Agentic implementation can improve margins. But the mechanism is not “more AI”. It is less avoidable work per completed outcome, better decisions at the point of execution, and a cost base that grows more slowly than revenue.
For an operating studio and fund, that distinction matters. A convincing demonstration does not establish a better business. The investment case begins when a workflow changes and the economics can be measured.
From assistance to completed work
A copilot helps a person perform a task. An agentic workflow can pursue a defined outcome across several steps: retrieve an order, check the applicable policy, propose a resolution, execute an authorised action and record what happened. The useful distinction is not the label. It is how much work is actually completed, with what authority and at what cost.
Consider an e-commerce operation. An order exception may require a support colleague to inspect payment, stock, shipment tracking and customer history before contacting a warehouse. Connecting those steps can remove handoffs and waiting time. But if every decision still needs extensive review, the apparent automation may simply relocate the work.
We would start with a narrow process whose completion criteria are unambiguous. “Produce a response” is not enough. “Resolve an eligible delivery exception without a repeat contact, within policy” is closer to an economic unit.
Three routes to better margins
Lower delivery cost. Routine work can require fewer paid hours or less outsourced processing. The cash benefit appears only when expenditure is actually avoided or reduced. Time released inside an unchanged payroll is capacity, not an immediate saving.
More contribution from the same operation. A team may process more orders, quotes or qualified enquiries without equivalent growth in staffing. That can create operating leverage. It is not unlimited: acquisition costs, fulfilment capacity and service quality still constrain the result.
Less leakage. Faster exception handling, more accurate product information and better follow-through can reduce avoidable refunds, errors or missed opportunities. These are hypotheses to test, not benefits to assume. Measure contribution after discounts, returns, payment fees and fulfilment—not just sales.
Count the whole system
The relevant cost is not the price of a model call. It includes integration, maintenance, monitoring, human review, tool access, failed attempts, escalations and the consequences of mistakes. A cheap agent that creates expensive rework is expensive.
A practical comparison is:
Net operating benefit = cash costs avoided + incremental contribution − implementation and running costs − additional rework and losses.
This is a decision framework, not an accounting standard. Whether a cost affects gross margin or operating margin depends on the business and its accounting treatment. Do not treat every productivity gain as an improvement to both.
Compare the same workload over the same period. Track cost per correctly completed case, repeat-contact rate, escalation rate, turnaround time and customer outcomes. Include the ongoing cost of the people supervising the system. Where feasible, compare similar cases handled by the old and new processes rather than crediting the agent for every improvement that happened during rollout.
Authority is an economic choice
Autonomy should expand only when the evidence supports it. Reading inventory is different from changing a price. Drafting a refund recommendation is different from issuing money. Approving a supplier payment is different again.
Begin with explicit permissions, monetary limits, a human owner for exceptions and an audit trail. Keep irreversible or high-impact decisions behind approval. The aim is not maximum autonomy. It is the highest useful level of autonomy at an acceptable total cost and risk.
Underwrite the business, then prove the upside
For an acquisition, the existing economics should make sense before agentic improvements. Automation potential is upside to validate, not a reason to pay today for savings that do not yet exist.
For a new product, the same discipline applies: does the agent complete something a customer values, reliably enough to support the price and cost structure?
The strongest implementations look less like a spectacular chatbot and more like a quieter operation: fewer handoffs, fewer exceptions and more completed work. Agents do the work. The margin comes from the operating design.
This is Rebellion Capital's operating perspective, not a claim of achieved results in our portfolio. No financial uplift is promised.