Should You Build a Custom AI Tool or Just Use ChatGPT and Copilot?

Last updated: August 2026

For most mid-size companies in 2026, the answer is buy, or more precisely, configure, not build. MIT's 2025 research found that only 5% of enterprise AI pilots reached production with measurable value, and custom internal builds were among the most likely to fail. Off-the-shelf tools like ChatGPT, Microsoft Copilot, and Claude, configured with your context and data, beat bespoke builds on cost, speed, and maintenance for the large majority of business use cases. Building your own is justified only in a narrow set of situations. This post lays out what the data says, why custom builds age badly, what buying actually means in 2026, when building is genuinely the right call, and a framework for deciding.

What the Data Says

The most-cited number in enterprise AI this year comes from MIT's The GenAI Divide: State of AI in Business 2025, produced by its NANDA initiative. Across analysis of 300 public deployments, 150-plus executive interviews, and a large employee survey, it found that 95% of generative AI pilots delivered no measurable profit-and-loss impact. Only about 5% reached production with real value.

The reason is the part that should shape your build-versus-buy decision. MIT concluded the divide was not about model quality or regulation; it was about approach. The pilots that failed most often were custom-built systems that could not retain feedback or adapt, while generic tools like ChatGPT reached over 80% adoption for everyday tasks. Crucially, the successful 5% overwhelmingly got there through workflow integration and external vendor partnerships, not by building models in-house. The lesson is blunt: building your own is statistically the losing move for most organizations.

Takeaway: MIT's data shows custom builds are where most AI money dies. Success came from configuring and integrating off-the-shelf tools, not building bespoke ones.

Why Custom Builds Age Badly

There is a structural reason beyond the statistics. Frontier models improve every few months. A capability that required a custom engineering effort eighteen months ago is now often available out of the box in a $20-to-$30-per-seat subscription. That means a bespoke tool starts depreciating the moment it ships, and the maintenance burden of keeping it current competes directly with a vendor whose entire business is staying at the frontier.

We see this repeatedly in the field. A family office CFO told us they had tried and failed to build a custom platform to aggregate holdings and generate returns; the lesson was not to rebuild it, but to use configured off-the-shelf tools that had since caught up to what the custom build was meant to do. The pattern is common: companies invest heavily in a build, the models leapfrog it, and they are left maintaining something the market now offers for a subscription fee.

Takeaway: In a market where models improve every few months, a custom build is a depreciating asset from day one. The frontier will usually catch up to your bespoke feature faster than you can maintain it.

What "Buy" Actually Means in 2026

Buying does not mean handing everyone a chatbot and hoping. The real work, and the real value, is in configuration. The successful pattern is to take a capable off-the-shelf tool and shape it to your business: load projects with your context, save skills that encode your repeatable workflows, and connect your own documents and data so the tool answers from your reality rather than the open web.

This is the context layer, and it is the actual lever, not the underlying model. It is also what MIT found separates the 5% that succeed: deep workflow integration. A company that configures Copilot or Claude around its own documents, processes, and standards gets most of what a custom build promised, at a fraction of the cost and with none of the maintenance drag, because the vendor keeps the model current for you.

Takeaway: Buying well means configuring off-the-shelf tools with your context, data, and workflows. The configuration is the project; the model is a commodity that keeps improving on its own.

When Building Is Actually Justified

The 5% is real, and some companies belong in it. Building or heavily customizing is worth serious consideration when you have a proprietary, high-volume workflow that no off-the-shelf tool addresses and that runs at enough scale to justify the investment; when you have a genuine data moat that a custom system can exploit for durable advantage; when regulatory or security constraints require an architecture you fully control; or when you are automating a specific, repeatable agentic process with a clear and measurable return.

Even then, the honest sequence is to prove the value with configured off-the-shelf tools first, then build only the narrow piece that off-the-shelf genuinely cannot do. Most "we need to build" instincts, examined closely, turn out to be "we need to configure properly." The build decision should survive a hard look at whether a well-set-up subscription tool already does the job.

Takeaway: Build only when a proprietary, high-volume workflow, a real data moat, hard security constraints, or a clear agentic ROI justifies it, and even then, prove the value off-the-shelf first.

The Pricing Reality

The cost comparison is lopsided once you count everything. Off-the-shelf business tools are inexpensive per seat: Microsoft 365 Copilot runs roughly $18 to $30 USD per user per month depending on tier, as an add-on to a qualifying Microsoft 365 licence; ChatGPT's team tier sits around $25 USD per seat monthly; Claude's team tier around $20 USD per seat. One note for 2026 budgets: agentic products have moved to usage-based pricing, with Microsoft's Copilot Cowork and Anthropic's enterprise plans both shifting to a base fee plus metered consumption, so budget for usage on agent-heavy work rather than assuming flat seats.

Against that, a custom build carries development cost, ongoing maintenance, and the hidden cost of model churn, keeping your bespoke tool current as the frontier moves. For the vast majority of use cases, the total cost of ownership of a build dwarfs the subscription, and the build is often behind the subscription within a year. The math only flips in the narrow situations above.

Takeaway: Per-seat subscriptions are cheap; custom builds carry development, maintenance, and model-churn costs that usually exceed them by a wide margin. Budget for usage-based pricing on agentic work.

A Framework for Deciding

At Solway we run this decision through our 4 Questions Framework, Strategy, Training, Build, Maintain, and our Opportunity and Risk Matrix, which sorts every AI idea into Quick Wins, Quality Lifts, Strategic Upgrades, and Not Yet. In practice, most "build a custom tool" ideas land in Not Yet or resolve into a Quick Win once you realize a configured off-the-shelf tool already does it. The genuine build candidates are the rare Strategic Upgrades that survive the filter.

The practical path for most companies is to audit where the real friction is, pilot a configured off-the-shelf tool against it, and only escalate to a build for the narrow slice that off-the-shelf cannot reach. This is exactly what Solway's AI Clarity Sprint is designed to produce: a clear-eyed opportunity map that tells you what to configure now and what, if anything, genuinely warrants building. Shaheer Tariq, Co-Founder of Solway, puts the rule plainly: configure first, and build only the narrow piece that configuration cannot reach. For Alberta companies, the training to get a team fluent on configured tools is eligible for the Canada-Alberta Productivity Grant, which reimburses 50% of eligible third-party training costs.

Takeaway: Run the decision through a structured filter before spending on a build. Most build ideas dissolve into configuration once examined; the few that survive are the ones worth the investment.

Frequently Asked Questions

Should my company build a custom AI tool or use off-the-shelf?

For most mid-size companies, use and configure off-the-shelf tools. MIT's 2025 research found only 5% of AI pilots reached production value, with custom builds among the most likely to fail. Building is justified only for proprietary, high-volume workflows, genuine data moats, hard security constraints, or clear agentic ROI, and even then only after proving value off-the-shelf first.

Do we need to train our own AI model?

Almost never. Frontier models from providers like OpenAI, Anthropic, and Microsoft offer strong intelligence out of the box, and they improve every few months. The value comes from configuring these models with your context and data, not from training your own, which is expensive, slow, and quickly outdated.

How much does a custom AI build cost versus a subscription?

Off-the-shelf business tools run roughly $18 to $30 USD per user per month. A custom build carries development cost, ongoing maintenance, and the cost of keeping it current as models advance. For most use cases the build's total cost of ownership far exceeds the subscription, and the build is often behind the subscription within a year.

What does it mean to configure an off-the-shelf tool?

It means loading projects with your business context, saving skills that encode your repeatable workflows, and connecting your own documents and data so the tool answers from your reality rather than the open web. This configuration is what MIT found separates the AI efforts that succeed from those that stall.

When is building genuinely worth it?

When you have a proprietary, high-volume workflow no off-the-shelf tool addresses, a real data moat, regulatory or security needs requiring full control of the architecture, or a specific agentic process with clear measurable ROI. Prove the value with configured tools first, then build only the narrow piece off-the-shelf cannot do.

Is our data safe if we use ChatGPT or Copilot instead of building?

Yes, with enterprise-grade tiers configured correctly. Use business or enterprise plans, disable training on your data, and set clear policies on what can be entered where. Microsoft Copilot in particular keeps answers grounded within your organization's data boundary. A short AI policy engagement settles these choices cleanly.

How do we decide without wasting money?

Audit where the real friction is, pilot a configured off-the-shelf tool against it, and escalate to a build only for what off-the-shelf cannot reach. Solway's AI Clarity Sprint produces exactly this opportunity map, and for Alberta employers the associated training is CAPG-eligible. Contact Solway to scope it.