Enterprise AI ROI: Build the Roads Before Buying a Faster Car

Why individual gains can disappear

We hear about people becoming dramatically more productive with AI while their companies struggle to demonstrate meaningful returns. I do not see that as a contradiction.

Give someone a powerful tool while leaving the organisation around them unchanged, and their improvement may remain local. They complete one task faster, then wait for the same approval, handoff or downstream bottleneck.

The organisation has created more output, but it may not have improved the outcome.

A Rolls-Royce in a jungle

Imagine a business that has spent years travelling through a forest on horses. Its paths and routines have evolved around that method of transport.

Then someone delivers a Rolls-Royce. It is faster on a road, but there are no roads. It gets caught between trees and blocked by terrain it was never designed to cross.

That is what happens when we bolt AI onto a workflow built around different constraints. The capability is real; the surrounding system prevents it from being useful.

For agents, the roads are accessible information, redesigned processes, clear authority and reliable handoffs. Building them requires decisions about the whole journey.

Start with the outcome and map backwards

Take a marketing process from briefing to approval. Faster drafting helps only if the brief is usable, reviewers have the necessary information and decisions can be made without repeated rework.

Before adding an agent, map the existing process. Where does work wait? Which decisions require judgement? What information is missing? Which checks are repeated because earlier stages cannot be trusted?

Then reimagine the process around what the technology can reliably do. Some tasks may disappear. Others may move earlier. A human may need to approve an exception rather than review every routine item.

This is a design exercise for the people who understand the work, supported by people who understand the technology.

Develop three capabilities together

I think marketing teams need a combination of subject-matter expertise, imagination and agentic AI knowledge.

Subject-matter expertise explains how the work actually happens. Imagination helps people see a different route to the outcome. Technical understanding helps them distinguish useful possibilities from brittle assumptions.

No single discipline owns all three. The opportunity is to build teams that can learn across those boundaries.

Measure the whole process

Time saved in drafting is useful, but it is not the same as money saved or a better campaign.

Establish a baseline before deployment. Then track the measures that matter for the complete process: elapsed time, quality, review effort, rework, cost and the business result.

If one team produces more material but another spends longer correcting it, that additional work belongs in the calculation. If saved time is redirected to a higher-value activity, make that decision visible too.

Evaluation is ongoing work

A successful demonstration does not finish the job. Inputs change, exceptions appear and the system's behaviour may shift.

An evaluation process should draw on the judgement of the people who use the outputs. When someone corrects an agent, that correction can become a future test case, even if the underlying model does not learn from the interaction itself.

Anthropic's work on agent evaluations is useful background for teams developing this discipline.

Buying the faster car is the easy part. Enterprise value comes from mapping the terrain, building the roads and improving the journey as you learn.