The AI Margin Squeeze: When Every New Customer Costs More to Serve

(Illustration of a profit margin squeezed between client price pressure and rising AI cost to serve in IT companies) Ask any IT services founder what a new customer means, and you’ll hear the same thing: growth. A signed contract, a bigger pipeline, a better quarter. But there’s a question that rarely gets asked in the celebration meeting: what does it cost us to serve this customer? For decades, the answer was predictable. A new client meant a few more engineers, a few more laptops, and a project manager. The cost scaled in a straight line with revenue, and as long as the rate card stayed healthy, so did the margin. AI has bent that line. And for many average IT companies, it has bent it in the wrong direction. What “Cost to Serve” Really Means Cost to serve is everything it takes to deliver value to one customer, beyond the obvious salary line. It includes the people who build the work, the tools they use, the cloud resources they consume, the managers who review the output, the support that follows the launch, and the rework when something goes wrong. Traditional IT firms tracked this loosely. Utilization and billing rates were the headline numbers, and as long as they looked healthy, nobody dug deeper. That worked when delivery was mostly human effort. It doesn’t work when delivery is a blend of people, models, pipelines, and usage-based bills that change every month. The Promise: AI Will Lower Our Costs Let’s be fair to the optimists. AI really does reduce effort in many places: Developers get first drafts of code in seconds Test cases can be generated automatically Documentation and release notes take minutes instead of hours First-level support can be handled by chatbots Routine analysis and reporting can be automated Used well, these gains are real. The trouble is what happens to them after they appear. The Reality: Costs Didn’t Disappear. They Moved. Here’s the part that surprises many leaders. When you remove one kind of cost, you rarely end up with zero. You end up with a different kind. In AI-assisted delivery, the cost profile shifts in several ways. Costs that used to be fixed are now variable A developer’s salary is the same whether they write 100 lines or 1,000. AI usage isn’t. Every prompt, every generated output, every automated test run can carry a usage charge. Heavy use across a large team and multiple clients creates a bill that grows with activity, not with headcount. That makes costs harder to forecast, and much harder to price into a fixed contract. Review became the new bottleneck AI produces output fast, but someone must check it. Generated code can look correct and still hide security flaws, performance problems, or subtle logic errors. The person doing that checking is usually a senior engineer, the most expensive and scarcest resource in the building. So the work that used to take a junior developer two days might now take a machine ten minutes plus a senior’s two hours. Faster, yes. But cheaper in the way the spreadsheet promised? Often not. Data work is the hidden iceberg Most AI projects aren’t really about models. They’re about data: cleaning it, organizing it, securing it, and making it usable. In practice, this is where budgets go to die. Clients assume their data is ready. It rarely is. Every new customer arrives with a different mess, and each mess needs custom effort. That’s not a cost that shrinks with scale. Talent got more expensive and more specialized Engineers who can build, evaluate, and maintain AI systems command premium salaries. Meanwhile, the existing workforce needs training to stay relevant. Both are real expenses, and neither shows up as a line item on any customer’s invoice. Governance and risk added overhead Clients now ask hard questions: Where does our data go? Which model is being used? Who is accountable if the output is wrong? Answering them takes legal review, security assessments, documentation, and sometimes custom contractual terms. Each new customer brings its own version of these questions. The overhead scales with the number of clients, not with the size of the contract. The Squeeze: Two Forces Pressing Together Now put the two sides together. On one side, revenue per project is under pressure. Clients know AI exists. They’ve seen the demos. They expect lower rates, shorter timelines, and smaller teams. Procurement departments have started asking for “AI efficiency discounts” before the contract is even drafted. On the other side, cost to serve is getting more complex and often more expensive. Usage bills, senior review time, data preparation, specialist talent, and compliance work all add up. That’s the margin squeeze: price pressing down from above, costs pushing up from below, and the profit sandwiched in the middle getting thinner with each new customer. The painful twist is that growth makes it worse, not better. Landing a new client now adds more custom data work, more governance conversations, and more review load. If your contract is time-and-material or priced low to win the deal, each new customer can quietly become less profitable than the last. Why Average Firms Feel It More Than Others Large global players absorb these costs through scale. They have proprietary platforms, negotiated cloud rates, reusable accelerators, and dedicated governance teams. Their cost per additional customer really does fall. Small specialists survive by charging premium prices for deep expertise. Clients pay for the niche, not the hours. The firms in the middle have neither advantage. They have: Limited pricing power, because their services look similar to competitors’ No reusable platforms, so each project is built close to scratch Thin senior benches, which makes the review bottleneck painful Little visibility into true costs, since they never tracked them closely They’re growing revenue, sometimes impressively, while the profit per customer quietly shrinks. By the time it shows up in quarterly numbers, the damage is already done. Five Warning Signs You’re in the Squeeze How do you know if this is happening to you? Look for these signals: Revenue is up, but gross margin is flat or falling. The classic symptom. Your cloud and tool bills are growing faster than your headcount. Usage costs are creeping in. Senior engineers are overloaded. They spend more time reviewing and fixing than designing. Pilots and proofs of concept multiply, but few reach production. Effort goes in, revenue doesn’t follow. You can’t answer a simple question: “What did it actually cost us to deliver that project?” If two or more of these sound familiar, your cost to serve deserves attention before your next growth push. How to Break the Squeeze The good news is that this is a business model problem, and business models can be redesigned. Here’s what the stronger firms are doing. Measure cost to serve properly Start by breaking down delivery cost per customer: people time, AI tool licenses, compute usage, review hours, rework, and support. You can’t fix what you can’t see. Many leaders discover that one or two “great” accounts are actually losing money once AI costs are counted. Price for value instead of effort If a task now takes a third of the time, billing by the hour punishes you for being efficient. Fixed-price and outcome-based contracts let you keep part of the productivity gain. The client gets a predictable result, and you get paid for delivering it rather than for how long it took. Build reusable assets Every integration, workflow, or accelerator you build once and reuse lowers your cost for the next customer. This is the only reliable way to make cost to serve fall as you grow. Treat repeatable work as a product, not a project. Control usage like a utility bill Set budgets for AI usage per project. Choose the right model for the right task instead of defaulting to the most expensive one. Monitor consumption weekly, not annually. A little discipline here prevents unpleasant surprises at month end. Invest in training as a cost-control measure An untrained team uses AI tools inefficiently: vague prompts, repeated attempts, unchecked outputs, and expensive rework. A trained team gets better results in fewer iterations and catches mistakes earlier. Structured training in practical AI skills often pays back through lower waste alone. Qualify customers more carefully Not every customer is worth winning. Some bring messy data, unclear goals, and unrealistic price expectations. Learn to spot them early and either price the risk properly or walk away. A smaller book of profitable customers beats a large book of draining ones. Move toward higher-value work As routine delivery gets cheaper, the value moves to judgment: architecture decisions, business understanding, risk management, and accountability for outcomes. Firms that build this capability can charge for it, and it’s the part AI is least able to replace. A Quick Self-Check for Leaders Before your next growth meeting, try answering these questions honestly: What is our gross margin by customer, after AI-related costs? Which three accounts cost us the most to serve relative to what they pay? How much of our delivery work could be packaged and reused? Do our contracts reward efficiency, or punish it? Does every team lead understand what AI usage costs the company? If you can’t answer most of these, you’re not alone. But the firms that can will be the ones still profitable when the pressure intensifies. The Bottom Line AI hasn’t made IT services a worse business. It has made the old accounting of that business unreliable. When delivery shifts from pure human effort to a mix of people, models, and usage bills, the cost of serving each customer stops following the old patterns. The danger isn’t that AI is expensive. It’s that average companies keep chasing revenue growth without noticing that each new customer is getting more expensive to serve. The winners won’t be the firms that adopt AI fastest. They’ll be the ones who understand their numbers, redesign how they price and deliver, and build assets that make the next customer cheaper than the last. Growth is still the goal. But in the AI era, the question to ask isn’t just “How many new customers did we win?” It’s “What did each one really cost us to serve?”