Buy AI Tools, Use ChatGPT, or Build Agents? How to Decide

Buy versus build is not one decision. It is a decision you make per workflow — and companies that pick a side once, company-wide, get the worst of both. Most of what you should do is buy. A layer of what you should do is assist. A small, valuable slice is worth building. The strategy is the sort.

When should you buy a point AI tool?

Buy when the workflow is a commodity — when your version of the problem looks like everyone else’s. Contract review, meeting notes, support deflection, expense coding: a vendor solving one job across a thousand companies has seen more edge cases than your team ever will. The success data backs this up. MIT’s GenAI Divide study found purchased, specialized AI tools succeed about 67% of the time, while internal builds succeed roughly a third as often. The same research found roughly 95% of enterprise generative AI pilots deliver no measurable P&L impact — and the deployments that worked disproportionately came from specialized vendors, not internal builds.

The point-tool market is not a transitional phase, either. Even as coding agents make custom software cheaper to produce, a16z’s Notes on AI Apps in 2026 argues third-party apps stay vital — the hard problem has shifted from how to build to what to build. Vendors who already answered the “what” keep their edge.

One warning before you swipe the card. Every point tool is a new contract, a new data surface, and a new invoice nobody owns. Per KPMG’s Global AI Pulse, reported by ITPro, 42% of companies have only partial visibility into their AI spend, and a third say token-based pricing confuses them enough to stall agent deployment. Buy deliberately, keep a kill list, and read our breakdown of controlling token, seat, and tool-sprawl costs before the line items compound.

When is ChatGPT, Claude, or Copilot enough?

A general assistant is the floor, not the strategy. Every employee should have one. The drafting, research, and analysis gains are immediate, and governed enterprise seats beat the shadow accounts your people are already using. Buy the seats.

Then be honest about what you bought. An assistant makes each person faster inside the existing process; it does not change the process. MIT’s researchers named the mechanism a learning gap: generic tools do not adapt to your workflows, so value stalls at scattered personal productivity. McKinsey’s State of AI survey shows where that ceiling sits — 88% of organizations now use AI, but only 39% report any EBIT impact at the enterprise level. Adoption is nearly universal. Value is not. Per McKinsey, the gap closes through workflow redesign, not more licenses.

If your AI rollout is a license count and a lunch-and-learn, that’s not a strategy — that’s hoping. Assistants raise the floor. Something else has to raise the ceiling.

When should you build internal agents?

Build when the workflow is your moat — when the way you do the work is the reason customers pick you. Underwriting judgment. Merchandising calls. The service playbook competitors keep trying to copy. No vendor has seen your edge cases, and handing a differentiating workflow to a point tool puts your moat on someone else’s roadmap.

A built agent earns its keep by carrying what generic tools cannot: your context, your permissions, your data plumbing, your evals. That is also why it is harder than the demo suggests. An agent is a thin model call wrapped in a thick operating layer — workflow tracing, guardrails, measurement that tells you whether the output actually worked. Skip the layer and you have shipped a demo with an API bill.

And here is the prerequisite almost everyone skips: someone must own the workflow before anyone automates it. Not the tool — the workflow. If you cannot name the person whose numbers change when the agent ships, you are not ready to build it.

Why do internal builds fail so often?

Because companies build the agent and skip everything around it. The pattern repeats: a capable team ships an impressive prototype, nobody redesigns the process around it, no evals gate its output, adoption stays voluntary, and six months later it is a Slack channel nobody opens. The code was rarely the failure. The operating layer was.

Here is our position, and you can disagree with it: most companies should not build their first agents alone. Not because your engineers are not smart — because agent engineering is a young discipline with unfamiliar failure modes. Context rot. Silent regressions. Evals that drift away from the business outcome they were supposed to protect. The fastest way to learn a discipline like that is next to people who ship in it weekly, inside your stack, on your workflows — not from a vendor deck, and not from a six-month internal science project.

Is there a third path between buying and building?

Yes: build with embedded partners, then take the keys. Tenex cofounder Alex Lieberman has argued that engineering-as-a-service is a massive need — his estimate, north of a trillion dollars a year — precisely because most companies want build-grade fit without staffing a permanent agent team on day one. The trade is straightforward: outside builders bring the discipline, your team keeps the ownership.

That is the model our AI engineering practice runs. The builders are the 0.16% — 50 hired from 32,100 applicants, every one through a paid build trial. No leetcode theater. You build. We watch. And the work sits close to the labs: Anthropic and OpenAI are formal partners, and Anthropic named Tenex on the Claude for Small Business launch. We stay model-agnostic anyway — the partnership serves your result, not a logo.

How would Tenex run this decision inside your company?

In four moves — and the last one is leaving.

Embed. We sit inside your teams and trace the actual workflows, not the org chart. Each workflow gets sorted: buy it, assist it, or build it. Most land in the first two buckets, and that is the point — the sort protects your build capacity for the workflows that deserve it.

Ship. For the build bucket: first production system by week four, median. Production means evals gating output and a number the system is supposed to move — not a pilot collecting compliments.

Train. Pairing with your operators inside the new workflow, not a webinar. The bar we hold: 90%+ operator adoption, defined as weekly active at handoff — tracked, not hoped. (Illustrative until audited.)

Hand off. Repos, evals, agents, dashboards — transferred. Our access revoked. Dependency is a design flaw; if you still need us to run the system, we built the wrong system.

The buy-assist-build sort is one slice of a bigger rebuild — governance, cost control, adoption — that our AI transformation work installs around it. The sort tells you what to do with each workflow. The operating layer decides whether any of it survives contact with Monday morning.

When should you skip building entirely?

Skip building if you cannot name the owner, fund the maintenance, or state the number the agent moves. More bluntly:

• Do not build a commodity workflow out of pride. If five vendors already solve it, your in-house version is a liability with a payroll.
• Do not build without evals. An agent nobody measures is not automation; it is unaudited risk with good manners.
• Do not build to dodge the adoption problem. A built agent operators ignore dies exactly like a point tool nobody logs into — we wrote about moving pilots into everyday operating processes because the graveyard is the same either way.
• Do not build six things at once. One workflow, in production, measured. Then the next.

The reverse failure costs just as much: buying a point tool for your moat workflow, then discovering your differentiation now ships on a vendor’s release schedule. The sort has to be honest in both directions. We publish calls like this every Tuesday in ultrathink, our newsletter for executives and builders who want the tactical version.

Here is the challenge. Name the one workflow you would defend as your moat, then check what is running it today — a point tool, an assistant seat, or nothing. If the answer is nothing, bring it to us. We read every message. When it’s a fit, a partner reaches out within a couple of days.