Building the Business Case for AI

The $200,000 Mistake Nobody Caught

Picture this. Leila is a product manager at a mid-size e-commerce company. She just sat through a dazzling vendor demo — AI that writes product descriptions automatically. The sales rep showed it generating beautiful copy in seconds. Leila's boss loved it. The vendor quoted $200,000 for a custom model.

Leila was about to sign the purchase order... until a colleague asked one simple question:

"Wait — can't ChatGPT already do this?"

Turns out? Yes. Yes it can. They integrated an off-the-shelf API in two weeks for under $3,000. Same result. $197,000 saved.

This module gives you the framework Leila should have used before that meeting ever happened. Because the hardest part of an AI business case isn't the technology — it's knowing when to build, when to buy, and when to walk away.

(Illustrative scenario. The cost figures are representative of real-world build-vs-buy decisions in AI product development.)

Build vs. Buy vs. Partner: The Coffee Machine Analogy

Think about coffee for a second. You have three options every morning:

Here's the thing most people get wrong: you only build your own coffee machine if coffee IS your business. If you're a law firm, you buy coffee. If you're Blue Bottle Coffee, you build your own roasting process because that's your competitive edge.

The same logic applies to AI.

The Decision Tree (Your New Best Friend)

This flowchart is going to save you from so many bad decisions. Tape it to your monitor. Seriously.

Start at the top. Ask yourself: "Is this AI feature the core reason customers pick us over the competition?" That's what "competitive moat" means — it's the thing that makes you hard to replace, like a castle surrounded by water.

If the answer is no, you're already on the "Buy" side of the tree. Most teams skip this question and jump straight to "let's build something cool!" — and that's how you end up spending six months on a custom model when a two-week API integration would have worked fine.

Here's the key insight: only ONE path leads to "Build." You need all three yeses — moat, ML team, deep customisation. Every other combination points you toward Buy or Partner.

There Are No Dumb Questions

Q: What exactly is a "competitive moat"? A: It's the unique advantage that keeps your customers from switching to a rival. Think of it like an actual moat around a castle — it's the thing that protects you. For Netflix, it's their recommendation algorithm trained on billions of viewing hours. For your average law firm, AI document review is useful but it's NOT a moat — any firm can buy the same tool.

Q: What if I'm not sure whether something is a moat? A: Ask yourself: "If a competitor shipped this exact same AI feature tomorrow using the same API, would we lose customers?" If yes, it might be a moat. If no, it's a commodity. Most AI features are commodities.

Q: We don't have an ML team but we want to build anyway. Can't we just hire one? A: You can, but that's really a "Partner" decision disguised as a "Build" one. Hiring an ML team takes 6-12 months. In the meantime, buy or partner. Revisit "Build" once the team is up and running.

Three Companies, One Vendor, Three Very Different Answers

Here's a story that makes the decision tree real. Three companies all got the same sales pitch in Q1: "We can automate your document review with AI." Watch how the same pitch leads to three completely different (and correct) decisions.

Company A: Meridian Legal (the "Buy" story)

Meridian is a 50-person law firm. Dana, their head of operations, noticed paralegals were spending 20 hours per week on contract review. That's painful. But here's the question: is AI contract review the reason clients hire Meridian? Nope. Clients hire them for their lawyers' expertise. Contract review is a chore, not a selling point.

Decision tree path: Moat? No. Commodity task? Yes. Answer: Buy — Claude API with custom prompts.

A contractor built it in six weeks for $12,000. Paralegals went from 20 hours/week to 6 hours/week on contract review. That's 14 hours freed up for higher-value work — worth about $95,000/year in labor savings (at an estimated fully loaded cost of ~$130/hour for paralegals — illustrative rate; actual costs vary significantly by geography and firm size). Against a $12,000 build cost, that's a payback period of roughly seven weeks.

Company B: WorkforceIQ (the "Build" story)

WorkforceIQ sells enterprise HR software. Marcus, their CPO, made AI the centerpiece of their sales pitch. They have an ML team of eight engineers. They sit on ten years of proprietary HR adjudication records that no competitor can replicate.

Decision tree path: Moat? Yes. ML in-house? Yes. Deep customisation needed? Yes. Answer: Build — fine-tune on their proprietary data.

Six months later, in this illustrative scenario, their model outperforms general-purpose APIs by 31% on their benchmark. That performance gap now anchors every enterprise sales deck. This is what a legitimate "Build" decision looks like: proprietary data + in-house team + a measurable performance delta that turns into revenue.

Company C: The e-commerce company (the "$200,000 mistake" story)

You already met Leila. Her company wanted AI-generated product descriptions. No ML team. The vendor wanted $200,000 for a custom model. But product descriptions are textbook commodity work — every competitor will match this capability within six months using the same off-the-shelf APIs.

Decision tree path: Moat? No. Commodity task? Yes. Answer: Buy — a commodity LLM API (e.g., GPT-4o or equivalent), integrated in two weeks for under $3,000.

Leila's real job here wasn't to start a project. It was to stop a $200,000 mistake. Owning the decision framework saved more money before sprint one than most features earn after launch.

There Are No Dumb Questions

Q: Company B spent six months building. Wasn't that risky? A: Sure — but the risk was justified because they had all three ingredients: a moat, a team, and proprietary data. If any one of those was missing, the risk wouldn't be worth it.

Q: What if my company is somewhere between Company A and Company B? A: That's most companies! If you're unsure, start with Buy. You can always upgrade to Build later once you've proven the use case works. Going the other direction (build first, realize you should have bought) is way more expensive.

The Hidden Cost Trap: TCO (Total Cost of Ownership)

Here's where even smart people get fooled. They compare the sticker price of Build vs. Buy and forget about everything else. That's like comparing the price of a puppy ($500) to a robot dog ($200) without considering that the real puppy needs food, vet visits, training, and someone to walk it every day for 15 years.

Total Cost of Ownership means counting ALL the costs, not just the upfront ones.

Here's the trap Company B fell into (even though their decision was correct): that "Build" decision didn't end at launch. They now have a permanent team cost for maintenance, retraining cycles, and evaluation infrastructure. Their leadership knew this going in — and that's the difference between a smart Build and a reckless one.

Rule of thumb: When someone pitches "Build," multiply their cost estimate by 3x for year one and add 40% of the build cost per year after that for ongoing maintenance. If the business case still works at those numbers, it might genuinely be a Build.

The ROI Question: "What's the Return on This?"

Your stakeholder will ask this. Guaranteed. And "it'll be really cool" is not an answer.

To answer honestly, you need exactly three numbers:

1. Baseline metric — How does the current process perform today? (e.g., "Paralegals spend 20 hours/week on contract review")
2. Measurable delta — How much will the AI feature improve that metric? (e.g., "AI reduces it to 6 hours/week, saving 14 hours")
3. Total cost of ownership — What does it cost over the payback window? (e.g., "$12,000 build + $500/month API costs")

Then the math is simple:

ROI = (Value of the Delta - Total Cost) / Total Cost

For Meridian Legal: ($95,000 savings - $18,000 first-year cost) / $18,000 = 428% ROI in year one.

That's a business case. Not "we should use AI." Not "competitors are doing it." A real number that a CFO can evaluate.

<strong className="block text-slate-800">1. Define the problem</strong>
<span className="text-sm text-slate-600">What specific decision or task is taking too long, costing too much, or producing errors? Be precise — "improve customer experience" is not a problem definition.</span>

<strong className="block text-slate-800">2. Quantify the cost of the status quo</strong>
<span className="text-sm text-slate-600">Hours per week × headcount × loaded cost. This is your baseline. Without it, you can't calculate ROI.</span>

<strong className="block text-slate-800">3. Estimate AI impact</strong>
<span className="text-sm text-slate-600">Conservative: 20–40% time reduction on the specific task. Don't assume 100% automation — assume augmentation.</span>

<strong className="block text-slate-800">4. Calculate cost to build/buy</strong>
<span className="text-sm text-slate-600">Engineering time + API costs + ongoing maintenance. Include 3x for hidden costs (testing, monitoring, edge cases).</span>

<strong className="block text-slate-800">5. Identify the risk</strong>
<span className="text-sm text-slate-600">What happens when it's wrong? Who reviews outputs? What's the rollback plan?</span>

There Are No Dumb Questions

Q: What if we can't measure the delta yet? A: Then you can't calculate ROI yet — and that's okay! Run a small pilot first. Use the Buy path to test cheaply. Measure the actual delta. THEN build the business case with real numbers instead of guesses.

Q: What counts as "value" if the AI doesn't directly save money? A: Think about time saved (converted to salary costs), revenue unlocked (e.g., sales team can handle 30% more leads), risk reduced (e.g., fewer compliance violations at $50k each), or customer retention improved. Everything maps to dollars eventually.

Q: Why do AI product margins tend to look worse than traditional SaaS margins? A: As a rough industry benchmark, traditional SaaS gross margins tend to run 80–90%, while AI-powered products typically run lower (often 50–70% — illustrative range; actual margins vary widely; see public filings from AI-native companies for current benchmarks) due to inference costs — though this varies significantly by product and scale. Because every user query costs you money in inference (API calls). Traditional SaaS serves a web page — nearly free. AI SaaS runs a model — not free. Your levers to improve margins: cache common outputs so you don't re-compute them, use smaller/cheaper models for simple tasks, and make your prompts more efficient so each call uses fewer tokens.

Putting It All Together: Your Business Case Checklist

Before you walk into any meeting to pitch an AI investment, make sure you can fill in every row of this table:

If you can't fill in a row, you're not ready to pitch yet. And that's fine — go run a pilot, measure the baseline, get the numbers. A delayed pitch with real data beats an early pitch with vibes.

Now Try the Full Decision Tree Yourself

Back to Leila

Leila walked out of the meeting having not spent $200,000.

She spent $3,000 instead. Two weeks of integration. Same result.

Her colleague's question — "Wait, can't ChatGPT already do this?" — was the most valuable thing anyone said in that room. Leila built the decision tree into every vendor evaluation after that. Not because she wanted to be the person who says no, but because she wanted to be the person who knows when "yes" costs $197,000 more than it needs to.

Key Takeaways

• Start with Buy. You can validate any AI use case cheaply by buying off-the-shelf first — before investing in custom models. Think Starbucks before building your own roastery.
• Your moat isn't the model. Your real competitive advantage lives in your data, your UX, and your domain knowledge — almost never in the model itself. The model is the coffee bean. Your product is the entire cafe experience.
• Build = permanent commitment. Every time you choose Build over Buy, you're also signing up for ongoing maintenance, retraining, and evaluation infrastructure as a permanent team obligation. Make sure leadership sees that line item before the first model ships.