A diligence framework
Five conservation laws that predict where automation savings actually land in labor-based businesses, and the fallacy that inflates every ceiling.
The frame
Every task stamped in seconds. Every queue observable. Every handoff logged. It is the most densely instrumented labor system that exists, which is why the constraints are visible there first. The same laws govern every services business built on human throughput.
Arrival variance, occupancy, abandonment.
Volume relocates; it does not vanish.
Averages hide the mix that drives cost.
Who is permitted to decide, and sign.
Most operators just don't have a stopwatch on the work, so they discover the constraint when the margin bridge fails in year two.
The framework
Immutable constraints on labor-based throughput. They hold regardless of vendor, model, or vertical.
Automation takes the cheapest work first.
Deployment targets the most structured, highest-volume, lowest-variance work, which is by definition the cheapest in the portfolio. The residual mix is harder, slower, and more expensive per unit. Cost-out is capped by the cost of the tier removed, not the share of tasks removed.
Work relocates; it is not destroyed.
Displaced volume reappears as exception handling, rework, oversight, or across the boundary onto the client's own staff. A savings case that shows no offsetting line item has not been modeled. It has been asserted.
Some decisions require a licensed human.
A CPA signs the return. A clinician signs the note. A licensed producer binds the policy. This ceiling is regulatory and liability-driven, and it does not move when the model improves. It is the one constraint a better model cannot lift, and it is systematically underpriced in exactly the regulated verticals sponsors are consolidating.
Aggregated variance is the real efficiency.
In most roll-ups the value creation is Erlang, not AI. Aggregating variable demand across a larger pool is a genuine, durable gain, available today with no model risk. The AI thesis requires deflection rates to hit plan, containment to hold, and the vendor to keep shipping. The pooling thesis requires a routing change. But the gain only exists if the work is genuinely fungible: separate tooling, client contracts forbidding shared agents, or correlated seasonal peaks and you don't have a pool of 100, you have four pools of 25.
Every handoff is a queue with its own loss.
Each human-to-AI seam carries its own latency, error rate, and abandonment. Automation does not remove seams, it adds them. And because AI scales concurrency without limit while humans do not, handoff arrivals are burstier and more correlated than the organic traffic they replaced. Staffing follows peak simultaneous handoffs at SLA, never average volume.
Skimming and Conservation break the synergy model. Jurisdiction sets the ceiling. Pooling is where the durable value sits. Friction is what nobody staffs for.
The named fallacy
Half my volume is refunds, refunds automate, therefore I remove half my labor cost. Three unstated assumptions, each of which fails.
The claim and the outcome differ by a factor of three, before a single deployment constraint is applied. Illustrative cascade; eligibility, containment, and cost-weight ratios are measurable in any instrumented operation. Ask for all three.
Applying the laws
1,000 units at a $50 blended fully-loaded cost. The sponsor case assumes 40% of tasks are automatable, and that deployment reaches all of it.
Roughly a third of the modeled synergy. And this bridge still grants the sponsor their 40% starting point. Illustrative model; tier costs, exception rates, and review burden are vertical-specific and should be sourced in diligence.
The empirical gap
Roll-up theses routinely underwrite automation rates that deployments do not reach. That is not a rounding error. It is roughly half the thesis.
You reach less of the tier than modeled.
And the tier you do reach is cheaper than average. The ceiling gap and the margin bridge compound.
Two independent haircuts, and sponsors apply neither. Bands reflect commonly underwritten ranges and commonly observed plateaus; verify against target-specific deployment data.
Application
The one-line thesis
Underwrite the physics first. The technology assumption is the easy part to fix later. The structural one is not.
Who wrote this
I run operations for a CX company delivering AI-augmented service at scale, with P&L responsibility across more than 200 concurrent client programs.
That is an unusual vantage point, and not because contact centers matter more than other services businesses. It is because they are measured. Every task stamped in seconds, every queue observable, every handoff logged. When an automation case fails there, it fails in high resolution and on a short clock. An accounting firm or a revenue cycle shop hits the same wall, but finds out two years later in a re-forecast, without the instrumentation to explain why.
So the five laws are not a theory I brought to operations. They are what I watched happen across a few hundred real programs before I had names for them, in the one place where you can see the savings move rather than infer it afterward.
Diligence support for sponsors underwriting services roll-ups, operating reviews for management teams, and post-close instrumentation. If you are looking at a target where the automation case is carrying more weight than the model can support, I would like to hear about it.
Josh Magsam
Founder, Bardwise LLC
josh@cxphysics.com