Build vs Buy AI Agents: A Weighted Scorecard and 3-Year TCO Model

Buy an AI agent when the workflow is common, lives mostly inside one platform you already pay for, and volume stays moderate. Build when the workflow is how you compete, crosses several systems, or carries data a vendor cannot hold. Most enterprises land in between: buy the platform, build the workflows that matter.

This build vs buy AI agents framework gives you a weighted scorecard with thresholds and a 3-year cost model using current list prices. Score first. Then price.

The four real options, not two

“Build vs buy” hides four paths. Each puts ownership in a different place, and ownership is what you pay for later.

Buy. You license a platform agent: Salesforce Agentforce, Microsoft Copilot Studio, ServiceNow, Intercom Fin, or a vertical vendor. The vendor owns the orchestration (the loop that decides which step runs next), the model, and the logs. You own configuration.

Hybrid A: buy the platform, build the workflows. You license the platform and write custom actions for the parts that make your process yours. The vendor still owns the runtime and the meter.

Hybrid B: partner-build. An outside team builds a custom agent on your cloud and model accounts, then hands over code, prompts, and test sets. You own the asset and rent the skills for a while.

Build. Your team writes the agent against model APIs. You own everything, including every model migration and outage.

The core mechanism: buying turns engineering cost into a usage fee, and building does the reverse. Which is cheaper depends on volume, the billing meter, and the people it takes to keep the agent healthy. If your question is who builds, not whether to build, read AI agency vs in-house team.

What does the evidence say about buying vs building?

The most cited data point is MIT NANDA’s The GenAI Divide (July 2025, research period January to June 2025). In its chart, strategic partnerships (“buy”) reached 66% and internal builds 33%, with success defined as deployment beyond the pilot phase with measurable KPIs. The text says pilots built through strategic partnerships were “2x as likely to reach full deployment” as internal ones.

Read the caveats before you quote it. MIT’s “buy” bucket means “procure external tools, co-develop with vendors,” so it already includes some hybrid work. The pure hybrid bucket had “insufficient data to quantify.” The percentages come from 52 organizations, and success definitions varied. The authors warn the gap “may reflect organizational capabilities rather than implementation approach alone.” It is a correlation, not proof that buying causes success.

Gartner’s June 2025 forecast predicts over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, or inadequate risk controls. It also estimates only about 130 of the thousands of agentic AI vendors are real. Many others are rebranded chatbots and RPA, which Gartner calls “agent washing.”

My read: use MIT as a prior that favors buying or partnering for a first agent. Then let the scorecard below overrule it.

Which variables decide build vs buy for AI agents?

Eight variables decide it. Five are about the work, covered here. Pricing, exit, and total cost get their own sections.

Workflow differentiation. Would a competitor running the same vendor agent erase your edge? Order status is not an edge. Underwriting or pricing logic built from your history might be. Renting your edge puts it on someone else’s roadmap.

Integration depth. Count the systems the agent writes to. Platform agents are strongest inside their own data. Each write outside that boundary needs its own connector, permissions, and failure handling. My threshold: three or more external write targets pushes you to Hybrid A or further. Copilot Studio vs custom agents shows where that boundary sits for Microsoft.

Data sensitivity and residency. Buy, and you inherit the vendor’s regions, retention, and subprocessors. Build, and you choose them and pay for it. Anthropic charges a 1.1x multiplier on all token prices for US-only inference on Claude 4.6 and later. Data sovereignty for AI agents lists the questions to ask.

Time to value. MIT found mid-market top performers averaged 90 days from pilot to full implementation. Enterprises took nine months or longer. A custom build adds hiring before that clock starts.

Team capability. Building means owning model churn. Anthropic gives at least 60 days’ notice before retiring a publicly released model, and calls to retired models fail. Without a test set to rerun against the replacement, 60 days is tight. No test set means you are not ready to build. Start with agent acceptance tests.

How do agent pricing models change the math?

Every meter rewards one agent design and punishes another. These are list prices I checked on October 15, 2026.

Read the fine print. Salesforce’s usage examples exclude Data 360 credits and other consumption services. Intercom bills an “assumed resolution” when a customer leaves after Fin’s last answer without asking for more help, even if they gave up.

I could not find published list prices for ServiceNow’s agent tiers, so get a written quote with the meter defined. My rule: model every quote at 1x, 2x, and 4x expected volume, and at double the actions per task. If any case costs more than a built agent, negotiate a cap. AI agent pricing models covers quote language.

Lock-in and exit rights: what the law covers and what it does not

In the EU, the Data Act has applied since September 12, 2025, and it covers switching between data processing services. Article 29 says that from January 12, 2027, providers “shall not impose any switching charges” for the switching process. Article 25 caps the notice period at two months. The transition is normally 30 calendar days. If that is technically infeasible, the provider must justify it within 14 working days and may set up to seven months. The customer can also extend the transition once.

Now the limits. The no-charge rule does not apply where most main features were custom-built for one customer and are not offered at broad commercial scale. A Digital Omnibus proposal would add transitional relief. As Greenberg Traurig summarized in July 2026, the Council’s compromise text would exempt adapted standard services and SME and small mid-cap providers of non-infrastructure services, but only for contracts concluded on or before September 12, 2025. Infrastructure providers would not benefit, and the Article 29 switching-fee regime would still apply. The text would also expressly permit proportionate early termination penalties. It was still under negotiation.

Exportable data does not mean a portable agent. Prompts, topics, action definitions, and test sets in a proprietary format may export as files nothing else can run. My rule: before you sign, ask for a sample export of one working agent and try to read it. AI agent ownership and exit rights lists the clauses to request. This is not legal advice.

The weighted scorecard (copy this)

Score each criterion from 1 (favors buying) to 5 (favors building). Multiply by the weight, add, and divide by 100. The weights are my defaults. Change them before you score, not after.

Use 2 and 4 in between. For TCO, 2 means 1.1 to under 1.3, and 4 means 0.7 to under 0.9.

Thresholds. Below 2.3, buy. From 2.3 up to but not including 3.5, go hybrid: Hybrid A when one platform is the system of record for most of the workflow, Hybrid B when no platform dominates. At 3.5 or above, build. If team capability scored 1 or 2, build through a partner with a handover clause.

Knockout rules override the score. If data cannot leave your environment and the vendor cannot contractually run where it must, buying is off the table. If nobody can own evaluations, model migrations, and incidents for three years, building is off the table. If the vendor will not show you a working export, treat buying as blocked until it does.

A 3-year TCO worked example

Everything here is hypothetical except list prices and BLS figures. A company runs service on Salesforce. An agent handles 10,000 customer conversations a month in year 1, 15,000 in year 2, and 20,000 in year 3. It averages 4 actions per conversation and writes to one ERP outside Salesforce. Existing Salesforce licenses are the same in every column, so they are excluded.

Labor rate. BLS puts the median software developer wage at $135,980 (May 2025). Wages were 68.5% of employer costs for full-time private industry workers in June 2026. That ratio is economy-wide, so this is approximate: $135,980 ÷ 0.685 = about $198,500 per loaded FTE.

Buy path, Agentforce Flex Credits. 4 actions x 20 credits = 80 credits, or $0.40 per conversation. Implementation partner: $80,000 one-time (assumed). Admin: 0.3 FTE.

Build path, Claude Sonnet 5.5. 4 model calls per conversation, each 8,000 input and 600 output tokens, half the input cached at $0.10 per million. That is $0.032 + $0.0016 + $0.024 = $0.0576 per conversation, ignoring cache-write charges. Engineers: 1.5 FTE in year 1, 0.75 FTE after. Hosting, logging, and eval tooling: $2,000 a month (assumed).

Hybrid A. The buy path, plus 0.5 FTE in year 1 and 0.25 FTE after, to build custom actions into the ERP.

By year, credits are $48,000, $72,000, and $96,000; tokens are $6,912, $10,368, and $13,824.

The build-to-buy ratio is 1.47, so TCO scores a 1. On cost alone, buy wins. Now put the same agent on the $2 per-conversation meter. Usage becomes $1,080,000, the buy total $1,338,650, and the ratio 0.52. Same agent, same volume, opposite answer.

Break-even. At steady state, build carries $113,325 a year more fixed cost ($148,875 + $24,000 - $59,550). On Flex Credits it saves $0.3424 per conversation, so break-even is about 27,600 conversations a month. On the per-conversation meter it saves $1.9424, and break-even falls to about 4,900 a month. At 8 actions per conversation with double the tokens, the Flex break-even drops to about 13,800.

Scored, the example comes out: differentiation 2, integration 3, data 2, pricing 2, lock-in 3, time 2, team 2, TCO 1. The weighted total is 2.15, so buy. On the per-conversation meter, volume sits over 2x break-even, so pricing becomes 5 and TCO 5. The total rises to 2.85: hybrid. At this volume, engineers dominate build cost, not tokens. The real cost of building AI agents in-house covers where that time goes.

How build vs buy decisions go wrong, and how to spot it early

Each failure shows up in a number before a budget review.

Credit creep on a bought agent. Action meters scale with tool calls. A prompt change that adds one lookup to a 4-action conversation raises cost by 25%. Copilot Studio has a sharper edge: Microsoft’s billing page says general overage enforcement disables custom agents at 125% of capacity unless a pay-as-you-go meter covers the overage. Detect it with a weekly chart of credits per conversation and an alert at 80% of capacity.

Outcome inflation. On per-outcome pricing, a customer who leaves without asking for more help can count as a resolution. Each month, sample 50 billed outcomes and check them for reopens or repeat contacts within 7 days. If more than 1 in 10 look like abandonment, renegotiate the definition.

Agent washing. Ask the vendor to run your 20 hardest real cases live, with tool calls visible. A scripted demo proves nothing.

The stalled build. Version one ships, then a model retirement notice arrives and there is no test set to check the replacement. Detect it in month two: if nobody can rerun the acceptance tests in under a day, you have built a liability.

The platform that became the product. Hybrid A drifts when custom actions grow until the platform is an expensive host. If your custom code outgrows the configuration, rescore.

Bottom Line

Score the workflow before you compare vendors. Differentiation, integration depth, and data rules decide most cases.

Then price all three paths over three years, on the meter you will actually be billed on. The meter can flip the answer.

Next step: score one workflow with its owner in the room, and calculate your break-even volume before your first vendor call.

FAQ

Can I mix Agentforce per-conversation and Flex Credit pricing?

Not in the same org, per Salesforce’s pricing FAQ. You can move to Flex Credits, but you must swap all existing conversation SKUs. Unused Flex Credits do not roll over into later subscription terms, so an oversized prepurchase is lost money.

Do Copilot Studio agents cost extra for people who already have Microsoft 365 Copilot?

Only employee-facing use is free. Microsoft’s billing page lists no charge when the user holds a Microsoft 365 Copilot license ($30 per user per month, paid yearly) and the agent runs under that user’s authenticated identity. Customer-facing agents, unlicensed staff, and agent flows started by other triggers draw credits.

Do token estimates from older benchmarks still hold?

Not exactly. Anthropic says Claude 4.7 and later models use a newer tokenizer that produces about 30% more tokens for the same text, depending on content. If your cost model came from a pre-4.7 model’s logs, add about 30% to token counts before applying new prices.

Does the MIT 95% failure figure apply to AI agents specifically?

No. The report covers generative AI broadly and says 95% of organizations are getting zero return. It drew on 52 organization interviews, 153 survey responses from senior leaders, and over 300 public initiatives. A task-specific tool counted as successful when users or executives remarked on a marked and sustained productivity or P&L impact. Treat it as a base rate, not a forecast.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect