Why AI Can't Make Your Rota for You

Of course AI can make your rota for you.

Give a capable model your opening hours, a list of employees, some availability and a rough idea of how many people you need, and it will happily produce something that looks suspiciously like a rota.

Hell, it might even be a pretty decent one.

That still doesn't mean you should let it run your week.

Scheduling looks like an ideal AI problem from the outside. There are people, hours, roles, rules and constraints, all waiting to be shuffled into the most efficient combination. Feed enough information into a machine, ask it to optimise the outcome and surely the computer can take one more miserable admin job off the General Manager's desk.

I understand the appeal. I use AI constantly, and GlassRota itself would probably not exist in its current form without AI accelerating parts of the work behind it.

But there is a fairly important distinction between using AI to help someone make a decision and allowing AI to quietly make employment decisions for them.

A rota affects somebody's paycheck, sleep, mental and physical wellbeing, career development, workload, and home life. It also determines which skills are present when the venue is busiest, which manager carries the close, who gets an opportunity to step up and whether your Saturday night has any realistic chance of working.

That is a surprisingly personal set of consequences to hand over to something that has never actually worked a Saturday night.

The rota looks like a maths problem until you get to know the team

Imagine you give your best pal Claude (or any other AI model) all the context it needs to make a compliant schedule.

Sarah is available Monday to Saturday.

Dave can work bar and floor.

Mary is a supervisor.

Tom wants 30 hours.

The bar needs five people during Friday peak.

The kitchen needs four.

Easy enough.

Now add the things an experienced manager knows without necessarily writing them down.

Sarah has technically marked herself available on Saturday, but she's worked the last three full weekends and asked you a fortnight ago if she could get this one back.

Dave can work the floor, although on a packed Saturday he's dramatically stronger behind the bar.

Mary can close by herself and has done several times, but you're deliberately pairing her with your strongest manager this week because you want to move her towards an Assistant Manager role.

Tom wants 30 hours, yet he's just returned from two weeks off sick and you'd rather not immediately give him six consecutive days.

Two of your bartenders are individually excellent and, for reasons known only to God, seem to be sharing 5 braincells between them whenever they work together.

None of that makes the original data wrong.

It makes the operation richer than the data.

Experienced managers carry thousands of these tiny pieces of context. Some are objective, some are judgement calls, some are temporary, and a surprising number would look utterly ridiculous written inside an employee database.

That's hospitality.

The machine can know the rules better than the manager ever will. The manager will know their team better than a machine ever could.

The more context a system has, the more useful its recommendations can become. There will still be things a good manager knows that either haven't been entered into software or probably shouldn't be.

A rota is too personal to become a black-box decision

This connects directly to something we explored extensively in Why Your Company Values Mean Nothing to Your Team.

The rota is one of the most tangible ways an employee experiences the business. It determines when they work, how much they might earn, which weekends they lose, which opportunities they receive and how often the company asks them to carry the difficult shifts.

Now imagine an employee asks:

"Why am I closing again this weekend?"

And the manager replies:

"The AI scheduled you."

Terrible answer.

Try another:

"Why has James got 35 hours and I've got 18?"

"That's what ChatGPT decided."

We've now taken one of the problems that creates bad rota cultures in the first place, employees feeling like numbers on a spreadsheet, and industrialised it.

Automation doesn't remove accountability.

A manager might use software to calculate, check, recommend and surface information. The moment they publish the rota, though, the consequences belong to them.

If somebody's income changes significantly, an actual human should be capable of explaining to them why.

If somebody has been given another exhausting weekend, somebody should have considered the pattern.

If the only answer available is “the AI considered a few variables and this was the most optimal”, the technology has actually made the decision harder to challenge rather than better.

Black box or Glass box?

This was one of the earliest philosophical decisions behind GlassRota.

A black-box system takes information in and produces an answer.

You ask for a rota.

It gives you a rota.

Perhaps it scores hundreds of possible combinations internally. Maybe it uses a language model, an optimisation engine or a collection of algorithms that nobody at venue level particularly understands.

The result can still be excellent.

The problem is when the manager doesn't know why it happened.

A glass-box approach tries to expose the thinking.

Instead of simply saying:

Move Sarah off Saturday.

A useful system should be able to tell you what it noticed.

"Sarah has worked three consecutive full weekends.

Her scheduled hours are above the range you've set.

James has the required skill, is available and currently has fewer hours."

Saturday still retains the level of experienced coverage you've told the system the venue needs.

Now the manager has something to work with.

They may still ignore the suggestion.

Perhaps Sarah specifically asked for Saturday because she needs the money. Maybe James has something going on that has never been entered into the system. The software cannot know that unless somebody tells it.

What it can do is make sure the manager sees the trade-off before pressing publish.

That is a very different relationship between person and machine.

GlassRota doesn't imagine what your venue needs

This is where the distinction becomes slightly technical, so let's keep it in layman's terms.

GlassRota's Validation Engine works with the information the venue gives it.

You define things such as operating hours, service periods, roles, staffing requirements, availability, employee hours, skills, wage information and the other boundaries that shape the operation.

The engine then checks the rota against those inputs using mathematical rules and logic.

If someone has insufficient rest, the maths can identify the gap.

If Saturday peak requires three bartenders and you've scheduled two, the engine can see the missing coverage.

If one employee is above the hours you've set for them while somebody else is short, that can be surfaced.

If a particular period is overstaffed against the requirement you've defined, the system can show the potential cost.

The useful bit is that those conclusions come from your inputs.

GlassRota shouldn't decide that your cocktail bar requires four bartenders because it once read something about cocktail bars on the internet.

If you tell it three, the calculation starts from three.

Give the system more relevant context and the checks can become more useful. Give it poor information and you should expect poorer output.

There is no magic escape from garbage in, garbage out.

Personally, I find that reassuring.

Why it feels much more 'AI' than it actually is under the bonnet

Auto-Assign is probably the GlassRota feature most likely to make someone say:

"Oh, so this is AI."

Not really.

You can ask it to build out a schedule and watch it make a surprising number of decisions very quickly. It considers things such as availability, skills, hour boundaries, coverage requirements and other constraints, then tries to find a sensible arrangement.

It certainly feels intelligent.

Underneath, however, think less Skynet has taken control of the rota and more Excel if it could speak plain English, immediately after downing three Red Bulls, and juggling while riding a unicycle.

It's just maths. Lots of maths, interdependent logic, priorities, and rules being evaluated much faster than a person could ever dream of doing manually.

Smart Suggest works on a similar principle. When you need to cover a shift, it can evaluate eligible people against the information already in the system and recommend sensible candidates. If you remember learning IF/AND/OR statements in IT at school, imagine that, but with thousands of data points, across hundreds of shifts, across months of scheduling information, for up to 250 employees per venue. Riveting stuff.

The aim isn't:

"The computer has chosen Sarah. All hail the computer."

It's closer to:

"Here are the people who appear to make sense for this shift, and here are the reasons why. You know the team. You make the call."

A manager can disagree.

In fact, sometimes they absolutely should.

Generative AI has another problem: it is extremely good at sounding right

Modern language models are remarkable partly because they generate convincing answers.

That is incredibly useful.

It is also exactly why incorrect answers deserve care.

A plainly broken spreadsheet formula announces itself fairly quickly. The number looks ridiculous and somebody asks what happened.

A fluent AI answer can be more dangerous because it can sound completely reasonable.

Imagine an AI scheduler confidently explaining that Sarah has been assigned Sunday because she worked fewer weekend hours than the rest of the team.

Sounds thoughtful.

Except Sarah actually worked a private event last Sunday which wasn't included in the data you gave it.

The reasoning is logical.

The conclusion is wrong.

The model isn't malicious or stupid. It simply operated on incomplete information and produced the best-looking answer it could from what it had.

Managers do that too, incidentally.

The difference is that we expect managers to recognise when they don't know something, speak to the team and apply context.

The acceptable tolerance for a machine quietly guessing should be much lower when somebody's livelihood sits on the other side of the answer.

If a chatbot recommends the wrong pizza restaurant, everyone survives.

If your scheduling system invents confidence around an employment decision, that deserves considerably more scrutiny.

Software can recommend. Managers remain accountable.

This is the part I think hospitality needs to understand as AI becomes more embedded in management software.

Delegating work and delegating responsibility are different things.

A calculator can work out wage percentage for you. You're still responsible for understanding what the number means. If you want to play with the numbers yourself, our free Hospitality Wage Percentage & RPH Calculator does exactly that.

A forecasting system can predict Saturday's sales. The manager still decides whether the rota makes sense against the reality they can see.

An AI assistant can summarise twenty pages of information in thirty seconds. Somebody still needs to care whether the summary is accurate before acting on it.

Scheduling should work the same way.

Technology can remember more than a manager.

It can run calculations faster.

It can compare hundreds of combinations without getting bored, distracted or interrupted by somebody asking where the spare till rolls are.

All very useful.

The final judgement still needs a human being.

Particularly when the decision is about another human being.

I'm not anti-AI, I promise!

This would be a painfully hypocritical article to write if I were.

AI has extraordinary potential in hospitality, especially in an industry where managers spend far too much time processing information and far too little time using it.

Imagine a system reading a month's worth of your end-of-night reports and pointing out the recurring problems nobody had noticed. That is exactly the sort of problem we're exploring with EON Report.

Or taking thousands of guest reviews and identifying that complaints about Friday service have increased specifically between 7pm and 9pm.

It can help managers interrogate P&Ls, summarise long internal documents, draft training material, translate communications for multilingual teams, analyse feedback, prepare handovers or pressure-test a difficult conversation before they have it.

Forecasting is another fascinating area. Historical sales, bookings, weather, local events and other signals create far too much information for a busy GM to manually process every week. AI could help turn that noise into something a manager can actually use.

Those are exciting applications precisely because they give the person more useful information to make a decision with.

There's a simple principle to apply:

Use AI to reduce the amount of information a manager has to process, not the amount of judgement a manager has to exercise.

That's where I think technology becomes genuinely powerful.

Memory. Maths. Monotony.

There are three things machines have become spectacularly well suited to in rota management.

Memory.

A manager shouldn't need to remember that Dave has worked three full weekends, Sarah is approaching her maximum hours and Mary's availability changed last Tuesday.

Software can remember for you.

Maths.

A human being shouldn't have to mentally calculate dozens of overlapping shifts, wage costs, hour balances, staffing requirements and rest periods while trying to finish the rota.

Software can calculate for you.

Monotony.

Checking the same rules across 150 shifts is exactly the sort of repetitive job computers were invented to do.

Let them.

Then preserve the three things hospitality managers bring that can't be reduced to a formula:

Meaning. Nuance. Judgement.

The machine can flag that Sarah has worked three weekends.

The manager understands what those weekends actually felt like.

The machine can see that James has fewer hours.

The manager knows James is deliberately working fewer hours while finishing his exams.

The machine can recognise that Mary meets the requirements to close.

The manager knows whether this Saturday is the right night to let her do it.

Those layers work brilliantly together.

Problems begin when one pretends it can replace the other.

Some context should probably never go into the software

There is another uncomfortable implication of fully automated scheduling.

To make a machine perfectly reproduce an experienced manager's judgement, how much information would you need to give it?

Sarah is going through a divorce.

James is anxious about his first management shift.

Tom and Dave had an argument last weekend.

Mary's confidence has taken a knock after a guest shouted at her.

You suspect another employee is preparing to resign.

Someone's parent is ill.

Suddenly, in pursuit of the perfect automated rota, we've created a system containing a disturbingly detailed model of everyone's personal life.

No thanks.

Some context belongs in software, some belongs in a conversation, and some should remain in a manager's head because employees are entitled to be human beings without every piece of their life becoming another scheduling variable.

There should always be room for judgement outside the system.

A very smart program built by very normal people

There is a temptation in technology to make everything sound more magical than it is.

Artificial intelligence.

Machine learning.

Optimisation engines.

Predictive this.

Autonomous that.

You half expect the scheduling software to become self-aware and ask for a pay rise.

GlassRota came from somewhere much less glamorous.

It came from real restaurant, bar, club, and pub operators getting frustrated writing rotas at 11pm the night of the deadline.

Frustrated at how much information you have to hold in your head while writing a rota.

Frustrated at noticing stupid mistakes after you've already published it.

Frustrated that trying to save labour in the wrong place can make the entire service worse.

Frustrated that your strongest people who quietly carry the team are quitting because they're exhausted too.

Frustrated at typing the exact same shifts into another system after you've already perfected them once.

We didn't need a computer that pretended it had run a bar.

We needed one that was extremely good at remembering the boring stuff, doing the maths and pushing back on things we might want to look at again.

That's what GlassRota is trying to be.

Very smart software, certainly.

Still just software.

Built by very normal operators who have spent enough Saturday nights wondering what idiot wrote the rota, only to remember it was them.

See what a rota that argues back feels like

GlassRota won't quietly decide your week and ask you to trust it.

Build the rota yourself, give the system the operational context you want it to use, and let it challenge the bits that deserve another look.

Try GlassRota free for 14 days →

You can ignore its suggestions.

Occasionally you probably should.

That's kinda the point.

The goal was never to replace the manager

Hospitality doesn't need fewer people exercising judgement.

If anything, we need to give managers enough time and information to exercise it better.

AI will absolutely become part of that future. Anyone betting against the technology entirely is probably going to have an uncomfortable decade.

But the opportunity isn't to outsource every decision simply because a machine can produce an answer.

The opportunity is to understand which parts of management are genuinely improved by automation and which parts derive their value from having a human being involved.

A rota sits right across that line.

The calculations can be automated.

The rules can be checked.

Patterns can be surfaced.

Options can be suggested.

The final responsibility should still belong to the person who knows the team, understands the operation and will be standing beside them when Saturday night arrives.

GlassRota doesn't want to make your rota for you.

It wants to make sure you've thought about the bits you might have missed.

Very smart program. Very normal operators. Human judgement still required.