The Digital Workforce That Never Arrived
We paved the cowpath, laid it in tarmac, and painted lines on it, and in doing so guaranteed that no one would ever build the road.
Executive Summary
By the early autumn of 2021, almost every large organisation had three things: a robotic process automation programme, a slide that counted how many bots were “live,” and a quiet, growing suspicion that the number on the slide and the number in the ledger had stopped agreeing with one another. The promise had been a digital workforce — tireless software robots taking the drudgery off human hands and the cost off the balance sheet. The reality, two or three years in, was more ambiguous: pockets of genuine relief, a surprising amount of maintenance, and savings that had a habit of shrinking every time someone recalculated them honestly.
This essay is a reflection on that gap. It is not an argument that the automation of clerical work was a mistake — in bounded places it worked, and worked well. It is an argument that the framing was the mistake: that a tactical layer laid over ageing systems was sold as a strategic destination, that motion was mistaken for progress, and that automating a process is not the same as fixing it. The pattern persisted not because practitioners were naive but because a set of structural forces — vendor incentives, executive appetite for a clean cost story, vanity metrics, and the sheer seductive ease of automating the surface rather than the substance — all pulled in the same direction. The automation wave was, in the end, less a story about robots than a mirror held up to how organisations actually pursue change.
The Slide and the Ledger
The scene recurs with such regularity that it has become a genre. A steering committee, a darkened room — or by 2021 more often a gallery of faces on a video call — and a slide with a large number on it. Two hundred and forty bots in production. The equivalent of three hundred and fifty full-time roles. A digital workforce, scaled. There is applause, or its remote-working equivalent, a scattering of approving nods and a chat window filling with thumbs-up.
Eighteen months later the same programme is harder to talk about. The bots are still there, mostly. But so is a team of people whose entire job is keeping them running, and that team has been growing. The three-hundred-and-fifty-role figure has not appeared in a business case for some time; when finance goes looking for it in the actual cost base, it proves elusive. Nobody is quite lying. The bots really did automate the tasks. And yet the money that was promised has developed the quality of a horizon — always visible, never arrived at.
What we are fluent in, as a profession, is the launch. We are far less fluent in the second act, the long unglamorous stretch where a technology either compounds into real value or slowly consumes the savings it was supposed to create. The automation wave of the preceding years is, above all, a study in that second act.
What We Were Promised
It is worth being fair to the promise, because it was genuinely attractive and, in 2021, genuinely timely. The pitch for robotic automation had a rare elegance: you did not need to replace your systems, rewire your architecture, or wait three years for a platform programme to deliver. You pointed a piece of software at the screens your people already used, recorded the clicks and keystrokes of the dull, repetitive, rules-based work — the swivel-chair shuffling of data from one window to another — and you handed that work to a bot. Fast to build. Cheap relative to a systems replacement. Reversible if it went wrong. Change without the pain of changing anything underneath.
The moment amplified all of it. The pandemic had delivered an operational shock that made the case almost rhetorical: volumes spiking in some channels and collapsing in others, back-office teams suddenly dispersed to kitchen tables, processes that had quietly depended on someone physically walking a form to another desk now visibly broken. Cost pressure arrived in the same breath. And by the middle of 2021 a tightening labour market was making it harder and more expensive to hold on to precisely the transactional staff whose work was most automatable. A tireless digital worker that did not need a laptop shipped to its home, did not resign for a better offer, and did not mind reconciling ten thousand statements a night was not a hard thing to want.
Around the core idea an entire vocabulary had bloomed. Attended and unattended bots. Orchestrators. Centres of Excellence. Citizen developers who would build their own automations. And, freshly minted, the umbrella term for where it was all supposedly heading — hyperautomation, the notion that everything that could be automated eventually would be, stitched together with process mining, machine learning, and intelligent document processing. The direction of travel felt settled. What remained, the story went, was simply to scale.
The Second Workforce
Scaling was where the trouble lived. The first uncomfortable discovery was brittleness. A bot built by recording a human’s interaction with a screen is, by construction, coupled to that screen. It does not understand the invoice; it knows that the total sits at a particular position, that a particular button advances the case, that a particular field expects a particular format. Change the presentation layer — and presentation layers change constantly, through a vendor’s portal redesign, an operating-system update, a security patch that repaints a login page — and the bot does not adapt. It fails, often silently, sometimes by confidently doing the wrong thing at machine speed.
A digital workforce built to remove human effort quietly created a new category of human effort: the second workforce, whose full-time job is keeping the first one alive. The maintenance tax is not a defect of a particular implementation. It is the structural consequence of automating the surface of a process rather than its substance.
Consider a composite that will be familiar to anyone who lived through it. A finance shared-service operation automates its purchase-to-pay drudgery: some sixty bots handling invoice matching, statement reconciliation, and master-data updates, with a business case promising the release of forty-five full-time roles’ worth of capacity. Within nine months two ERP patches and a supplier-portal redesign have broken roughly a third of the estate. The Centre of Excellence, launched with four people, has grown to fourteen to keep pace with the breakages and the backlog of change requests. When someone finally reconciles the promise against the outturn, the genuinely released capacity is closer to twelve roles than forty-five — and most of that twelve has been redeployed to other work rather than removed from the cost base. The programme is not a failure exactly. It is simply nothing like the thing that was sold.
This is the arithmetic that undid so many business cases: they were built on gross task-time removed and quietly assumed it would fall through to net cost taken out. Between the two sat the maintenance headcount, the licensing, the orchestration infrastructure, and the awkward fact that automating four-fifths of a role rarely lets you remove a whole person. Capacity was released. Cash frequently was not.
Automating the Cowpath
The deeper failure was subtler than brittleness, and more damaging, because it looked like success right up until it didn’t. Robotic automation is extraordinarily good at doing a process faster. It is entirely indifferent to whether the process should exist. And so, again and again, organisations pointed their new tools at exactly the tasks that most deserved to be eliminated rather than accelerated — the reconciliation that existed only because two systems had never been integrated, the manual re-keying that compensated for a data-quality problem nobody had fixed, the four-way approval that governed a decision of trivial value.
Automating that work does something perverse. It takes a process that was visibly painful — and therefore a candidate for redesign — and makes it quietly bearable. The pain that would have eventually forced someone to fix the underlying cause is anaesthetised. The broken integration is now permanent, because a bot bridges it every night. We paved the cowpath, laid it in tarmac, and painted lines on it, and in doing so guaranteed that no one would ever build the road.
“The bot did not fix the broken process. It preserved it, at speed, and removed the last incentive anyone had to ask why the process existed at all.”
There is a genuine tension here that the honest practitioner has to hold. Sometimes the pragmatic bridge is the right call; not every broken seam can or should be re-engineered, and an automation that buys three years of relief while a platform is rebuilt is a perfectly defensible thing. The failure was not bridging. The failure was forgetting that a bridge is a bridge — treating the scaffolding as the building, and letting the temporary fix calcify into permanent infrastructure that no one owned, no one had budgeted to maintain, and no one dared switch off.
Why the Gap Persisted
If the shortcomings were this visible, the interesting question is why the pattern held across so many organisations for so long. The answer is not individual foolishness. It is that several structural forces were all pushing in the same direction, and none of them was pushing towards honesty about net value.
- Vanity metrics made activity legible and value invisible. A bot count is easy to put on a slide; realised net savings, verified by finance and net of maintenance, are hard, slow, and often disappointing. A programme measured by bots in production will produce bots in production, whether or not they are worth running.
- Vendor and adviser incentives rewarded expansion, not restraint. Licences were often priced per bot or per process. The commercial logic of the ecosystem favoured more — more automations, more of the estate covered — and rarely the sober counsel that a given process should be redesigned or retired rather than automated at all.
- Executives wanted a clean cost story, and the technology offered one. A headline about a digital workforce and a large full-time-equivalent number is a far more comfortable narrative than one about redesigning how work is done, which implicates leadership, structure, and years of effort. The bot narrative let organisations talk about transformation while changing very little.
- The Centre of Excellence carried a contradiction. It was asked simultaneously to evangelise automation and to govern it — to grow the pipeline and to police whether a candidate process deserved automating. In almost every case the growth mandate won, because that was what was celebrated.
- The path of least resistance ran through the surface. Automating a screen is easy and fast; fixing the process, the data, or the system beneath it is hard and slow. Under pressure, organisations reliably chose the easy layer, and then defended the choice.
Put together, these forces did not merely permit the gap between promise and reality. They actively manufactured and sustained it. The wonder is not that so many programmes drifted; it is that any of them held their discipline at all.
The Strongest Case for the Bots
It would be too easy, and unfair, to end there. The most serious defence of robotic automation deserves to be stated at full strength, because it contains a great deal of truth.
The defence runs like this. Robotic automation was never meant to be a strategy; it was a tactic, and it should be judged as one. Legacy estates in banking, insurance, and the public sector were riddled with systems that had no usable interfaces, no appetite for modification, and no realistic replacement date. Against that reality, a piece of software that could operate those systems the way a person did — reversibly, cheaply, without a multi-year integration programme — was not a delusion. It was often the only thing that could deliver relief inside a planning cycle. Judged as a bridge over genuinely un-crossable ground, it worked, and the organisations that used it as such got real, bankable value.
This is correct, and it is precisely why the technology is worth taking seriously rather than dismissing. But it also locates the failure exactly. Where automation was scoped as a deliberate, temporary bridge — with a named owner, a maintenance budget honestly carried, a genuine business case built on net value, and a hard rule that a badly broken process was fixed rather than automated — it tended to deliver what it promised. Where it was scoped as a destination, dressed in the language of a digital workforce and measured by how many bots were live, it tended to disappoint in the ways this essay has described. The concept was sound. The framing decided the outcome. The same tool, pointed at the same estate, produced value or theatre depending almost entirely on the discipline of the hand that held it.
What the Automation Wave Was Really Telling Us
Step back far enough and the robotic automation story stops being about robots. It is a particularly clear specimen of a much older organisational habit: reaching for a technology as a substitute for the harder, slower, more political work of changing how an organisation actually operates. The appeal of the bot was never really its cleverness. It was that it promised transformation while asking almost nothing of the structures, incentives, and processes that most needed to change. It let leaders point at motion and call it progress.
| The promise | The reality |
|---|---|
| A digital workforce that replaces effort | A first workforce of bots and a second workforce that maintains them |
| Cost taken out of the base | Capacity released, frequently redeployed rather than removed |
| Scale as a matter of adding bots | Scale as the point where brittleness and maintenance compound |
| Change without changing the systems beneath | The systems beneath, now permanently unfixed |
The most useful thing automation did, in the end, was diagnostic. A bot cannot be built for a process until that process is understood step by step — and the act of trying to automate work forced organisations to look, sometimes for the first time, at what their people were actually doing all day. What that gaze revealed was usually not a case for automation. It was a case for elimination, integration, and redesign. The tragedy of the wave is how rarely that finding was acted upon; the tooling that surfaced the broken process was then used to preserve it.
As to where this is heading, the honest answer from here is that it is uncertain. The newer additions to the toolkit — process mining that maps how work truly flows, machine learning that lets automation cope with less structured inputs, the low-code platforms promising that the business will build its own automations — are real advances, and they loosen some of the specific constraints that made the first generation so brittle. Whether they escape the underlying trap is a different question, and not one that can be answered from 2021. A smarter bot pointed at a process that should not exist is still automating something that should not exist. If the lesson of the first wave is not learned, better tools will simply let organisations make the same mistake faster and at greater cost.
The Discipline That Separated Success from Theatre
If there is a transferable principle in all this, it is unglamorous and it is this: automate a process only after you have earned the right to, by first asking whether it should exist. The organisations that got real value from the automation wave were not the ones with the most bots or the cleverest tooling. They were the ones that treated an automation candidate as a question rather than an opportunity — that ran every process through a simple sequence before a line of it was automated.
- Should this work exist at all? If it is compensating for a broken integration or a data-quality failure, the first-best answer is to fix the cause, and automation is at best a dated bridge with an owner and an expiry.
- If it must exist for now, what is the honest net case? Not gross hours removed, but cash or capacity taken out after maintenance, licensing, and the stubborn last fraction of a role are subtracted.
- Who owns it when the screen changes? Every automation is a standing liability the moment it goes live; an automation without a named maintainer and a carried budget is a future outage waiting for a patch.
- How will we know it worked? Measured in outcomes verified by finance, never in bots counted by the programme that built them.
None of this is exotic. It is simply the difference between using a tool and being seduced by one. The automation wave was not a failure of technology; the bots did, for the most part, exactly what they were built to do. It was a failure of framing — of mistaking the surface for the substance, activity for value, and a bridge for a destination. That failure is not specific to robots, which is why it is worth understanding well. The next technology to arrive wearing the language of effortless transformation will make the same offer, and the organisations that remember what the digital workforce actually taught them will be the ones best placed to tell the promise from the reality.