Ask a plant manager what she did on Tuesday. You will not hear about strategy. You will hear about the four systems that disagreed about the same order, and the spreadsheet she built at 6:40 p.m. to settle it.
That spreadsheet is the most expensive employee in the company. Nobody requisitioned it. Nobody can fire it. It has been on the payroll for years, funded by overtime and quiet attrition.
I have come to think of this as the work that shouldn't exist — not the people doing it, who are usually the most resourceful staff in the building, but the job itself. It exists only because two systems do not agree, and someone has to stand between them.
The job nobody requisitioned
Frends, an integration platform vendor, published a piece this year that names the mechanism better than most analyst writing. It calls the ongoing cost of systems that are not connected well enough for automation to work at scale the integration tax.
Its distinction is the useful part. Integration debt is the accumulation of shortcuts — the script written to meet a deadline, the point-to-point connection that solved today's problem. The integration tax is what you keep paying afterward. One is the cause; the other is the invoice that never arrives in a recognizable form.
Frends also coins the automation productivity gap: the distance between what leadership expects automation to deliver and what the architecture can actually support. When that gap is wide, partial automation does something worse than nothing — it creates new manual handoffs and leaves the organization paying the same tax while actively trying to reduce it.
I should be plain about the source. That page is vendor content, and it announced a benchmark — 600-plus IT decision-makers across six European markets, conducted by Sapio Research — whose figures were not available in what I could read. So the framing is useful and the numbers are still owed.
What the research does quantify
The most specific published figures come from Freshworks' Cost of Complexity report: a survey of 700 IT, CX, finance and operations professionals across six countries. Freshworks sells service software, so read it as a vendor finding. The internal consistency is what makes it worth a board slide.
| Measure | Figure | Status |
|---|---|---|
| Annual revenue lost to tech complexity | 7% average | Vendor survey, 700 respondents |
| Software budget spent on underused or poorly integrated tools | 20% | Same survey |
| Time lost per employee per week to context switching | About 7 hours | Same survey |
| Apps and channels juggled daily | About 15 apps, 4 channels | Same survey |
| Implementations that ran over budget in 12 months | 43% | Same survey |
| Respondents who did not receive expected software ROI | 53% | Same survey |
| Weekly hours lost chasing data trapped in silos | 12 hours | Agency analysis relaying third-party research |
| Revenue leakage as share of realized EBITA | 1% to 5% | EY, relayed secondhand |
The last two rows deserve a flag. The 12-hour figure appears in agency analysis citing DATAVERSITY research on data silos, and the EBITA range is attributed to EY through that same secondary chain. Neither underlying study was retrieved directly. Treat them as directional.
Freshworks' chief customer and marketing officer, Mika Yamamoto, described the human version of the loss in an interview with CX Today: employees lose almost a full workday each week to context switching. Her framing is the one I would keep — more tools do not equal more efficiency, and customers feel the impact the second an agent has to dig for answers.
Why the AI purchase comes before the architecture fix
Here is the pattern I find genuinely hard to defend, and I have watched it repeat.
A company has seven days of reconciliation pain. The architecture is the cause. But architecture work is invisible in a board deck: no demo, no logo, no analyst quadrant. An AI assistant, by contrast, demos in nine minutes and has a price per seat.
So the assistant gets funded. It reads the same inconsistent data. It produces a confident answer over four conflicting sources. And the reconciliation work not only survives — it now has a review step attached.
The Frends analysis describes the underlying problem with more patience than I can manage: AI agents need to read from and write to multiple systems, operate inside governance boundaries, and produce auditable outcomes. Where that infrastructure is fragmented, the agent hits a ceiling quickly.
An AI program cannot outrun the data model it inherits. That is not an argument against AI. It is an argument about sequencing.
The arithmetic of a job that shouldn't exist
Fastweigh, which sells scale-house software, published an example in September 2026 that is worth stealing because it is so plain: at 250,000 tickets a year, an average of one extra minute of handling per ticket exceeds 4,000 staff hours.
One minute. Not a broken process — a rounding error applied quarter of a million times.
Illustrative calculation, constructed to show the method. No real company or result is described. A distributor traces a single order-to-cash path and finds a coordinator rekeying 2,000 orders a month at four minutes each. That is roughly 133 hours monthly, about 1,600 annually. At a loaded $55 an hour, it is $88,000 of capacity — and most of it is released, not cashable.
That distinction matters more than the total. If you present $88,000 as savings, a CFO will ask which line in the P&L changes and the case collapses. If you present it as released capacity plus whatever actually leaves the cost base, the case survives contact with finance.
Three questions that separate work from rework
Who owns this field? For every field, one system is authoritative and the rest display it. If the warehouse holds the true on-hand count and the ERP shows it with a lag, any decision reading ERP inventory will be confidently wrong — and consistently wrong, which is harder to notice.
Who owns this exception, and how old is it? A queue without an owner and an age is a report, not a control. Exceptions belong where they occur, not in a month-end discovery.
What does this cost per completed transaction? Not per seat, not per token. Per finished unit of work. If that number rises as adoption grows, you have built a variable expense and called it a capability.
What I would put in writing
Four things, before a signature.
Unit economics, not licensing. A named owner, a measured baseline and a cost per transaction that either falls with volume or gets explained.
A reconstruction test. Ask whether a colleague could explain, from records alone, why a specific number changed. A dashboard showing that systems ran is reassurance. A queryable trail is evidence.
A named approver for anything consequential. A model that classifies a transaction and posts it has removed the human check from the one place controls exist to keep it. Staging a recommendation with its evidence costs almost nothing.
An exit path. A stable interface between business logic and whatever runs the model, so a provider change is configuration rather than a rewrite.
This is not a headcount story
I want to be careful here, because the wrong reading of this article is that the spreadsheet person is the problem.
She is not. She is the reason the order shipped. The problem is that a company with her judgment on staff spent it reconciling records instead of improving delivery, pricing or customer service.
Freshworks' own survey found 60% of employees considering leaving within the year, citing redundant tasks and unclear accountability — and nearly one in five reporting that someone on their team burned out or quit after a complicated software rollout. That is what the work that shouldn't exist actually costs. Not a salary line. The people who leave because the job was never the job they applied for.
The decision in front of the board
The work that shouldn't exist is not free, and it is not invisible to everyone. It is invisible to the budget, because it is paid in fragments: an hour here, a rekey there, a Saturday close, a customer who waited three days for an answer someone already had.
Fixing the seam is unglamorous work with a defensible return. Buying another assistant on top of the seam is a demo with a renewal date.
The question I would put to any leadership team is not whether AI can do more. It is whether the company can finally stop paying people to translate between systems it already owns.
Hashtags
#CustomSoftware #SystemsIntegration #EnterpriseAI #BusinessAutomation #ERPIntegration #DataGovernance #AIGovernance #OperationalExcellence #DigitalTransformation #ITLeadership #WorkflowAutomation #BusinessStrategy #CFO #CIO #RiskManagement #CostOptimization #EnterpriseArchitecture #BusinessIntelligence #ProcessAutomation #WorkforceProductivity