Autonomy Should Follow Reversibility, Not Ambition

Nearly every agentic AI roadmap we review has the same defect. It scopes agents by capability — what the model can plausibly do — and then bolts on controls once legal or risk gets nervous. The result is a portfolio of agents that are technically impressive and operationally unusable, because nobody can say what happens when one of them is wrong at 2am on a Saturday. Gartner's June 2025 prediction that over 40% of agentic AI projects will be cancelled by the end of 2027 — citing escalating costs, unclear value and inadequate risk controls — is not a verdict on the technology. It is a verdict on how scope is being chosen.

There is a better allocation rule, and it is boring enough to survive an audit committee: give an agent autonomy in proportion to how cheaply its actions can be reversed. Not how accurate the model is. Not how confident the vendor is. Reversibility.

Why confidence is the wrong gate

Confidence scores tell you about the model's internal state, not about your exposure. A 94%-confident agent that issues a duplicate supplier payment has created a recovery process involving treasury, the supplier relationship and possibly your auditors. A 71%-confident agent that drafts a credit memo for a human to approve has created a slightly annoying draft. Same model, same accuracy band, wildly different blast radius.

The MIT State of AI in Business work that circulated widely in 2025 — the source of the much-quoted finding that the overwhelming majority of GenAI pilots produced no measurable P&L impact — pointed at the same structural issue from the other side: value shows up when systems are integrated into real workflows and learn from them, not when they sit adjacent to work. Agents that can only draft never touch the workflow. Agents that can act but can't be unwound never get permission to. Reversibility is the axis that lets you thread both.

Build a reversibility ledger before you build agents

Before scoping any agent, inventory the actions it would take and score each one. This is a half-day workshop with process owners, not a research project.

• Undo cost. Can the action be reversed by a single system call, or does it require a human, a counterparty, or a journal entry? A price simulation is free to undo. A published price is not.
• Detection latency. How long before someone notices the error? Actions that surface immediately in a queue are safer than ones that silently corrupt master data for six weeks.
• Counterparty visibility. Did a customer, supplier or regulator see it? Anything that leaves your boundary — an email, an order confirmation, a filing — becomes practically irreversible even when it is technically reversible.
• Fan-out. Does the action trigger downstream automation? One agent-created delivery block can cascade through ATP, transport planning and invoicing before anyone reads the log.
• Concentration. Is the error idiosyncratic or systematic? A human makes one bad decision; a mis-prompted agent makes ten thousand identical ones before lunch.

Score those five, and your autonomy tiers assign themselves: full autonomy for cheap-to-undo, low-fan-out actions; propose-and-approve for anything crossing a boundary; hard human ownership for concentrated, slow-detection changes. You are no longer arguing about whether AI is trustworthy. You are arguing about specific actions, which is a conversation enterprises already know how to have.

Where the value actually concentrates

The counterintuitive part: reversible territory is large and underexploited. Reconciliation and matching, enrichment of incomplete records before they hit a system of record, triage and routing, simulation of pricing or configuration scenarios, drafting of structured artefacts that already pass through an approval step, monitoring and escalation. None of these make a keynote slide. All of them are high-volume, measurable, and safe to run at full autonomy on day one.

This is also where the unit economics work. Reversible actions need less review overhead, and review overhead — not tokens — is what kills the agentic business case. If every agent action requires ninety seconds of human verification, you have built an expensive way to create work. We see the same pattern in SAP landscapes: the agent that safely proposes a master-data correction for bulk approval outperforms the agent that dramatically closes a sales order, because one scales and the other needs a babysitter.

Make reversibility an engineering property

Reversibility is not fixed. You can manufacture it. Stage agent writes into a shadow table before promotion. Use compensating transactions rather than direct mutation. Give every agent action a correlation ID so a bad batch can be identified and reversed as a set rather than hunted individually. Add a delay window on outbound communication so a detected error never reaches the counterparty. Each of these converts an expensive category of action into a cheap one — which means engineering investment in undo capability directly expands the surface area where autonomy is affordable. That is a far better use of budget than another model evaluation. It is the kind of trade-off we work through in client architecture engagements, and it usually reorders the roadmap within a week.

The organisations that will still have agentic programmes in 2028 won't be the ones with the most capable models. They'll be the ones that made undo a first-class design concern, then let autonomy expand quietly behind it — one reversible action at a time.