Two ways software can do the work
Fixed rules. Someone writes down what to check and what to do about it: "if a shift is still open 45 minutes after the alarm arms, text that store's manager." The same facts always give the same result. You can read every rule, test it against a week you remember, and explain any result afterward by pointing at the rule that fired and the data it read. The limit is just as plain: it only handles what someone thought to write down. A case no rule covers is skipped, or left for a person.
An AI agent. A language model is given a goal, access to your data and a set of tools, and works out the steps itself: what to look up, what to compare, what to draft and, if it is allowed, what to change. It copes with work fixed rules can't do: an open question ("why was labor high at the north store last Tuesday?"), a messy invoice or email, a situation nobody wrote a rule for. The trade-offs are real too. The same request can get a different answer on a different day. It can be confidently wrong. And it can be steered by text it reads along the way, such as an instruction hidden in an email or a note in a record.
Neither is better in general. They fail in different ways, and payroll and cash are where the way a tool fails matters most.
| Fixed rules | An AI agent | |
|---|---|---|
| Same facts, same result? | Yes, every time | Not always |
| Explaining a result | The rule that fired and the data it read | The agent's account of its steps, which can be hard to reproduce |
| Something unexpected | Skips it, or stops and asks | Improvises |
| Testing before you trust it | Run every rule against past weeks | Spot-check samples, and again whenever the model or its instructions change |
| Typical failure | Misses what no rule covers | Acts on a wrong reading |
| Best at | Repeating checks with clear thresholds | Questions, summaries and messy documents |
What to let software do with payroll
Payroll mistakes land on people's paychecks and in your tax filings, and they are slow to unwind. A sensible split:
Software on its own
- Read hours from the time clock, the point of sale and payroll
- Match them person by person and store by store, and flag every gap
- List open shifts, missed clock-outs and punches outside the schedule before the week closes
- Track which stores are posted, clean or still blocking payroll
Software prepares, a person approves
- Fixing a punch. The software proposes the time and shows its evidence; the store manager approves it.
- Building the payroll import, balanced to the cent against the time clock, for a person to review.
- Explaining a variance. An agent is good at this, as long as nobody acts on the explanation without checking it.
A person only
- Approving and submitting payroll
- Changing a pay rate, adding someone to payroll or ending their employment
- Anything else that changes what someone is paid
Software can make payroll much faster to check, but the approval should be a person's, every run. That holds most of all for an agent: its strength is handling the unexpected, and an unexpected payroll is exactly the one you want a person to see.
What to let software do with cash and money movement
Reading money is cheap insurance. Moving it is a different job.
Software on its own
- Read balances and transactions through read-only access
- Match each store's deposits to its sales, allowing for weekends and card settlement lags
- Flag a missing or short deposit, and unfamiliar charges on company cards
Software prepares, a person confirms with a typed amount
- A transfer between your own accounts. The software can show the deposits it matched and the balances it read. The person types the amount. A typed amount is much harder to give on autopilot than a tap on "approve".
A person only
- Paying bills, vendors or card balances
- Sending money outside your own accounts
- Adding a payee, or changing bank users, limits or alerts
For an AI agent the line is simple: don't give it credentials that can move money. An agent that reads email and invoices can be fooled by a convincing fake ("our bank details have changed, please pay today"). If it can't move money, it can't be talked into moving it. Confirm any change to a payee's details by phone, on a number you already had.
Start read-only
Read-only access is the cheapest safeguard there is, and it works the same for either kind of software.
- A separate user for the software in each system. Never share your own login. You create the user, you own it and you can remove it at any time.
- View-only roles, checked. Find out what "view" means in each system. In some, a viewer can still export employee data or change settings.
- Bank data through a read-only connection. A read-only data connection or a view-only bank user, never online banking credentials that can send payments.
- Write access one automation at a time. When you later switch on something that changes a system, grant the access that one job needs. Approving one doesn't approve the rest.
Run it in shadow mode first
In shadow mode the software runs on its real schedule, on your real data, and logs what it would have done without sending or changing anything. Then you compare its log with what your people actually did.
- Long enough to see your unusual days: a payroll week, a busy weekend, a holiday, a store that closed early. Two weeks is a sensible minimum.
- Prove it was silent. Check that nothing was sent and nothing changed, not just that the send step was switched off.
- For an agent, run the same week twice. If two runs over the same data disagree, you have learned how far to trust either one.
There is more on stages and kill switches in Shadow mode: how to trust automation with your stores.
Insist on an audit trail
Every change software makes should leave a record you can read without the vendor's help:
- who or what asked for it, and who approved it
- what the software read before it acted
- what it changed, before and after
- what the system showed when it was read back afterward
For an AI agent, also ask for the instructions and inputs it worked from and each tool it used. If a vendor can't show you why a change happened, assume nobody will be able to on the day it matters.
"Done" should mean the software re-read the system and saw the new value. "No error" is not the same thing: systems accept a request and then drop it, save it to the wrong day, or get overwritten by a sync an hour later.
Questions to ask any vendor
- Which parts run on fixed rules, and which on an AI model?
- Is my business data sent to an AI model, and is it used for training?
- What can it change in my systems, and what is read-only?
- Can it approve payroll or move money on its own? The answer you want is no.
- Can each automation run in shadow mode first, and for how long?
- Can I switch off one automation without stopping the rest?
- Does it re-read the system after every change?
- Can I see and export the audit trail, and how long is it kept?
- What does it do when it isn't sure? It should stop and ask.
- Who finds out when a run fails, and how fast?
A sensible default
| Task | Who decides |
|---|---|
| Reading hours, sales, deposits and balances | Software, with read-only access |
| Flagging missed clock-outs, short deposits and odd card charges | Software |
| Fixing a punch | Software proposes; the store manager approves |
| Building the payroll import | Software prepares it, balanced to the cent; a person reviews |
| Approving and submitting payroll | A person, every run |
| Changing a pay rate | A person (software can flag a rate that drifted) |
| A transfer between your own accounts | Software prepares; a person types the amount |
| Paying bills, vendors or card balances | A person |
| Adding a payee or changing bank users | A person |
The table holds whether the software runs on fixed rules or an AI agent. What changes with an agent is how hard you test it first and how closely you read its log.
Where Poros Ops stands
Since you are reading this on our site: Poros Ops runs on fixed rules, and business data from your connected systems is not sent to an AI model. It connects read-only first, runs every automation in shadow mode for at least 14 days, re-reads the system after every change and logs every write. Payroll approval and moving money always stay with a person. The details are on Security & data.
Questions
Should software ever approve payroll?
No. Software can read the hours, match them across systems, build the import and flag what is wrong, but approving and submitting payroll should be a person's job every run, whether the tool runs on fixed rules or an AI agent.
Is it safe to connect software to my bank?
Read-only access is: the software can see balances and transactions but cannot move money. Don't give any tool online banking credentials that can send payments, and confirm any change to a payee's details by phone.
Which is better for restaurant operations, fixed rules or an AI agent?
It depends on the job. Fixed rules suit repeating checks with clear thresholds, like missed clock-outs or matching deposits to the bank. An AI agent suits open questions and messy documents. For payroll approval and money movement, neither should act alone.
What is shadow mode?
A trial run on your real data in which the software logs what it would have done without sending or changing anything. You compare the log with what your people did before you let it act.
See it on your stores
If you'd like to see what fixed-rule checks would catch on your own systems, in shadow mode and before anything writes, book a walkthrough.
Book a walkthrough