Beyond AI-First: Turning Artificial Intelligence into Business Performance

1. Questions every leader should be asking

Picture a company that has invested seriously in artificial intelligence over the last two years. And I do not mean only buying technology. It has rolled out or piloted generative AI assistants, built predictive models, tested agents, automated tasks in several departments and created dashboards to track all of it. Employees use the tools, some approved by the company and some brought in on their own. The technology or operations team delivers projects on schedule. The executive committee receives regular reports on adoption and progress. Everything seems fine.

Until someone, a board member skeptical of AI or the board itself, starts asking questions of a different kind. By how much did revenue per employee go up? By what percentage did gross margin improve? Did operating costs fall, or EBITDA rise? And, above all, can the CEO, operations or finance say which AI initiatives contributed to any of those results?

In my experience with boards and executive teams, the candid answer is usually “honestly, I'm not sure; we would have to measure it”. From there, some bet on improving the initiatives and others on cutting them. In both cases, the people who built or led them are not necessarily the ones who got it wrong: the mistake usually sits further back, in never having defined at the outset which result they were meant to move and how it would be measured.

And it is not just my impression. In its August 2026 global survey, McKinsey found that nearly 9 in 10 companies now use AI regularly in at least one function, yet only 37% attribute any impact on EBIT (operating profit) to AI, and just 6% meet the criteria McKinsey uses to classify a company as a high performer.² BCG estimated in 2025 that only 5% of companies were capturing substantial value.⁴ According to its new 2026 index, 7.5% capture value at scale and close to half obtain some value; the other half still obtains nothing significant.⁵ The gap between adopting and monetizing remains the rule.

From those same figures, BCG draws a more ambitious conclusion than the one I propose in this article: the companies it calls future-built grow revenue 2.4 times faster and EBITDA 2.8 times faster than laggards, and so it recommends adopting an AI-first operating model.⁵ The data is real, but it does not oblige anyone to follow that recipe, for three reasons. First, because those companies did not get there by chance: they already had better data, more management discipline and more capital before AI. Part of their advantage comes from what they already were, not from AI, and a survey cannot separate the two. Second, because what BCG measures in them — workflow redesign, results discipline, sustained investment — is exactly what this article calls capturing value. And that can be done without adopting the label or redesigning the whole company at once. Third, because the 5% of 2025 and the 7.5% of 2026 come from indices with different definitions: they show a trend, but the figures are not comparable with each other.

That is the central problem of enterprise AI today: organizations deploy ever more sophisticated technology without developing, with the same seriousness, the ability to convert it into measurable business results. In most cases of analytics and automation, the technology is not what fails. The exception is generative AI, which still invents answers in a confident tone and carries biases, and therefore demands human supervision and monitoring in the decisions that matter. But even there, what usually fails is what surrounds the technology: processes, rules about who decides what, and operating models that were never redesigned to take advantage of what AI can do.

This is why the conversation has to move beyond adoption. The goal is not to maximize the number of AI applications deployed. It is to decide where and how AI can improve the economic performance of the organization.

That distinction matters in the debate about "AI-first" companies, the ones that put AI at the center of their operating model. Not every company needs to become one of them. A mining company can reduce the variability of its plant processes, stop repeating operational errors and optimize maintenance and energy without rebuilding its operating model. The same holds for a manufacturer with its production and inventory. A professional services firm or a software company, on the other hand, may find that weaving AI agents into the way work itself is organized opens a much deeper transformation. Both paths can be economically justified.

AI-first is a strategic choice. Capturing value from AI is a business imperative.

The leadership job is to decide which level of integration creates the most value for your organization, and to make that value measurable.

2. Where the value is: indicators and mechanisms by industry

Productivity is the benefit most often attributed to AI, and the one most often measured badly. Counting hours saved says very little. For a leader, AI productivity only means something when it shows up in the operating and financial performance of the business. Four indicators cover most cases, and for each one it is worth going beyond the definition to the concrete mechanisms that push it up or down.

Revenue per employee. It rises when AI reaches the work that generates revenue. In a software or technology services company, a proof of concept for a client, which used to take weeks, now comes together in days. Development cycles (sprints) deliver more because part of the code, tests and documentation is generated with the help of AI. A new engineer shortens the learning curve because an agent hands over the context of the system in minutes. On the commercial side, salespeople manage larger portfolios and prepare more proposals. The most common warning sign is not that the indicator falls, but that it stays flat while the business grows: every additional dollar of revenue requires hiring in the same proportion. When that happens after two years of AI investment, AI is not reaching the work that generates the revenue.

Gross margin. It moves when AI lowers the cost of producing or delivering each unit. In mining, that means narrowing the variability of the concentrator plant, keeping metallurgical recovery close to its optimum and not repeating the same errors shift after shift. In manufacturing, less scrap and rework, and planning that reduces line changeovers. In retail, demand forecasts that avoid stock-outs and markdowns. In software, a lower delivery cost per project because the same team gets more work out the door. In capital-intensive businesses, small efficiency improvements have an outsized financial effect, because cost of goods sold is the largest line in the income statement.

EBITDA. AI reaches EBITDA through several routes at once: revenue growing faster than operating expenses, lower direct production or service costs, better asset utilization, less administrative spend and better decisions about profitability. What takes discipline is evaluating the net impact, deducting the additional cost of AI itself and not double-counting what is already in gross profit.

Operational productivity. The indicators depend on the industry and are worth naming precisely. In a mine: tonnes processed per effective hour, availability of critical equipment, deviation from plan. In a plant: units per labor hour, OEE, scrap rate. In a bank or an insurer: transactions or claims processed per employee and cost per transaction. In healthcare: the time clinicians spend documenting instead of with patients, how quickly the relevant information arrives when a treatment is being decided, and how clearly a diagnosis is explained to the patient and family. In all of those cases AI supports the clinical team; it does not replace it. In software: time from idea to prototype, defects per release, cycle time for a change to reach production.

The common thread is that AI productivity should be measured through business outcomes, not technology usage. Figure 1 traces that path, from the processes and decisions AI improves to the financial indicator that finally moves.

Figure 1. The value path runs from the processes and decisions AI improves to the financial indicator that moves; measuring AI at the left end of the chain says nothing about the right end.

Where to look, then? The most valuable opportunities are rarely the most technologically sophisticated. They tend to sit in three places.

Recurring processes. Financial reconciliations, production planning, order management, inventory replenishment, claims processing, and also the reports and presentations that eat hours of every week. These processes are attractive because improvements compound: a small reduction in unit cost, multiplied by a large volume, is a meaningful annual benefit. That said, accelerating one activity is not the same as improving the whole process: automating one step achieves little if other bottlenecks still set the pace.

Recurring decisions. These are the ones a company makes over and over: how much inventory to replenish, which orders to prioritize, which customers need commercial intervention, which equipment to maintain, which products are losing margin. AI can make them timelier, more consistent, better in quality and larger in scale. What is valuable is not more accurate predictions, but better decisions that turn into effective action.

Continuous performance management. Most companies have sophisticated dashboards for sales, profitability, production and inventory. They provide visibility, but visibility does not produce improvement by itself. A dashboard shows that margin is deteriorating; it does not say why, nor choose the corrective action, nor execute it, nor verify that it worked. Moving from monitoring to AI-supported performance management means combining the right capabilities — from business intelligence and predictive analytics to optimization and agents — and using each one only where it is already reliable.

Figure 2 arranges examples from these three categories by recurrence and economic impact. In the upper-right quadrant, a small improvement in each decision becomes a significant annual benefit.

Figure 2. Recurrence multiplies impact: the upper-right quadrant is where a modest per-decision improvement becomes a material annual benefit.

3. A case: taking gross margin from 30% to 31%

Consider a beverage company: soft drinks, juices and water, with two plants, its own distribution network and about 1,800 employees. It sells to supermarket chains, to tens of thousands of corner stores and liquor stores through its own trucks, and to restaurants and hotels. It generates USD 200 million in annual revenue at a 30% gross margin (gross profit of USD 60 million), operating expenses of USD 36 million and an EBITDA of USD 24 million, that is, a 12% EBITDA margin. The figures are hypothetical, but the profile is that of many mid-sized bottlers in the region.

Its margin is squeezed on three fronts. On the cost side, sugar, packaging resins and energy rise and fall with the markets. On the revenue side, the chains demand promotions and volume discounts that, if nobody tracks them, eat the margin of an entire category. And in the middle sits distribution: every truck that leaves half loaded, or visits a store that buys two cases, takes margin away from the whole route.

The company has an ERP, a CRM, production and routing systems and a business intelligence platform. Management reviews dashboards on sales by channel, plant costs, distribution cost, pricing and profitability every month. Despite that, gross margin has been slipping in some categories and the organization struggles to respond. The dashboards reveal the problem; the response gets stuck. Finance identifies the deviation; the commercial team reviews prices and promotions with the chains; production examines plant costs; distribution looks at its routes; procurement, at its sugar and resin contracts. Each on its own, with independent KPIs. Several meetings are needed to agree on an action and, by the time the company acts, more margin has eroded.

Management sets an objective (an OKR, for example): improve gross margin by one percentage point without reducing quality or compromising customer relationships. The one-point target is not arbitrary. A prior diagnosis showed that the company sits three points of margin below the top quartile of its sector, and that the first of those points can be recovered by acting on pricing and promotions, production planning, routing and procurement.

Notice what management does not do. It does not start by asking which AI model to implement. It asks: which recurring decisions determine our gross margin, and how can AI help us make them better? That reframing surfaces six opportunities: detect earlier that the juice category lost two points of margin to a chain promotion nobody offset; identify which format, which customer and which route explain the drop; estimate what happens if the list price of the family-size format goes up or if visits to the smallest-buying stores become less frequent; bring commercial, production, distribution and procurement together in a single decision; execute what is approved faster; and measure whether the actions improved profitability. Figure 3 contrasts the current state with the target state.

Figure 3. The target state does not remove people from the loop; it removes the waiting between detection, decision and action.

The business case. Assume an initial investment of USD 1.2 million in data integration, models and organizational change, plus USD 400,000 a year in operating cost, which is expensed and therefore reduces EBITDA. With revenue constant, one point of gross margin equals USD 2 million of additional gross profit. Net of the AI operating cost, that leaves USD 1.6 million of additional EBITDA per year, about 6.7% more than the USD 24 million starting point. If the point of margin were achieved immediately, the investment would pay back in roughly nine months; with a realistic six-month period to reach the full point, payback takes 12 to 15 months. And it pays to look first at what happens if the result is smaller than expected: at half a point of margin, incremental EBITDA falls to USD 0.6 million and payback stretches to two years; at just two tenths of a point, the AI operating cost absorbs the whole benefit. That exercise shows where the risk of the case sits: not in the technology, but in how much margin actually moves. Table 1 summarizes before and after.

Financial indicator	Before	After	Change
Revenue	USD 200.0M	USD 200.0M	—
Gross profit	USD 60.0M	USD 62.0M	+USD 2.0M
Gross margin	30.0%	31.0%	+1.0 pp
Operating expenses (excl. AI)	USD 36.0M	USD 36.0M	—
Incremental AI operating costs	—	USD 0.4M	+USD 0.4M
EBITDA	USD 24.0M	USD 25.6M	+USD 1.6M (+6.7%)
EBITDA margin	12.0%	12.8%	+0.8 pp

Table 1. Illustrative financial impact of a one-point gross margin improvement, net of AI operating costs. Not shown in the income statement: initial investment of USD 1.2M (payback ≈ 9 months at full run-rate, 12–15 with a six-month ramp). Sensitivity: at +0.5 pp, EBITDA +USD 0.6M and payback ≈ 24 months; at +0.2 pp, the AI operating cost absorbs the benefit.

These figures illustrate a possible outcome; they are not a forecast. Real results depend on which decisions get implemented, on how much of the improvement is really due to them and on the total cost of the solution. Some pricing actions may reduce demand. Some production improvements may require capital investment. Some margin gains may come from the market rather than from AI. A credible measurement compares observed performance with a baseline or a counterfactual scenario, that is, with what would have happened without the intervention.

The central point stands: what is being evaluated is not the AI system, but the improvement in economic performance that system makes possible.

4. What the frameworks say and where they converge

Three recent reports look at the same problem from different angles. Microsoft's 2026 Work Trend Index describes Frontier Firms as organizations where people and AI agents work as teams, and argues that what separates the ones pulling ahead from the ones stalling is not adopting AI (having it available) but absorbing it (making it part of how work itself gets done).¹ McKinsey found in 2025 that redesigning workflows is the factor that most increases the likelihood of reporting EBIT impact.³ In 2026 it added that nearly three-quarters of high performers have already done so.² BCG proposes designing the company for AI rather than AI for the company and, as we saw, goes further: it recommends being AI-first.⁶ What follows — the application of each framework to the margin case — is my interpretation, not those firms'.

Lens 1: human-agent teams (after Microsoft). In our beverage company, margin management is spread across four departments: commercial, production, distribution and procurement. A human-agent model would introduce specialized capabilities into that process: a monitoring agent that detects material margin deviations by category, customer, route and plant from validated data; a diagnostic agent that investigates drivers such as changes in the mix of formats sold, discounts to chains, sugar and resin costs, and under-loaded routes; a decision-support agent that prepares alternatives with their financial effect; and an execution agent that coordinates what is approved with the systems and the people responsible. None needs to be an autonomous decision-maker. Their authority should be graded by the risk of each action: a price or discount change for a supermarket chain requires commercial approval; an alert about trucks leaving below the load standard goes out on its own. Employees spend less time gathering information and more time evaluating decisions and managing exceptions.

Lens 2: workflow redesign (after McKinsey). Before thinking about agents, the company maps the whole flow: how deviations are detected, how information moves, who decides, who approves, how actions are executed and how results are measured. In most companies I have seen, that exercise shows that the problem is not the speed of financial reporting but the time between detecting a deviation, agreeing on an action and implementing it. The flow is redesigned in six stages — detect, diagnose, evaluate, decide, execute and measure — and each one gets an owner, a response time and an indicator. These six stages are the same links of the conversion chain I present in section 6, only seen from the process rather than from the value. AI enters where it improves each stage: predictive models to anticipate, optimization to evaluate alternatives, generative AI to explain findings, automation to coordinate. Success is measured by how long each decision takes and by the margin recovered, not by the number of models deployed.

Lens 3: organizational design (after BCG). Margin depends on decisions taken in four departments, but each department optimizes only its own part. Introducing AI into each one separately improves local productivity without solving coordination. The alternative is a cross-functional capability for managing margin: not a new department, but a shared process, with an executive who answers for the result, clear rules about who decides what, common data and analytics, and shared indicators. The organization is designed around the outcome, not around isolated applications.

Absorption on the ground. Suppose the company already has an assistant that analyzes margin reports and employees ask it which categories lost the most margin last month. That is adoption. The organization still depends on someone remembering to ask, interpreting, convening and executing by hand. Absorption, in Microsoft's sense, is that every morning deviations are detected on their own and causes are investigated without anyone asking; that owners receive prioritized recommendations, that actions are assigned and that results are compared with expectations. Dean Hardy, a consultant who writes from hands-on experience in his newsletter Past the Pilot, describes that same step: AI stops being a tool someone may use and becomes part of how the company runs.⁷

The three lenses converge on the same thing: map the flow, assign owners and decision rights, embed AI in the routine and measure the result. What none of them fully answers are two questions that take up most of my time with leadership teams. Why does an organization fail to implement these changes even when it has the technology? And how does management know whether the transformation is actually generating value? The first leads to what I call AI Leadership Debt (section 5); the second, to the conversion funnel (section 6).

5. Why value is not captured: AI Leadership Debt

In companies that fail to capture value, one pattern repeats: the technology arrives before the organization does. A company can have access to the best models on the market, experienced engineers, cloud platforms and sophisticated analytics while its leadership team lacks the knowledge or the capabilities to make good decisions about AI. I call this AI Leadership Debt: it accumulates when the potential of the technology advances faster than leadership's ability to understand, govern and convert that potential into value.⁸

It is recognized by concrete signals. An organization has it when several of these five hold:

1. There are AI initiatives approved without a financial indicator assigned.
2. Nobody answers for the end-to-end outcome and it is not defined who may act on an AI recommendation. The effect is that departments build competing solutions and managers resist, because accountability is unclear.
3. Leaders cannot tell what AI really knows how to do from what it merely appears to know, and swing between believing everything a generative assistant says and delegating the subject to IT.
4. Finance cannot distinguish the benefit that was actually obtained from the productivity that was only estimated.
5. Employees use AI tools nobody approved or supervises.

The more signals, the larger the debt. And like any debt, it charges interest: every initiative launched without an owner or a metric adds integrations, data and expectations that later cost more to unwind. Putting the house in order gets more expensive every month. Leadership debt is not the only reason value goes uncaptured: dirty data, integration costs, middle-management incentives and resistance from operating staff all weigh too. But leadership is the root cause of the others, because only leadership can assign owners, define decision rights and demand measurement. Microsoft's 2026 Work Trend Index points the same way: of the impact users attribute to AI, organizational factors explain twice as much as individual attitude (67% versus 32%), and only 26% of users say their leaders hold a clear and consistent position on AI.¹

The third signal deserves a pause. When leaders have fully delegated the use of AI, they tend toward one of two extremes. Some overestimate it: they take whatever a generative assistant produces as absolute truth, without asking where the figure came from, or conclude that the company will never need to hire another programmer. Others underestimate it and treat it as a technology matter that is not theirs to own. Both extremes produce poor investment decisions, and both are hard to correct from below.

Reducing the debt goes beyond teaching executives to write better prompts. It requires the leadership team to understand which business decisions AI can improve, where human judgment remains necessary, how workflows and responsibilities should change, and how to evaluate the return against competing investment alternatives. In the case of our beverage company, that means executives able to decide, on solid grounds, who is responsible for the process, how much autonomy the agents have, who has authority over prices and who answers for the financial result. Closing that gap is a leadership responsibility, not a technology project.

6. How to know whether value is captured: the conversion funnel

Even a well-designed flow does not create value by itself. The organization has to ensure that what AI produces — the intelligence — influences decisions, that decisions become actions and that actions produce measurable results. That conversion chain is the core of what I call AI Value Realization Theory (AVRT), a theory of how the value of AI is realized, presented here as a framework in development.⁸ It is also the thread that runs through this article: the six stages of section 4 and the seven steps of the next are two ways of walking the same chain.

AI → Intelligence → Decisions → Actions → Operational KPIs → Financial value

Value can leak at every link. A model can identify an opportunity and go no further. A manager can approve a recommendation that is never executed. And even an executed action may not improve profitability. Value is realized only when there is a measurable improvement attributable to the intervention.

To locate the leak, the framework proposes three conversion rates. They are calculated by cohort, that is, by the group of alerts born in the same month, not by calendar month: all the alerts of one month are followed until their effect is measured, even if that takes three months.

• Decision Conversion Rate (DCR): alerts that received a recorded decision (act, not act for a stated reason, or escalate), divided by the alerts that exceeded the relevance threshold set by finance. The leak is the alerts nobody looked at, not the ones dismissed with judgment.
• Action Conversion Rate (ACR): actions actually executed in the systems, divided by the decisions to act.
• Value Conversion Rate (VCR): actions with a positive effect relative to the baseline, divided by the actions executed. Actions with no effect and actions that were not measured are reported separately, because they are different leaks and are fixed differently.

An illustrative month in our beverage company: 40 material margin alerts by category, customer, route or plant; 25 with a recorded decision (DCR 63%; 15 were never reviewed); 18 actions executed (ACR 72%); and 12 with a positive effect (VCR 67%; 4 with no effect, 2 not measured). The twelve effective ones are things like a list-price adjustment on one format, a renegotiated discount with a chain, a change in the frequency of a route, a corrected parameter on a bottling line or a new resin supplier. The funnel says where value is lost: here, in the 15 alerts nobody looked at. And the figures square with the case in section 3: twelve effective actions a month, with an average effect of about USD 14,000 each, add up to the order of USD 2 million a year. These rates are still being validated and today serve to diagnose, not to compare across companies. In the meantime, it is worth complementing them with operational measures any management team can track from the first month: time between detecting and deciding, time between approving and executing, share of recommendations implemented, gross profit attributable to the actions and incremental cost of the solution.

More than creating dashboards, the goal is to establish traceability between AI capabilities, decisions, executed actions and economic results. Figure 4 brings together the chain, the funnel and the debt signals that interrupt it.

Figure 4. Value leaks at every link of the chain; the funnel locates the leak (here, 15 of 40 alerts never become a decision), and the leadership-debt signals explain it.

7. How to begin and how to sustain it

You do not need to redesign the whole business to begin. The more effective path is to start with one recurring process tied to an indicator that weighs on the result, and walk the conversion chain a first time in seven steps (Figure 5):

1. Choose an objective, financial or operational, that matters and set its baseline; do not start from a technology.
2. Identify the recurring decisions that determine it and prioritize them by weight in the outcome, frequency and feasibility.
3. Diagnose the current flow: how information becomes decisions today and where value is lost.
4. Design the target process with analytics, models, automation or agents, and with explicit responsibilities, approvals and escalation rules.
5. Implement a bounded pilot with operational and financial success criteria defined before deployment; a good pilot shows that AI-supported decisions change results, not just that the model works.
6. Measure economic value by distinguishing the benefit estimated at the start, what was actually implemented and the benefit actually obtained, net of costs.
7. Scale what works and reinvest.

Figure 5. Seven steps for the first cycle; the seventh is what turns a project into a capability.

Our beverage company starts with margin, but the same data integration, the same governance and the same operating practices later serve routing, maintenance of the bottling lines or demand planning. A fraction of the value captured is reinvested in those capabilities, and each new cycle starts from a higher point than the last. I call this Capability Compounding: the organizational ability to convert AI into results grows with every cycle of the chain, the way compound interest grows on accumulated capital.⁸ The idea is not new, and it is worth saying so: the strategy literature calls it dynamic capabilities (a company's ability to reconfigure itself) and absorptive capacity (its ability to take in outside knowledge), and here it is applied to the concrete cycle of AI. The effect is not automatic: reinvestment competes with other priorities and every additional investment has to be justified by what is expected of it. But it is what separates companies that run isolated AI projects from those that build a sustained capability. The competitive advantage will not come from owning a particular model or agent, but from that capability.

Conclusion

The AI-first debate is valuable because it forces companies to rethink how they operate. Microsoft, McKinsey and BCG contribute useful lenses, and none requires every organization to follow the same path. Some will benefit from targeted applications; others will redesign specific workflows; others will integrate AI progressively into their operating model; some will choose to become genuinely AI-first. The right strategy depends on the business, its opportunities, its constraints and its economic objectives.

What should not vary is the discipline of connecting every AI investment to a measurable outcome: more productivity, more revenue per employee, stronger margins, a higher EBITDA and better-quality decisions. That takes more than technology. It takes leadership capable of understanding AI, redesigning operations, governing decisions and measuring results.

The most important question is not whether your organization is becoming AI-first, but whether it is getting better, cycle after cycle, at converting AI into business performance.

References and further reading

1. Microsoft (2026). 2026 Work Trend Index Annual Report: Agents, Human Agency, and the Opportunity for Every Organization. Microsoft WorkLab, May 2026. microsoft.com/worklab
2. McKinsey & Company (2026). The State of AI in 2026: On the Road to ROI. Global survey, August 2026. mckinsey.com
3. McKinsey & Company (2025). The State of AI: How Organizations Are Rewiring to Capture Value. March 2025. mckinsey.com
4. Boston Consulting Group (2025). The Widening AI Value Gap. October 2025. bcg.com
5. Boston Consulting Group (2026). Applied AI Index 2026: AI Is Starting to Pay Off. 30 September 2026. bcg.com
6. Boston Consulting Group (2026). Design Your Company for AI, Not AI for Your Company. bcg.com
7. Hardy, D. (2026). The AI Frontier Firm Field Guide: Start Here. Past the Pilot newsletter, LinkedIn. linkedin.com
8. Barrientos, C. (2026). AI Leadership Debt, AI Value Realization Theory (AVRT) and Capability Compounding. Work in progress by the author. AVRT executive brief · AI Leadership Debt

Disclosure: the author is co-founder and CTO of Oncoger.ai, an AI company focused on oncology. The healthcare examples in this article are illustrative and do not describe that company's products.