looks odd, but the calculation and filing should stay in payroll software with fixed rules. Ask your accountant how your pay runs must be handled, and keep a person approving each pay run before it is final.
Often, for repeated work, because a rule costs very little each time it runs. An agent often costs more per task because each run uses a language model, and the total depends on how much it does. Costs vary by tool and volume, so ask for an estimate based on your own workload.
Usually not. If the old system holds your records correctly, the better step is to modernize it, tidy its data and connect it, then put an agent beside it for the reading and writing jobs. Replacing a system of record with an agent risks losing the exactness the records need.
