Build vs. buy AI agents: when to build, buy or forego

What are the three answers to build vs. buy?

Every AI opportunity has three possible answers. The usual build-vs.-buy debate covers two of them and skips the question that comes first: does this work need AI at all?

• Build. You build and run an agent. You own its logic, its tests and its upkeep.
• Buy. You license a product that already does the job. The vendor owns the product. You still own the integration, the checks and the result.
• Forego. You decide AI isn’t worth it for this work, for now. You keep a person on it, run it as plain code, or delete the step.

Many real choices sit between build and buy, such as a vendor’s product tuned to your rules, or an agent a partner builds with you. The useful test is who keeps it working. If your team will own the upkeep, plan for it as a build.

Foregoing can still change the work. Deleting a step or replacing it with a simple rule improves the process; it just isn’t an AI project. If you have sorted a workflow’s steps into the four buckets, three of them (delete, plain code and human) are forms of forego. Build vs. buy only applies to the agent steps.

Make the call one workflow at a time. The same company can build one agent, buy one tool and keep two jobs human, as in the worked example below.

Is it better to build or buy AI? What the evidence says

The evidence leans toward buying for common work. It is also thinner than the headlines suggest.

Start with MIT NANDA’s report, “The GenAI Divide” (July 2025). It says 95% of organizations are getting “zero return” from generative AI, and that just 5% of integrated AI pilots are extracting millions in value while the vast majority show no measurable P&L impact. On build vs. buy, it found that partnerships reached deployment about twice as often as internal builds.

Check how it was measured before you repeat it. The report presents itself as preliminary findings. Its numbers come from 52 interviews, 153 survey responses collected at four industry conferences, and a review of more than 300 public AI initiatives. Success meant moving beyond the pilot with measurable KPIs. Return was measured six months after the pilot, and outcomes were self-reported. The authors caution that the build-vs.-buy gap may reflect the organizations rather than the approach.

Others went further. Wharton professor Kevin Werbach said he could find no support in the report for the 95% figure, and his colleague Ethan Mollick wrote that the methodology “doesn’t allow general conclusions about failure rate.”

Menlo Ventures’ numbers come from larger surveys: 600 U.S. IT decision-makers in 2024 and 495 enterprise AI decision-makers in 2025. Menlo is a venture firm that invests in AI companies. Gartner adds that many use cases sold as agentic don’t require agents.

The fair reading: for common work, buying tends to beat building, and many AI projects shouldn’t start. None of these sources can tell you which answer fits a given workflow. A test on your own cases can.

How do you decide whether to build, buy or forego?

Take one workflow at a time and ask five questions, in order. The first answer that fits decides the call.

Buy comes before build because, for common work, a vendor has already paid for the engineering and spreads the upkeep across many customers. Build when the workflow is yours and you can keep it working.

Two of the questions can only be answered with evidence: whether people still beat AI on your cases, and whether a tool passes on them. One test answers both. A personalized benchmark runs people, an agent alone and an agent with a reviewer on the same past cases, scored by your team’s own definition of correct. Put the vendor’s product through the same test before you sign. See how to benchmark AI on your own work.

Which criteria should drive the decision?

Eight criteria do most of the work. Few workflows point one way on all eight, so weigh them together.

Five of them are easy to get wrong:

• Mature tools. A demo on the vendor’s data proves little. Run the tool on your own past cases, awkward ones included, and keep those cases for later.
• Data and integration. An agent can only judge what it can see. If the work depends on systems or context a tool can’t reach, build the access or forego for now.
• Rate of change. If the rules change often in ways only you know, a product built for everyone will lag behind. If they change the same way for everyone, as with formats or regulations, a vendor spreads that work across its customers.
• Evals and monitoring. Building means owning evals and monitoring for as long as the agent runs. If no one will own them, don’t build.
• Lock-in. Ask what leaving would cost in two years. Whichever path you take, keep your test cases, rules and data where you can take them with you.

What does it really cost to build or buy an AI agent?

Price each path over the same period, and per task done correctly. The license fee or the model bill is only the part everyone sees.

Building costs engineering up front, then model usage, hosting and monitoring, plus re-running your evals each time a model, a prompt or the process changes. The upkeep is easy to underestimate. Google engineers warned in 2015 that machine-learning systems are relatively fast and cheap to ship, but “maintaining them over time is difficult and expensive.” And when the builders leave, the know-how can leave with them, a form of key-person risk.

Buying moves engineering, models and hosting into the vendor’s fee. Much of the rest stays with you: integration, training and change management, review time, your own testing and, if you leave, switching. In Menlo Ventures’ 2024 survey, implementation costs were cited in 26% of failed pilots, while just 1% of leaders named price as a selection concern.

Either way, the model is the smaller part. EY estimates that tokens account for roughly a third of what an agent costs to run. The rest goes to subscriptions, infrastructure, governance, organizational change, expected failure and regulation.

Foregoing costs what the work costs today, plus whatever it takes to delete steps or write plain code. That is the baseline the other two paths must beat. Divide each path’s total by the tasks it will handle; cost per task makes all three comparable.

When is foregoing AI the right answer?

Forego when any of these is true:

• The step shouldn’t exist. Waits, re-keying and chasing are symptoms of a broken process, so fix the process. Michael Hammer’s advice from 1990 still holds: “Don’t automate, obliterate.”
• Rules can do it. If a step follows fixed rules, plain code is cheaper, faster and predictable. Even Anthropic, which sells the models many agents run on, advises starting with the simplest solution, which “might mean not building agentic systems at all.”
• Volume is low. A handful of cases a month can’t repay a build or a license, however well the agent performs.
• People still win. If people beat every AI setup on your own cases, keep them on the work. Expect this where judgment is high and mistakes are costly.
• The AI can’t see what people see. When a decision depends on context an agent can’t reach, forego until it can.

Forego isn’t forever. Revisit the call when volume grows, a tool matures, a model improves or the missing data becomes available. Keep the cases you tested on, so the next check takes days, not months.

Worked example: what to build, buy or forego

Here is the method applied to four workflows at the illustrative company from Nolli’s product demo. Each was benchmarked three ways on the same past tasks: people, an agent alone, and an agent with a reviewer.

• Ticket triage: build. An agent alone routed 97.4% of tickets correctly, against 93.5% for people, and a reviewer added cost without adding accuracy. Reading a messy ticket takes judgment, but the routes are the company’s own, a wrong route is quick to fix, and 3,700 tickets a month leave plenty of history to test against. On a small model, the agent still routed 97.1% correctly at a fraction of the cost, so it is cheap to run and easy to keep tested. Tier-2 cases stay with people.
• Invoice matching: buy. An agent with a reviewer matched 99.4% of invoices correctly, against 96.2% for people. Invoices arrive in many vendor formats, so rules alone fall short. But matching them to purchase orders works much the same everywhere, and mature accounts-payable tools already handle this company’s formats. The company buys one and keeps a reviewer on the exceptions.
• NDA redlines: forego. Counsel’s own redlines passed review 96.8% of the time, 2.6 percentage points ahead of an agent with a reviewer. There are about six a month, and each one turns on counsel’s judgment about non-standard clauses.
• Refund decisions: forego. People’s decisions were upheld on review 98.1% of the time; an agent working alone, 72.4%. Each decision weighs customer history the agent can’t see. OpenAI’s guide to building agents lists refund approval among the workflows to prioritize for agents, which is exactly why you test on your own work.

Triage and invoice matching take 210 and 120 hours of people’s time a month, worth about $166,000 and $95,000 a year at $66 an hour. That is where this company’s AI budget should go. NDAs and refunds get no new spend.

Sources

1. Aditya Challapally, Chris Pease, Ramesh Raskar and Pradyumna Chari. “The GenAI Divide: State of AI in Business 2025.” MIT NANDA, via a copy hosted by MLQ.ai, July 2025.
2. Tim Tully, Joff Redfern, Deedy Das and Derek Xiao. “2025: The State of Generative AI in the Enterprise.” Menlo Ventures, December 9, 2025.
3. “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027.” Gartner, June 25, 2025.
4. Kevin Werbach. “Post questioning the MIT NANDA report’s 95% figure.” LinkedIn, August 23, 2025.
5. Ethan Mollick. “Post on the methodology of the MIT study of AI pilots.” LinkedIn, September 2025.
6. Tim Tully, Joff Redfern and Derek Xiao. “2024: The State of Generative AI in the Enterprise.” Menlo Ventures, November 20, 2024.
7. Shawn Smith, Steve Wanner, Hyong Kim, Sachin Lulla, Dan Diasio, Andrew Phillips and Gregory Daco. “Agentic AI ROI: Can AI pay for itself?” EY, September 21, 2026.
8. D. Sculley, Gary Holt, Daniel Golovin and seven co-authors at Google. “Hidden Technical Debt in Machine Learning Systems.” Advances in Neural Information Processing Systems 28 (NIPS 2015), 2015.
9. Michael Hammer. “Reengineering Work: Don’t Automate, Obliterate.” Harvard Business Review, July–August 1990.
10. Erik S. and Barry Zhang. “Building effective agents.” Anthropic, December 19, 2024.
11. “A practical guide to building agents.” OpenAI.