--- canonical: https://stredgepartners.co/leadership-tools/lacas/ai language: en-GB author: Bruno Hounkpati published: 2026-11-11 --- # The A in LACAS: AI that focuses on tomorrow Growth and margin last when AI forecasts the decisions that make money, and someone checks every forecast against what happened. A faster report of last month does neither. ## In brief 1. Judge an AI investment by the decision it changes before the event, not by how fast it reports after it. 2. Growth stops at the smallest of three numbers: profitable demand, compliant output and delivery capacity. Put AI on the decision that limits the smallest one. 3. A gain lasts when each forecast is checked against results in writing, owned below the CEO and funded through SPARK. Otherwise it leaves with you. ## Two screens that saw everything after it happened When I ran logistics in Nigeria, we built a control tower. Every truck fitted with an in-vehicle monitoring system reported to it. On one screen we followed routes, deliveries and delays across the network. Our on-time, in-full deliveries rose to more than 30% above the service level we had committed to our customers. It was a good tool. It showed us everything that had already happened. What it could not foresee was the next driver to cause the next road crash. Or the next failed delivery to a key account because of a production issue at the plant. We saw both on the screen, after they happened. As a CEO, in a business I ran, I asked my finance team for a customer dashboard. They built a solid one. Every customer on one screen: details, activity level, three years of purchase history, payment patterns. What it could not predict with any precision was the next customer to default on its credit terms, stop paying and become a bad debt. Two good tools, in two different functions. The same blind spot. Both described the past very well. Neither told us what was coming next. That gap is the AI element of LACAS. ## The test: reading yesterday faster, or sensing tomorrow LACAS asks one question of a leader: what will still work after you leave? Its AI element has a test question of its own. Is AI a transformation and growth lever, sensing tomorrow, not just reading yesterday faster? Much of the AI money I see goes to the second half of that sentence. Monthly packs produced in two days instead of ten. Meeting summaries. Dashboards that answer questions nobody acts on. Useful, sometimes. A faster report of last month still changes nothing that happens next month. A US survey published by Wharton in October 2026 points the same way. Among 1,007 managers and executives surveyed in July 2026, the most common way to judge an AI investment was hours saved (40%). Net-new revenue was the least used of the major measures (27%). Yet 35% of departments tracking at least one growth measure reported strongly positive returns, against 13% of those tracking efficiency alone. The returns are self-reported and the link is an association: companies already growing with AI may simply measure growth. Hours saved measures how fast yesterday gets read. It says nothing about tomorrow. The AI element earns a good rating when three conditions hold. It forecasts something that moves revenue, cost or safety. The forecast feeds a named decision, with a person authorised to act on it. And every period, someone checks the forecast against what happened, in writing. Rate yourself on the LACAS scale: "1 depends entirely on you, 3 works with your regular attention, 5 works without you." A company whose AI produces monthly reports faster, with no forecast checked against results, rates 1. The lead indicator is simple to state and hard to fake: the key decisions built on AI forecasts that are checked against results. No evidence, no rating. ## Growth: put AI on the smallest of three numbers Additional revenue needs three things in the same period. Customers who will buy at a profitable price. Output that meets specification. Capacity to deliver it. The smallest of the three sets what you can sell. Take an illustrative plant. Better process control could add 80,000 compliant tonnes a year. Customers exist for 60,000. Dispatch can move 50,000. The business case is 50,000 tonnes, and only at full ramp. The other 30,000 tonnes stay an option until demand and dispatch catch up. A model that adds output while dispatch is the constraint moves the bottleneck from the mill to the loading bay. **Exhibit 1 · Additional sales are capped by the smallest number (illustrative)** | Compliant output | Profitable demand | Delivery capacity | Sellable | |---:|---:|---:|---:| | 80,000 t | 60,000 t | 50,000 t | **50,000 t** | So the first question in an AI growth case is which of the three numbers is smallest, and which decision limits it. The model comes after. Mining gives a clean example. BHP reports that a digital tool at Escondida, its copper mine in Chile, used AI to increase revenue by US$18.9 million in FY2024 by helping drill and blast teams choose a better blast pattern. The same technology helps relieve coarse-ore restrictions at the grinding mills. At its BMA coal operations, improvements to process set-point optimisation increased production of on-specification steelmaking coal and revenue by US$15.5 million. These are company-reported revenue figures. They are not profit, and they come without full cost. What transfers is the mechanism: the AI sits on the decision that limits saleable output. The commercial side follows the same rule. A customer request that arrives by email, with a specification, a quantity and a date, can be read by AI, matched to approved products and checked against stock, credit and capacity before a person approves the offer. That is growth only when it becomes delivered orders that would not have happened otherwise. Moving the same orders from the telephone to a chat window is a channel change. ## Profitability that lasts: optimise the delivered result Margin leaks between functions. The plant optimises its cost per tonne, logistics its cost per trip, and nobody owns the cost per tonne delivered and paid. Two sources for the same product, the same customer and the same net price of €110 per unit (illustrative): | € per unit | Current source | Cheaper plant | |---|---:|---:| | Plant | 72 | 68 | | Distribution | 12 | 18 | | Service | 6 | 6 | | Delivered cost | 90 | 92 | | Contribution | **20** | **18** | The plant saves €4. The business loses €2 on every unit delivered. This is where forecasting and optimisation do their best work. Forecast demand by product and location. Let a solver allocate orders to plants and routes at the lowest delivered cost, within quality, safety and service limits. The planner sees the allocation and its constraints, and approves it. Two disciplines keep the gain honest. **Separate price from usage.** If energy cost rises from €10,000 to €12,000 for the same 100 MWh because the tariff moved from €100 to €120 per MWh, the plant floor lost no efficiency, and changing settings will save nothing. AI can flag what remains unexplained. Arithmetic does the reconciliation. **Turn a warning into planned work.** A model that detects abnormal vibration on a motor is worth something when the inspection is planned, the parts are on site and someone has the authority to stop the equipment. The measure is completed interventions. And never fund the AI programme by cutting maintenance. SPARK is clear on this: pay from bad cost. Process control shows what a credible result looks like. In March 2023, ENEOS Materials and Yokogawa reported that an autonomous control AI cut steam consumption and CO2 emissions by 40%, measured against the steam previously used to maintain liquid level in a distillation column under manual control. The field test began with 35 days of continuous operation, and the AI was then formally adopted for use at the plant. Forty percent of one duty, against a stated baseline. The plant's total energy bill is a different number. A claim that precise can be checked. Ask for the same precision in your own plants. ## Profit is not cash An executive committee approving an AI case for 2027 should read two lines, not one. **Exhibit 2 · One illustrative operating case, year one, before tax and financing** | | € thousand | |---|---:| | Extra volume sold | +625 | | Energy saved | +300 | | AI run cost | −180 | | One-off set-up | −120 | | Depreciation | −50 | | Additional operating profit | **575** | | Depreciation added back | +50 | | Capital expenditure | −600 | | Additional working capital | −473 | | Incremental cash | **−448** | | Peak funding | **913** | The case adds saleable volume from July with a three-month ramp. Operating profit is positive in the first year. Cash is negative, and the peak funding need is larger than the profit. A positive EBIT line does not settle the funding decision. Credit risk shows why. A customer who buys and never pays appears first in revenue, then in a provision for bad debt, and never in cash. The committee needs the monthly cash calendar, the downside if adoption slips, and the names of the people who must stay in place to deploy the change. AI has a role on both large cash outflows. On capital projects, a forecast of spend and progress against plan flags an overrun before it lands, and site reports and photos read by AI flag quality and safety issues early. On working capital, a credit forecast can cut overdue receivables and a demand forecast can cut stock. ## What makes it last This is the part most AI programmes skip, and it is why the gains leave with the sponsor. Four things, written down. 1. **A forecast ledger.** For each forecast that feeds a key decision: what was predicted, what happened, the gap, and what changed in the model or in the decision as a result. Reviewed with the numbers, beside EBIT, every quarter. Credit shows how the check works. Of the customers the model flagged, how many fell behind? Of the customers who fell behind, how many had it flagged? The first answer tells you what false alarms cost in tightened terms and lost sales. The second tells you how much bad debt it missed. 2. **An owner below the CEO.** One person accountable for each forecast loop, with the authority to act on it. 3. **Decision rights.** What the AI may read, what it may write, which systems it may trigger, and what always needs a person's approval. In a plant, no language assistant writes to process control. In credit, the model flags a customer and the credit committee decides the terms. AI agents that execute raise the stakes: they repeat a wrong forecast as fast as a right one. Give a forecast the right to act only when its ledger justifies it. 4. **A lead indicator that passes SPARK.** Fund it only if no score is below 3 and the average is 4 or more. Score it with its owner and with one person who does not own it. Then rate the AI element again. If it still depends on you, it is a project. It is not yet part of your platform. ## What it takes: data, models, people, process Four decisions sit with the CEO, and none of them is choosing a model. 1. **Data.** One owner for each data set a forecast needs, and a decision on where that data stays. 2. **Models.** The job first, the model second. Forecasts of demand, credit risk or process response come from machine learning trained on your own history, often paired with a solver that tests the options within your constraints. Language models (LLMs) do the text work: contracts, service cases, site reports. All of it is AI, and it helps a decision maker for the same reason: it shows the person who decides what is coming, early enough to act, and the ledger shows whether to trust it. 3. **People.** One owner per forecast, below the CEO, and training for the people who act on it. 4. **Process.** The step each forecast feeds, who may override it, and how overrides are recorded. Leave the choice of tool to the CIO, or to a chief AI or data officer if you have one, against five tests: accuracy tested on your own history, where your data goes, what it can do without your sign-off, cost per use and exit terms. ## The arithmetic of ambition If revenue rises from 100 to 130 while total costs fall from 80 to 60, operating profit goes from 20 to 70. That is 250% more profit, and a margin that moves from 20% to 53.8%. The arithmetic works. Scattered pilots will not deliver it. A target of that size needs new offers, new routes to market or new capacity, with AI inside the decisions that make them work. Define the target precisely too: a 25% margin improvement takes a 20% margin to 25%, not to 45%. ## Five questions for your next executive committee 1. Which of our key decisions runs on a forecast today, and who checks it against results? 2. Between profitable demand, compliant output and delivery capacity, which is our smallest number, and which decision limits it? 3. Which cost do we optimise inside one function when we should optimise it delivered? 4. For each AI case in the 2027 budget, what does it do to EBIT and what does it do to cash, month by month? 5. If the sponsor left tomorrow, which AI forecast would keep running, and who would check it? ## Pull quote "A faster report of last month changes nothing that happens next month." ## What to instrument (evidence block) - **Forecast against actual on key decisions:** predicted, observed and gap, each period, by owner. - **Delivered contribution per unit:** net price after rebates, less plant, distribution and service cost. - **Completed interventions per alert:** alerts that became planned, completed and verified work. - **Bad debt flagged in advance:** the share of write-offs on customers the model had flagged at least one quarter earlier. ## FAQ (visible on the page) **Is generative AI part of the AI element of LACAS?** When it feeds a forecast or a decision that someone acts on, yes. Drafting and summarising faster is reporting, and it rates as reporting. **Where should an industrial company start?** At the smallest of its three numbers (profitable demand, compliant output, delivery capacity), with one decision and one owner. Choose the technology afterwards: forecasting, optimisation or an assistant. **How do we know an AI gain will last?** When its forecasts are checked against results in writing every quarter, an owner below the CEO is accountable for the loop, and its lead indicator passes SPARK. ## Continue reading - LACAS, the legacy test: https://stredgepartners.co/leadership-tools/lacas - SPARK, the funding test for indicators: https://stredgepartners.co/leadership-tools/spark ## Call to action Rate the AI element of your company with LACAS: https://stredgepartners.co/leadership-tools/lacas ## Sources - BHP Group, Annual Report 2024 (Form 20-F, FY2024), Operating and Financial Review, "Technology and innovation". Company-reported revenue figures. https://www.sec.gov/Archives/edgar/data/811809/000119312524210297/d812514d20f.htm - ENEOS Materials and Yokogawa Electric, joint release, 30 March 2023, "Autonomous control AI officially adopted for use at an ENEOS Materials chemical plant", footnote 4. https://www.eneos-materials.com/english/information/2022/20230330_01.html and https://www.yokogawa.com/news/press-releases/2023/2023-03-30 - Wharton Human-AI Research and GBK Collective, "The Adaptive Advantage: Who's Winning with AI and How", 2026 AI Adoption Report, October 2026, pp. 11 and 31. US online survey, 9 to 29 July 2026, managers and above (n=1,007). Returns are self-reported. https://knowledge.wharton.upenn.edu/wp-content/uploads/2026/10/Wharton-GBK-AI-Adoption-Report-2026.pdf - LACAS and SPARK, Bruno Hounkpati, Stredge Partners, version 1.0. https://stredgepartners.co/leadership-tools - Exhibits 1 and 2, the delivered-cost table, the energy example and the ambition arithmetic are Stredge Partners illustrations, not client results or forecasts. © 2026 Bruno Hounkpati · Published by Stredge Partners with his permission · Free to use with attribution, without modification.