Field note

The difference between using AI and operating AI-natively

May 2026.6 min read

Audience: Founders, COOs, and operating partners running AI adoption programmes.

At the start of the year, the working assumption was that licences would do most of the work. Buy enterprise AI, roll it out by function, and the productivity arrives. Three quarters in, with the same model versions and broadly the same licences, two patterns are visible inside the same companies. Some teams produce three times the output they did before. Other teams, holding the same tools, report no change. The variable is not the model. It is whether the operating model around the tool has been redesigned. This is the line between using AI and operating AI-natively.

The expectation going in was that the gap would close with training. It did not. McKinsey’s State of AI 2025 puts 88% of organisations using AI in at least one business function, with only 6% qualifying as high performers where AI contributes more than 5% of EBIT (McKinsey, 2025). The decisive variable, McKinsey reports, is workflow redesign, and its effect on EBIT impact is larger than any other dimension in the rewire (McKinsey, 2025). MIT’s Project NANDA found 95% of enterprise generative AI pilots delivered no measurable financial return across roughly 30 to 40 billion dollars in committed enterprise spend (MIT NANDA via Fortune, 2025). The cohort that did deliver was the cohort that bought from specialised vendors and built integrated operating partnerships, at roughly twice the success rate of internal builds. That is not a productivity story. That is an operating-model story.

What surprised most, working inside a small AI-native team rather than auditing one, was how early the redesign needs to start. Using AI means the workflow already exists and the tool is dropped into it. Operating AI-natively means the workflow does not exist yet and is designed around the tool. The two require different opening questions. The using-AI question is what task we can offload to the model. The operating-AI-natively question is what decision the model is reliable enough to make, what capital we will reallocate when it does, what we will stop measuring because the measurement was a proxy for the labour the model now performs, and what we will newly measure because the model exposes a variable the old workflow hid. The using-AI question produces incremental gain. The operating-AI-natively question produces a different company.

A simple test makes the line visible. Look at the calendar of the leadership team. If the calendar still treats the AI work as an additional project on top of the operating model, the company is using AI. If the calendar has rebuilt the cadence around an AI loop, with a weekly review of model performance, a monthly review of the workflows being collapsed, and a quarterly review of capital reallocation against the workflows that have been removed, the company is operating AI-natively. The first version of this loop runs imperfectly. The second version exposes what the third version needs to measure. The fourth version starts to compound. There is no shortcut, and there is no version of operating AI-natively that does not require senior time on the loop.

What changed as a result. The hiring profile rebuilt first. The next senior hire was no longer the head of a function the company already had. It was an operator who could design and maintain the AI loop inside an existing function, with the authority to retire the workflow the model now performed. The pricing decision rebuilt second. The unit a customer paid for stopped being the seat and started being the outcome the model produced, because the seat was now decoupled from the work being done. The product roadmap rebuilt third. The features the small team chose to ship were the ones that hardened the model loop, removed labour that the loop was already performing manually, and shortened the cycle time between a model decision and a customer outcome. The capital plan rebuilt last, because the previous three changes made it impossible to keep funding the old cost shape.

The hidden cost of skipping the operating redesign is that the AI line on the P&L lifts revenue and erodes margin at the same time. Bessemer’s State of AI 2025 put LLM-native gross margin at roughly 65% with 400% year-on-year growth (Bessemer, 2025). For traditional SaaS layering AI onto an 80-dollar seat, the same analysis shows roughly 15 dollars of variable cost added per seat, dropping gross margin from 80% to 65% on that seat (SaaS Magazine, 2026). The reading from a small team is the same as the reading from a public 10-K.

Using AI moves the cost line. Operating AI-natively moves the unit the company is paid for.

Companies that confuse the two end up worse off than companies that ignored AI entirely, because they have grown the cost line without rebuilding the revenue unit it has to be financed by.

The leadership implication is small, exact, and unusually concrete. Operating AI-natively is not a capability to acquire. It is a sequence of operating decisions to make, in the order that lets each unlock the next. Redesign the workflow before deploying the tool. Redesign the hiring profile after the workflow. Redesign the pricing unit before the next renewal cycle. Redesign the product roadmap to harden the loop. Redesign capital allocation to reflect the workflows that no longer exist. The mistake is not in any single decision. The mistake is treating the five as parallel projects. They are not. They are sequential, and the sequence is the operating skill.

The early pattern, stated as a rule of thumb for other operators. If you can name the workflow the model is now performing and the human work it has replaced, you are operating AI-natively. If you can only name the tool you have rolled out, you are using AI. The first pattern compounds. The second one does not. The next test, for the next quarter, is the cycle time between a model decision and a customer outcome. The teams whose cycle time falls inside the quarter are the teams whose operating model is now the AI loop.

Sources

AI-native, operating-model, workflow, margin, leadership

Back to writing