We are handing agents the work humans used to do, and then wondering why the bill climbs while the results stay stubbornly incremental. The instinct is to blame the agent — its reasoning, its guardrails, its model. But look at what we are actually asking it to do. We are pointing it at a workflow whose every handoff was built for a person: a PDF to read, a dashboard to scan, a form to interpret. The work was shaped for human eyes, and we have asked a machine to read it.
That is not an agent problem. It is a design problem — and a much older one. Operating models were built for a purely human workforce, with work structured around functional roles, handoffs, and decision bottlenecks; simply inserting AI into that model only delivers incremental gains [1]. Adoption tells the same story: organisational AI use has reached 88%, yet actual agent deployment sits in the single digits — activity is everywhere, structural change is rare [3].
Built for eyes, not for machines
The artifacts we hand between steps are quietly the problem. Each was optimised for a human to consume — and each forces a machine to do expensive, error-prone reconstruction to get back to the data underneath.
Built for Eyes, Not for Machines
The Handoffs Shaped for a Human Reader
Tap an artifact to see why it costs a machine — and what to reshape it into
Select an artifact above to explore it
This is why bolting an agent onto the old flow disappoints. In finance, rule-based automation clears the standard invoice, but the value leaks on the case that only shows up across several documents — exactly the reading a human-shaped artifact makes hard [4]. Point an agent at the PDF and you have paid tokens to rebuild what a structured handoff would have simply stated.
The next-reader test
So before deciding whether a step should be agentic, ask a simpler question: who reads this handoff next, and does it need judgment? Two answers sort every handoff into one of three lanes — and, usefully, the leading operating-model research already describes the same three modes: some work stays human-led where trust and accountability dominate, some becomes AI-assisted, and some runs machine-led under human supervision [1]. Crucially, humans stay embedded wherever judgment, accountability, ethics, or regulatory oversight are genuinely required [2].
The Next-Reader Test
Where Does This Handoff Belong?
Answer for a real handoff in one of your flows
Who reads it next?
Does it need judgment?
Judgment, accountability, ethics, or relationship
Rows: needs judgment (top) / no judgment (bottom)
This handoff is
Machine-native
A human-as-router — the highest-value reshape
A person acts next, but only to read, transcribe, or route — no real judgment. This looks like a human touchpoint; it is really a human-shaped interface. Convert it to a machine-native handoff and free the person entirely.
Near-zero marginal token cost, and no parse errors to catch.
One cell does the most work. When a person reads the handoff next but the step needs no judgment — they are only reading, transcribing, or routing — that is not a human touchpoint at all. It is a human‑shaped interface wearing the costume of one. Those are the highest‑value steps to reshape: convert them to a machine‑native handoff and you free the person entirely, without touching a single decision that actually needs them.
The economics of re-reading
This is where the token argument turns counter‑intuitive. Inference keeps getting cheaper, yet compute and infrastructure spend is hitting record levels as agentic workloads scale [3]. Cheaper tokens don't save you if you point ever‑more agents at ever‑more human artifacts to re‑read. The structural saving comes from removing the re‑read — making the routing handoffs deterministic so no agent has to parse them at all.
With 8 handoffs, 60% shaped for a human reader and 55% of those pure routing, at $10 per million tokens, reshaping saves USD 4,224 per month. A USD 60,000 reshape pays back in about 14 months, so reshaping is the rational call.
Parsing cost, as-is
USD 7,680
per month
Saved by reshaping
−USD 4,224
per month
Payback
14 mo
on USD 60,000 of work
Reshape it. The one-off work repays itself inside 24 months, and every reshaped handoff becomes deterministic — the parse errors go too.
Reshaping removes 55% of the parsing cost — but the honest question is the payback, not the saving. Drag the token price toward zero and the crossover point moves; it never disappears, because what doesn't get cheaper is verification, accountability, and the cost of being wrong.
Illustrative model. Assumes ~8k tokens to parse and reason over one human-shaped artifact, and counts token cost only — it excludes the cost of catching a bad parse.
Notice what the model rewards — and where it refuses to. Reshaping wins on volume and durability, because the one-off engineering has to repay itself out of the monthly saving. Drop the token price far enough, or the volume low enough, and the model will tell you to leave the artifact alone and let the agent read it. That boundary is real and worth taking seriously.
What survives the boundary is everything the token price never touches: a deterministic handoff cannot misread, and nobody has to check it. Cheap tokens buy you the parse; they don't buy you the confidence that the parse was right.
Reshaping the flow
Reshaping is not a technology project; it is an operating‑model move, and it runs in a deliberate order — because wrapping AI around fragmented, human‑shaped workflows simply accelerates the fragmentation [2]. The value comes from redesigning so intelligence is embedded in the flow of work, not layered on top of it [1].
The Reshape Sequence
How to Reshape a Flow — In Order
Deterministic wins now compound into agentic gains later
Walk one real flow end to end and, at every handoff, name the next reader — a system or a person — and whether the step needs judgment.
Why it matters
You cannot reshape what you have not named. Most flows have never been audited for who actually consumes each artifact.
In practice
List each handoff with its artifact (PDF, dashboard, form) and apply the next-reader test. The "person, no judgment" rows are your quick wins.
The sequencing is the whole point. Making routing handoffs deterministic is the incremental value available now; it also lays the machine‑native rails on which agents compound later. Not everything needs to become agentic — and the flows where much becomes deterministic are precisely the ones where the remaining agents run cheapest and most reliably.
The best of both worlds
The debate is usually framed as a choice: automate with agents, or re‑engineer the process. The framing is false. The organisations pulling ahead do both, in order — and the evidence is unambiguous that the constraint is design, not tools. Only about 5% of organisations capture AI value at scale, and becoming AI‑first is roughly 30% technology and 70% people and operating‑model design [1]. Meanwhile the biggest measured productivity gains land in structured, machine‑legible work [3] — which is exactly what reshaping a human‑shaped handoff creates.
So use the transformation mandate for more than agents. Use it to ask, at every handoff, who reads this next — and to reshape the answer. Make the routing deterministic. Point agents at the ambiguity. Protect and feed the human decisions that genuinely need a person. That is not choosing between people and machines. It is giving each handoff to its right reader — and finally collecting the value that the old, human‑shaped flow left on the table.
Sources
- Boston Consulting Group. Design Your Company for AI, Not AI for Your Company. 2026.Two transformation case studies. Operating models were built for a purely human workforce — functional roles, handoffs, decision bottlenecks — so simply inserting AI delivers only incremental gains; value comes from a ground-up redesign with AI embedded in the flow of work, not layered on. Only ~5% of organisations capture meaningful value at scale; the shift is 30% technology, 70% people and organisation.View source
- EY-Parthenon. How agentic AI can help unlock enterprise value at scale. June 2026.As much as 70–75% of value creation potential is trapped in the silos between functions; AI wrapped around fragmented workflows simply accelerates fragmentation; humans remain embedded where judgment, accountability, ethics, or regulatory oversight are required.View source
- Stanford HAI. The 2026 AI Index Report — Economy. 2026.Compute costs and infrastructure spending are reaching record levels (Google reporting >$150B annual capex in 2025); organisational AI adoption rose to 88%, with generative AI used in at least one function at 70% of organisations, yet AI agent deployment remains in the single digits; productivity gains are largest in structured, measurable work.View source
- McKinsey & Company. AI in finance: How finance teams are putting AI to work today. 2025.Rule-based automation handles repetitive tasks (invoices, reconciliations); agentic AI is needed to orchestrate exception-laden workflows such as the close, and to catch value leakage that only appears across multiple documents.View source