88% of companies are using AI in some form. Only 39% attribute any EBIT impact to it, and of those, most put the figure below 5%. That's from McKinsey's State of AI 2025, and it captures the real state of the market: AI deployed almost everywhere, value captured almost nowhere.
The technology is ready.
Evidence that AI moves margin and volume is published, audited and replicated. What fails is how it gets deployed. The companies that see EBIT are the ones combining AI with teams that redesign their work around the tool. Those that treat AI as a mechanical replacement for people end up with nice-looking pilots that never touch the P&L.
What follows are ten proven levers, each with real numbers published by the companies themselves or by peer-reviewed studies, and the common pattern linking them: AI returns money when it amplifies people; when it's used to replace them, pilots go nowhere.
What the aggregate data actually says
McKinsey measures impact at the function level, not just at the enterprise level. And there the numbers look very different: 10–20% cost reductions in software and manufacturing; revenue uplifts above 10% in marketing, R&D and sales. The money is real. What fractures the enterprise-level result is the company's ability to translate those functional uplifts into consolidated EBIT, which demands the hard work: workflow redesign, data governance, time reallocation.
Another number worth pinning down: 80% of companies set efficiency as the objective of their AI initiatives, while the ones capturing the most value also add growth and innovation as goals. Translation: use AI only to cut costs and at best you get the savings. Use it also to sell more and ship faster, and you unlock the double effect.
Front office: support, sales and marketing that convert more
1. AI-augmented customer service with humans on the hard tail
The most-cited case is Klarna. In its first month the AI assistant handled 2.3 million conversations, two-thirds of all service chats, with CSAT on par with human agents and a 25% drop in repeat enquiries. The company attributed $39M in P&L improvement in 2024, updated the figure to $60M on its Q3 2025 earnings call, and saw cost per transaction fall from $0.32 to $0.19 in two years.
The more interesting data point came later: in 2025 Klarna reintroduced human agents for complex cases after finding that full automation degraded the experience on the high-value tail. The operating headline still holds (lower cost per transaction, higher volume handled), and the lesson is new: optimal economics means AI on first contact, well-qualified humans on the hard tail.
Academic evidence points the same way. The Brynjolfsson, Li and Raymond study (Stanford HAI and MIT, published in the QJE in 2024 on 5,179 contact-centre agents) found an average 14% productivity gain with generative AI assistance, with a sharply asymmetric distribution: +34% for novice and low-performing agents, close to 0% for veterans. AI transmits the tacit knowledge of the best to everyone else, without lifting the ceiling of the best themselves. For a COO the read is direct: AI flattens the learning curve and stabilises service quality; it doesn't replace the agent who already knew how to do the job.
2. Marketing content at half the cost, twice the speed
Unilever has spent two years rebuilding its content production chain with generative AI and digital twins on NVIDIA Omniverse. The result, on the record from its Chief Growth and Marketing Officer: product imagery produced twice as fast, at 50% of the cost. In one published case, TRESemmé Thailand cut content creation cost by 87% and lifted purchase intent by 5%. The company now runs more than 500 AI applications across marketing, R&D, supply chain and innovation.
The point of Unilever's case isn't "we replaced the agency". It's that the internal brand team moves from approving one shoot per quarter to iterating daily variants of the same digital asset across markets. The creative still decides; AI removes the physical friction of the set.
3. B2B pricing that recovers basis points left on the table
McKinsey estimates that AI pricing tools add 2 to 5 percentage points of EBITDA in B2B when applied where leverage is highest: 60th–80th volume percentile customers, products with high discount dispersion. A documented case at a specialty-chemicals company with more than six business units put the impact at around $100M in additional earnings after micro-segmenting customers across 100+ attributes. Wilbur-Ellis, an agricultural distributor, reported a 2% margin uplift after activating real-time pricing on 6,000 SKUs with PROS.
Again, the pattern is not purely algorithmic. The engine proposes; the human seller closes. And the real unlock of those basis points sits in sales comp: if reps keep getting paid on gross volume, they'll leave margin on the table regardless of what the model suggests.
Operations: lines, supply and assets that keep running
4. Predictive maintenance that turns downtime into capacity
Siemens Senseye, deployed across more than 100,000 machines in 650+ sites, documents unplanned-downtime reductions of up to 50%, maintenance-efficiency gains of up to 55% and ROI in under three months. At BMW Regensburg, the system has saved more than 500 minutes of disruption per year on robotic lines, with the Siemens-BMW digital-twin integration delivering what BMW itself quantifies at up to a 20% production-efficiency gain.
For a CFO the maths is direct: every hour of stopped line at a vehicle or food manufacturer translates into five- or six-figure numbers. Halving unplanned downtime is, in effect, buying capacity without buying more fixed assets.
5. Demand and inventory forecasting at granular precision
Walmart's public documentation on ML applied to supply chain reports up to 90% inventory-accuracy improvements, 30 million driving miles avoided through route optimisation, and a 68% success rate on automated supplier negotiations with an average 3% cost saving. The compounding effect McKinsey and the NRF estimate in retail: 8% annual profit growth in 2023 and 2024 for those who deployed ML across forecasting and pricing, versus competitors who didn't.
The real lever for a mid-market company isn't to replicate Walmart. It is to understand that a one-point improvement in gross margin through reduced shrink and stock-outs translates into multiples of that point in operating profit. And for mid-market, that improvement is reachable today with off-the-shelf tools on top of SAP, Oracle or Microsoft Dynamics.
6. Software engineering that ships faster
GitHub's controlled experiment on Copilot measured a 55% speed-up in time to complete a development task (71 minutes vs. 161) and an 8-point-higher completion rate (78% vs. 70%). The piece usually left out of the quote is the other one: freed-up time doesn't automatically turn into more shipped features. It turns into whatever the team decides to do with it.
The companies that capture it well redirect the margin into tech-debt reduction, better test coverage and more design time. Those that redesign nothing watch the 55% evaporate into extra meetings or more mediocre pull requests. AI accelerates; direction stays human.
Back office: finance, legal and talent
7. Financial close and reporting, from weeks to days
A MIT/Stanford study on AI adoption in finance functions found that teams using it cut 7.5 days off the monthly close, raised report detail by 12% and shifted 8.5% of time from routine processing to analysis. The Hackett Group's 2026 benchmark places mid-market companies with agentic workflows in production at a close cycle of 1.8 days vs. 6.2 days sector average.
Documented cases: Tower Federal Credit Union closes in two to three days; Ralph Lauren compressed a four-week, multi-person process into "one person, less than a day". For a CFO the impact isn't just headcount efficiency; it's that the board gets firm numbers to decide on one or two weeks earlier, and that changes the quality of every capital decision.
8. Contract review and legal analysis
JPMorgan put COiN (Contract Intelligence) into production in 2017 to analyse commercial loan agreements. The system removes 360,000 hours of manual review per year, estimated at $144M in labour cost, and processes in seconds what used to take weeks of lawyer and loan-officer time. Clause-extraction accuracy beat the human baseline on the bank's internal benchmarks.
What matters for mid-market is that the same capability is available today in commercial legal-tech tools at standard software pricing, with no need to replicate COiN. If your legal team still reviews NDAs line by line, you are paying for contractor work your P&L no longer needs to absorb.
9. Recruiting with less friction and more diversity
Unilever has published precise numbers on its recruiting redesign with HireVue and Pymetrics: two million applications a year, 30,000 hires, graduate hiring cycle compressed by 90% (from four-plus months to four weeks), over £1M in annual savings, 50,000 hours reclaimed and a 16% lift in workforce diversity because the model weights cognitive and behavioural traits over resumes.
A CHRO should frame this for the CEO in direct terms: if we take 90 days to fill a role and our competitor takes 15, the competitor wins the candidate who answers first. AI doesn't change your employer value proposition; it changes response time, which in tight markets is literally worth the hire.
Innovation: R&D that compresses the cycle
10. Discovery and product design
Moderna rebuilt its discovery pipeline with AWS to go from manufacturing 30 mRNAs per month to more than 1,000, with robotic automation and AI models handling QC and analysis. Pfizer reports compressing discovery cycles "from years to 30 days" on selected cases and saving 16,000 hours annually in research alone, with 10% manufacturing-yield gains and 25% cycle-time reductions. Eli Lilly has deployed an NVIDIA DGX SuperPOD with 1,016 Blackwell Ultra GPUs dedicated to compressing pharmaceutical-discovery timelines.
The pharma case is extreme and doesn't transfer directly to a mid-market company. The signal does. Any product-engineering process with expensive iterations (industrial design, chemical formulation, materials development, firmware testing) is susceptible to the same compression. What takes three sprints today can take one if the engineer works with models that simulate and prune options before ever reaching the test bench.
The common pattern: why "AI alone" doesn't win and "AI + team" does
What separates the companies capturing EBIT with AI from those just stacking up pilots?
Rereading the ten cases, a clean, repeating pattern emerges.
The results hold when AI enters as a multiplier of human capabilities; when it enters as a pure substitute, the numbers fade. The novice call-centre agent reaches the veteran's level with AI whispering in their ear. The Senseye-equipped plant engineer prioritises better which asset to inspect, and still performs the inspection. The B2B rep uses the pricing engine to propose, and closes the deal themselves. The JPMorgan lawyer confirms what COiN has flagged red; they don't sign without reading.
The failures, when they happen, follow the inverse pattern. When Klarna pushed automation to 100%, it had to bring humans back for the hard cases. McKinsey documents that companies focused only on efficiency capture less value than those that also chase growth. The Veracode GenAI Code Security Report 2025 finds that AI-generated code carries 2.74× more vulnerabilities when accepted without human review.
"AI doesn't replace senior judgment. It distributes it. Operating profit shows up when the people who couldn't previously execute with discretion now can, and the senior team is freed to decide where the company is actually played."
That is the operating thesis of this piece: AI produces margin because it raises the floor, not because it lowers the ceiling. It raises the floor for the junior, the novice, the generalist, the team that was under water. It frees the senior for higher-leverage calls. What it does not do, in any case documented with serious numbers, is remove the need for qualified human judgment.
AI is a tool, and this blog is no exception: here we openly use AI to research, fact-check and support drafting. The ideas, the stance and the judgment stay human. That's exactly the relationship we recommend inside the companies we advise.
What a C-level should move this quarter
If you run a mid-market business and want to move from noise to P&L, here are the concrete moves for the next 90 days, by lever and with a clear owner per role.
1. CEO: pick only two levers. Companies that capture EBIT pick two priority functions and redesign them deeply. Those that capture nothing launch transversal "AI initiatives". Take the list of ten, cross-reference with where your margin pressure or volume bottleneck is worst, and sign off on two written objectives with executive sponsorship.
2. CFO: demand the business case per lever, never aggregated. Each initiative has to reach the board with an operating metric tied to P&L: gross-margin points, DSO days, cost per transaction, close days, time-to-hire. The enterprise-level "value generated by AI" number without traceability is, per McKinsey's data, the clearest indicator that there's no real value.
3. COO: redesign flows, don't layer AI on top of today's process. McKinsey's finding is clean: the companies that see EBIT are the ones that changed the workflow. Putting Copilot on top of a QA pipeline still measuring tickets-closed-per-head returns nothing. You have to change what gets measured.
4. CHRO: design the team's capability curve. If AI adds +34% for novices and 0% for seniors, the career-track problem changes: the junior learns faster, and the senior needs a clear step up toward judgment, decisions and coordination. Without that redefined curve, the team flattens, senior attrition climbs and the organisation loses its scarcest asset.
5. The four together: write down what stays human. Klarna learned the hard way that certain tiers (delicate claims, regulatory-risk cases, retention conversations with high-value clients) need a human hand. Listing those tiers in writing before deploying avoids having to walk things back nine months later.
6. Secure the data before spending on the model. None of the ten cases works on an incoherent data lake. Before signing a generative-AI licence, have the data team validate that the sources feeding the model are clean, versioned and traceable. It's the least visible item on the list, and the only one without which the rest is noise.
If you make only these six moves, the next strategic planning cycle will arrive with hard data on what works in your house and what doesn't, rather than with a dashboard of pilots no one can read.
The market-level picture will look the same one more year: almost everyone deploying AI, few capturing value. The difference between staying at the 61% with no EBIT impact and joining the 6% of high performers comes down to having fewer initiatives, deeper ones, redesigned around people who know what they are doing.
If you want to work through which two levers are the right ones for your company and what the flow redesign looks like, our advisory programs are built for that conversation, with decisions your board can sign off on Monday.
