Build vs. Buy AI Tools: Mid-Market Framework

Key Takeaways

• MIT's "buy, boost, or build" framework defines three AI adoption paths: adopt off-the-shelf (buy), customize with your data (boost), or develop from scratch (build)
• "Boost" is the sweet spot for most mid-market use cases, delivering most of custom-build differentiation at a fraction of the cost and risk
• External vendor partnerships succeed 67% of the time vs. 33% for internal custom builds, according to MIT's NANDA Initiative (2025)
• Five factors determine your path: competitive differentiation, data readiness, technical talent, time-to-value, and total cost of ownership. The interactive decision tree below walks you through all five.
• Over-buying is as dangerous as over-building: companies that only purchase tools end up with sprawl, redundant spending, and zero differentiation

The Buy, Boost, or Build Framework Explained

MIT CISR researchers Nick van der Meulen and Barbara Wixom published the buy, boost, or build framework in September 2025. McKinsey's Eric Lamarre independently described the same three paths as "take, shape, and make." The convergence is telling: two of the most respected institutions in business strategy landed on the same taxonomy.

Here's what each path means for a mid-market company:

Buy means adopting an off-the-shelf AI solution. The vendor provides, runs, and maintains the model. You get fast deployment and low risk, but every competitor using that same tool gets the same capabilities. Think Microsoft Copilot, ChatGPT Team, or Salesforce Einstein.

Boost means starting with a vendor's base model and enhancing it with your proprietary data. This is where techniques like RAG (retrieval-augmented generation) and fine-tuning come in. You don't need ML engineers for this. You need clean data, a clear use case, and someone who knows how to connect the pieces.

Build means developing, running, and maintaining the entire AI solution in-house. You control everything from model selection to deployment infrastructure. Maximum differentiation, maximum cost, maximum risk.

Save as PNG

The framework isn't a one-time choice. Most companies that succeed with AI use all three paths for different use cases. You buy commodity capabilities, boost where you have unique data, and build only where AI is the product or the core differentiator.

Five Questions to Determine Your Path

First, Check Whether You Actually Need AI

When a mid-market ops team tells me they need "an AI solution," roughly half the time what they actually need is a workflow that routes data between systems they already own. Lead routing, invoice processing, onboarding checklists: an n8n workflow, Zapier flow, or Power Automate sequence handles these without touching AI. S&P Global found that 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the year before, often because the problem never required AI in the first place.

Automate the routine stuff first. The use cases that genuinely need AI become obvious once you do.

For those remaining use cases, I use a five-question decision process. The questions are sequential: each answer narrows the options.

1. Does this use case touch a core competitive differentiator?

If the AI application directly creates competitive advantage, such as proprietary pricing algorithms, specialized analytics, or customer experience that defines your brand, buying off-the-shelf won't cut it. You need boost or build. If it's a commodity function like email summarization or meeting transcription, buy and move on.

2. Do you have clean, structured data specific to this use case?

Boost and build paths both require your data to be organized and accessible. If your CRM is a mess, your documents live in 14 different shared drives, and nobody has touched data hygiene in years, start with the AI readiness checklist before making any tool decisions. In my experience, data preparation eats 20–30% of any AI project budget and 60–80% of the timeline.

3. Do you have technical talent on staff?

You don't need a machine learning team. One strong developer or a technical operations person can manage boost implementations. Build requires deeper expertise: either an AI engineer on staff or a fractional resource who can architect and deliver the solution. No technical talent at all? Buy is your safest path until you address that gap.

4. Do you need results in weeks or can you invest months?

Buy deploys in days to weeks. Boost takes 4–12 weeks for a typical mid-market implementation. Build runs 3–12 months before you see production value. If the CEO needs results this quarter, building from scratch isn't realistic.

5. What's your realistic budget for this initiative?

A mid-market AI tool stack (buy path) runs $5,000–$10,000/month. Boost implementations range from $5,000 for small integrations to $100,000 for a full RAG deployment. Custom builds start at $50,000 and can exceed $500,000 before you account for ongoing maintenance. Match the path to what you can actually fund for 12+ months, not just the initial project. Before locking in your budget, check whether grants and tax credits for AI projects apply. The R&D credit alone can return 6–10% on qualifying build and boost work.

Save as PNG

How I Apply This Framework to My Own Tool Decisions

I run every tool decision through these same questions. When I need to connect AI to common SaaS platforms like Salesforce, HubSpot, or Google Workspace, I buy pre-built integrations. The use case isn't differentiating, pre-built options work well, and my time is better spent elsewhere.

When I need AI to work with client-specific databases and legacy systems, I build custom solutions. That's high-value work where the integration logic itself creates something no off-the-shelf tool can replicate. The UpSkalr case study shows three production applications built this way.

For AI workflow automation, where multi-step orchestration logic is intellectual property, I build. For governance frameworks, I build templates once and deploy them repeatedly. Same logic applies to any scale: if it's commodity, buy it; if it needs your data or your process logic, boost or build it.

What Each Path Actually Costs at Mid-Market Scale

Cost is where most "build vs. buy" articles get vague. Here are real numbers scaled for a mid-market company with roughly 100 employees.

The Buy Path: $5,000–$10,000/Month

A typical mid-market AI tool stack includes Microsoft Copilot at $30/user/month (50–100 seats: $1,500–$3,000/mo), ChatGPT Team or Claude Team at $25–$30/user/month (20–50 seats: $500–$1,500/mo), and Salesforce Einstein at $50–$125/user/month (10–20 sales reps: $500–$2,500/mo).

That's $5,000–$10,000/month before you factor in that, according to Zylo's 2025 SaaS Management Report, 52.7% of SaaS licenses sit idle. You're likely paying for seats nobody uses.

The Boost Path: $5,000–$100,000 Setup + $2,000–$20,000/Month

Boost implementations vary widely based on scope. Small integrations connecting AI to your existing tools run $5,000–$25,000. RAG deployment on your company data costs $45,000–$100,000. Fine-tuning a model with industry-specific documents runs $500–$5,000 per model. Ongoing operational costs add $2,000–$20,000/month.

Before you measure AI ROI, factor in the hidden cost: data preparation eats 20–30% of your budget and 60–80% of your timeline.

The Build Path: $50,000–$500,000+ Development

Custom AI development costs depend on complexity. Simple AI like chatbots or basic automation runs $50,000–$100,000. Medium complexity projects with ML-powered recommendations or analytics cost $60,000–$250,000. Advanced builds involving deep learning, NLP, or computer vision range from $150,000–$500,000.

The real cost shock comes after launch. A minimum viable in-house AI team runs $520,000–$825,000/year fully loaded. Annual maintenance adds 15–30% of the initial build cost every year. Outsourcing development can cut costs 30–60%, but you still own the maintenance burden.

Save as PNG

Not sure which path fits your budget? Take the free AI readiness assessment to see where your data, team, and infrastructure stand today.

Why Most Mid-Market Companies Should Start with Boost

For the use cases that genuinely need AI (not just the workflow automation we filtered out above), the data overwhelmingly favors starting with boost. According to McKinsey's 2025 Global Survey on AI, 88% of organizations now use AI in at least one business function, but only 6% qualify as "AI high performers."

The 82% in between mostly stalled because they treated AI as a purchasing decision instead of an integration problem. They bought tools without connecting them to the workflows, data, and business rules that make AI actually useful. Boosting fixes that.

MIT's research found that external vendor partnerships (buy and boost) succeed 67% of the time. Internal custom builds succeed just 33% of the time. That's a two-to-one success rate advantage for approaches that start with vendor capabilities rather than blank-slate development.

For most mid-market companies I talk to, the wins come from connecting AI to real workflows, not from the AI itself. What that looks like in practice, without a machine learning team:

• RAG on your CRM and knowledge base: This is the boost project I build most often. Connect an AI model to your company's actual data so it gives answers grounded in your processes, products, and customer history, not generic answers from the internet. The difference in answer quality is immediate, and a typical deployment takes 4-6 weeks.
• Workflow-level customization: Configure AI tools to follow your specific business rules, approval chains, and output formats. This goes beyond the default "turn it on and hope" deployment.
• API integrations: The value isn't in the AI model itself, it's in the connection between the model and your business systems. Building these integrations so data flows automatically between tools is where most boost projects deliver their ROI.
• Industry-specific fine-tuning: Train a base model on your documents, terminology, and domain knowledge. A general-purpose AI that knows insurance underwriting guidelines outperforms a brilliant AI that doesn't.

Boosting works for most of these companies because the math is straightforward: you get most of the differentiation of a custom build without the six-figure development cost or the 12-month timeline. A well-scoped boost implementation reaches production in 4-6 weeks, fast enough to show results this quarter. And those results tend to be concrete: a sales team that gets answers in seconds instead of searching across three systems, or a weekly report that used to take 15 hours of manual work now generated automatically. When boost is scoped correctly, the first project pays for itself quickly enough to fund the next one.

A Fractional AI Director can architect and deliver the boost implementation, identifying which use cases warrant customization, building the integrations, and getting you to production without the six-month ramp of a full-time hire or the bias of a vendor partnership steering the recommendation.

When to Build, When to Buy, and the Danger of Only Doing One

Most build-vs-buy articles only warn against building. That's half the picture.

When to Build

Build makes sense in four specific situations:

1. The AI application IS your product or core differentiator. If the AI capability is what your customers are paying for, buying someone else's version defeats the purpose.
2. You have proprietary data that creates a defensible moat. If your data is so unique and valuable that enhancing a vendor model isn't enough, building gives you full control over how that data is used.
3. The orchestration logic is intellectual property. Multi-step AI workflows where the sequence, decision points, and business rules create value that competitors can't replicate by subscribing to the same tool.
4. You need custom integrations with legacy systems. When your technology stack is unique enough that no vendor integration covers it, building becomes the practical path. The legacy system integration guide covers five patterns for connecting AI to older infrastructure without replacing it.

Smaller organizations can sometimes justify custom builds more easily than large enterprises. The costs are lower (LLM subscriptions plus engineering time vs. enterprise infrastructure), the approval process is faster, and the iteration cycles are shorter. The blanket "don't build" advice assumes enterprise-scale complexity that doesn't always apply.

When NOT to Build

According to the RAND Corporation's research on AI project failure, the five root causes are misunderstood problem definition, insufficient training data, technology-first thinking, inadequate infrastructure, and problems too difficult for current AI. If three or more of those apply to your situation, building is premature.

Don't build for commodity functions. Don't build without clean data. Don't build if you need results in 60 days. And don't build because it sounds impressive. Technology-first thinking is how 80%+ of AI projects fail.

The Overlooked Risk: When NOT to Only Buy

I've walked into companies running ten or more AI tools where none of them talk to each other. Every tool solved a narrow problem in a vendor demo. None of them connect to the company's actual workflows. The result is a growing SaaS bill and zero measurable impact.

Zylo's 2025 research backs this up: 52.7% of purchased SaaS licenses sit completely idle. No competitive differentiation. Every competitor has access to the same vendor tools. Nobody has built anything proprietary. The AI budget grows every quarter while the business impact stays flat.

Tool sprawl without integration does as much damage as a failed custom build. And it's usually a symptom of a deeper problem: nobody owns the AI strategy, so every team buys their own tool. One department subscribes to a chatbot, another builds a custom solution for the same problem, and a third doesn't know either exists. Mid-market companies are usually small enough to avoid this, but I've seen it start creeping in around the 300-employee mark. Before you spend anything on AI tools, get your leadership team aligned on which path you're taking for each use case. That includes understanding the data privacy differences between AI platforms, since tool tier and data handling policies vary significantly across vendors. The what to look for in an AI consultant guide covers how to find someone who'll give you honest buy-vs-build advice without a vendor agenda.

The right approach for most mid-market companies is a hybrid: buy commodity capabilities, boost where you have differentiating data, and build only where the AI application creates competitive advantage you can't get any other way.

Frequently Asked Questions

When should a mid-market company build custom AI vs. buying off the shelf?

Build custom AI only when the application directly creates competitive advantage, you have proprietary data to train on, and you have technical talent (at least one strong developer) to maintain it. For commodity functions like email summarization, document search, or meeting transcription, buying is faster, cheaper, and less risky. The middle ground, boosting a vendor tool with your own data, fits most mid-market use cases where you need some differentiation without the full build commitment.

What is the buy, boost, or build framework for AI?

The buy, boost, or build framework was published by MIT CISR researchers Nick van der Meulen and Barbara Wixom in September 2025. "Buy" means adopting off-the-shelf AI as-is. "Boost" means customizing a vendor's AI with your proprietary data through techniques like RAG or fine-tuning. "Build" means developing the entire solution in-house. McKinsey independently named the same three paths "take, shape, and make."

How much does it cost to build vs. buy AI tools?

At mid-market scale (~100 employees), the buy path runs $5,000–$10,000/month for a typical AI tool stack. Boost implementations cost $5,000–$100,000 upfront plus $2,000–$20,000/month in operational costs. Custom builds range from $50,000–$500,000 in development, with an in-house AI team adding $520,000–$825,000/year. The hidden cost in all three paths is data preparation, which consumes 20–30% of any AI project budget.

What does "boosting" an AI tool mean in practice?

Boosting means enhancing a vendor's base AI model with your company's proprietary data. In practice, this looks like connecting an AI model to your CRM or knowledge base through RAG, fine-tuning a model with your industry-specific documents, or building API integrations that connect off-the-shelf AI to your business systems. You don't need machine learning engineers for this. You need clean data and someone who understands how to scope and guide the implementation.

How do I evaluate which AI approach is right for my company?

Start with five questions for each use case: Does it touch a core differentiator? Do you have clean data for it? Do you have technical talent? Do you need results in weeks or months? What's your realistic budget? An AI Strategy Assessment maps every use case to the right path so you invest in the approach that fits your data, team, and budget rather than defaulting to whatever a vendor is selling.

Ready to Map Your AI Use Cases to the Right Path?

The build vs. buy decision isn't one choice. It's a portfolio of decisions across every AI use case in your organization. Getting it wrong means either building something expensive that a $30/seat tool could handle, or buying generic software that gives you zero competitive advantage. An AI Strategy Assessment maps every use case to the right path, so you invest in the approach that fits your data, team, and budget rather than defaulting to whatever a vendor is selling.

Take the free AI readiness assessment to see where your data and team stand today, or book a free 30-minute AI strategy call to walk through the buy, boost, or build decision for your specific use cases.