A few thoughts on why algorithms get bad press.
We came across a news article last week (“Here’s What Happens When an Algorithm Determines Your Work Schedule”, Vice, 24 Feb 2020), in the vein of others we’ve seen before (eg “Scheduling Technology Leaves Low-Income Parents With Hours of Chaos”, NY Times, Aug 2014), that made us sad. 😔 If you’re thinking of making the leap from ‘manual’ rotas to optimised, automated scheduling in your business, then maybe such press coverage of our industry also makes you wonder whether this is intrinsically a change that can improve efficiency only at the cost of employee wellbeing.
It’s painful for all of us at Rotageek to hear stories of employees struggling with their schedules. Working “long and unpredictable hours”, with “no way to guarantee week-to-week certain days off”, and “shifts being added same-day or dropped mid-shift" - this is why most of us got into this business. We believe there’s an opportunity, now, to use the power of modern computing and apps to improve the experience of being a shift-worker employee (alongside other wins). These are real problems in shift-worker employer businesses that need solving, and whilst part of our proposition is about capturing that top 5-10% of labour inefficiency that comes from not being fully optimised, our starting point is always the ‘lose-lose’ elements of poor processes that result in a more painful experience for the employee than it needs to be.
But beyond that, I thought there were four main reasons we don’t think it’s a “zero sum game” between employees and employers trying to achieve better rota efficiency with the help of algorithms, based on our experience over the last five years at Rotageek. 🚀
(1) The ‘exploitation’ mindset isn’t what we’ve seen from our adopters
I guess the first thing to say is that, when algorithmic scheduling is perceived by employees as optimising at the expense of their wellbeing via things like “shorter shifts during the week in favour of longer weekend shifts”, it’s obviously entirely possible to configure the algorithm in a way that simply doesn’t allow, or prioritises the avoidance of, such things. (At least, that’s possible in ours). So if and where that’s happening, then it’s (on some level, at least) a choice being made by the business that is implementing it.
So I think an important myth to bust, before we get any further into what’s technically possible with algorithms (and of course what “unintended consequences” might befall), is that this a common agenda behind adoption of such tools in the first place. In all honesty, our experience to date has mostly been of employers massively focusing on using the tool to improve fairness and workability for staff, well ahead of improving customer metrics or cost. That’s visible in both the objective “weightings" placed on rules in pretty much every deployment we’ve ever done (EWTD, key HR policies and observance of things like employee unavailability tend to be programmed as ‘hard constraints’; and then most fairness and workability constraints that management and surveyed local leadership can think of get ranked as ‘high’ priorities; and only as a sort of tie-breaker between solutions that comply with all of that does ‘fit-to-demand’ get considered), but also mindsets in the way that project goals are set.
I can’t speak for all of the organisations referenced in the Vice and NY Times (and similar) articles, but this really doesn’t ring true for the employers that Rotageek has so far worked with. So either the outcomes reported in those sorts of articles are (a) inadvertent, (b) unusual, (c) not really to do with the algorithmic side of things in the first place, or at the very least (d) not something you have to emulate in your own implementation. Part of the difference may be down to the power of the algorithm chosen, to achieve what the employer wants to achieve: we’re aware that many of our competitors’ algorithms “automate as opposed to optimise”, including for example creating the shifts to be assigned first and then doing their best to assign them as a separate second step, which can result in far inferior results in terms of both fairness and supply-matching, and some tougher trade-offs for the implementer to make. But with the right mindset and a sophisticated enough algorithm, it doesn’t have to be a zero-sum game.
We’ve actually been amazed over the last five years by the extent to which companies adopting our solution seem to prioritise employee wellbeing over pure efficiency gains, as they go through this change - it never seems to be perceived as an opportunity to tip the employer-employee power balance.
(2) Wellbeing can actually be optimised too
There is often a very measurable potential to improve the “fit to demand” of a store manager’s manual schedule, by, for example, starting a few shifts later here, a few earlier there, going down to 15-minute granularity rather than always have people start and finish on the hour, avoiding simply pasting in the same schedule every week or having people rotate through 4 or 6 or 8 day repeating ‘patterns’: the ‘optimised’ version, unshackled by any of those heuristics, will often give you more staff at times when today queues are building up, fewer at times when today staff are a little under-utilised, and thereby allow you to “do more with less”. And that’s of course a key reason to move to automated scheduling, and something people generally have no trouble believing the computer is ‘better at’ than the human. There are just an awful lot of options to consider, once you start to entertain all the possibilities; the computer can systematically cover, quantify and remember many more of them.
But the same is largely true of employee wellbeing considerations too. For example, it’s easy for a human to miss that they’ve scheduled someone to close up the shop at 10pm and be back in to open it at 7am (a breach of the European Working Time Directive, an amount of rest between shifts that tends to be unpopular with employees and likely also won’t have them performing at their best), and we see instances of that pretty often in the “pre Rotageek” rotas of customers who trial our product. But you can easily set the algorithm to simply never even consider doing this, and it never will. If, under company norms or HR policy or just from asking and finding that this matters to your particular population of employees with the kind of work that they do, you think it’s valuable that full-time employees who work 5 days in 7 should have their 2 days off ‘consecutively’ rather than singly, then you can (and we have customers who do) weight this highly, and the algorithm will ensure that this happens 60%, 80%, 100% of the time. If the reason that some or many of your managers are scheduling people to those 8-day patterns is it’s the only reliable way to ensure each one works their ‘fair share’ of the ‘Lates’ or the weekends, then you should consider how easily an algorithm could ensure the exact same thing over an 8-week cycle, without the blunt instrument (and use the new-found flexibility to hit demand better, or to also see how well it can give people consistent start-times in a run of consecutive working days, for example).
The idea that humans are great at fairness and stability and good adherence to labour laws/HR policies, but the computer is only good for optimising to demand, is just the wrong premise. This change can be as much about improving wellbeing as maintaining it (or not) while moving the dial on efficiency.
(3) Efficiency per se is also in employees’ interests
Insofar as part of the objective of moving to data-driven, optimised scheduling for the employer is to do that “fit to demand” bit better (sell more products, reduce labour costs), employees shouldn’t necessarily see this as a bad thing. Efficiencies improve the business’s chance of surviving and thriving, improving their job security and satisfaction. And of course, being rushed off your feet in an understaffed team at busy times while feeling under-utilised and unengaged at others, up and down in the course of a single day or week, isn’t how anyone wants to work.
But if this comes at a ‘cost’ of worsening wellbeing for employees, then it stops being a no-brainer. One complaint about optimised scheduling that was quoted in the most recent article is “unpredictable hours”. For sure, a key tool in the armoury of the optimised scheduling algorithm is the ability to break away from “predictable” hours, in the sense of, the same every week (or pattern) - the forecasted busy-ness of the store, and the availability of other team-mates, given leavers and joiners and holiday and sickness, isn’t the same profile every week, so unless it’s specified as a rule going in, the optimised schedule is highly unlikely to produce identical rotas from one period to the next. You might think this is the tit-for-tat that comments like this are getting at. But what we generally see (and our research supports) is that when employees talk about the value of “predictable” hours, they tend to either mean the ability to earmark times as “unavailable” (the same slots every week or ad hoc in advance), or a decent amount of forward visibility (schedules published well in advance). But patterns and copy-pastes aren't the only way to solve those problems! By digging a little deeper into what employees really value, we find we can typically define solutions that allow businesses to pursue efficiency without inherently reducing employee control over their commitments (to the extent that the company has ever and is willing to grant that), ability to plan their lives, and dependability of their earnings.
(I didn’t see many other examples in the article, actually, among the employee ‘pains’, of things which were clearly done in the interest of better-satisfying “seasonal sales patterns, customer trends, and even the weather” and at the detriment of employee wellbeing. Many of the issues were with things like “shifts being added same-day or dropped mid-shift", which sound to us more like issues with the business process around it, including tolerance of what sound like ‘bad’ manager behaviours, despite these themselves being open to better prompting and monitoring once the business was on a digital system, with or without the optimisation. And actually, some things like fewer last-minute changes to the rota or allowing more ad-hoc preferences for days on or off for the employee, would rather be enabled by the flexibility of a pattern-less, better-optimised-at-the-outset, schedule).
Both employer and employees have an interest in the business surviving and thriving - you just have to define the limits of this, which of course rule-based software does allow (help) you to do.
(4) In reality, you can usually improve both at once
These are complicated jigsaws, to staff roles in a way that wastes as little as possible and gives the business the best chance of survival and growth, at the same time as taking into account the full gamut of employee wellbeing considerations (booked Leave and agreed Unavailability but also the days they “Prefer” and “Prefer Not” to work, the history of lates and weekends worked versus their comparable peers, all the little things that make a schedule less ‘sustainable’ over time such as volatile waking-up times, too much variety or too much similarity in the roles being assigned, too many or too few rest days coming consecutively between runs of days on-shift, and so on). In our experience, computers are simply much better at them than humans.
Exactly what you take into account, and exactly how you prioritise each dimension of each aspect, is a configuration choice. The key is to embrace an algorithm that can consider enough of the things that matter to your employees (and customers), with enough subtlety, to minimise the “unintended consequences” of optimising for demand but inadvertently introducing features into schedules that employees will notice as a change and won’t like.
In reality most of our users achieve both at the same time - I’m not sure we’ve ever once seen a worsening of employee wellbeing as the source of the ‘gains’ for better meeting customer demand. We’ve never believed at Rotageek that deploying AI in scheduling inherently tips the balance in either direction - it simply gives you a better tool to optimise for either or both. 🤖✨