How to Decide Whether to Build or Buy a Marketing Agent Platform
A build vs buy framework for AI marketing agents: billing units from vendor pricing pages, upkeep, data control, a worked cost comparison and a platform scorecard.
On this page
The build vs buy decision for an AI marketing agent platform in 2026 turns on four things you can check before any demo: the unit each route bills in, who carries the upkeep when models change, where prompts and outputs are stored, and how much of the work you could move to another vendor. This piece prices both routes from published pages, read on 7 October 2026, and works one comparison under labeled assumptions. In that example buying wins at low run volumes, building wins once volume climbs, and the crossover shifts a long way with two inputs the pricing pages we read do not give you. For the definitions underneath, start with our explainer on what agentic AI is.
What does each route put on the invoice?
Building AI agents for a marketing platform on an LLM API and buying an agent platform in 2026 put different units on the invoice, and the units matter more than the headline prices.
| Route | Publisher | Billing unit | Published price |
|---|---|---|---|
| Build on a model API | Anthropic | Tokens | Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, with 50% off both through the Batch API |
| Build on a hosted agent runtime | Anthropic | Tokens plus runtime | Claude Managed Agents, in beta, at $0.08 per session-hour of running time on top of standard token rates, a rate published on the pricing page |
| Tracing for a custom build | LangChain | Seats and traces | LangSmith Plus at $39 per seat per month, with 10k base traces a month included |
| Buy a workflow orchestrator | n8n | Workflow executions | Pro at 50€ a month for 10K executions and Business at 667€ a month for 40K, both billed annually |
| Buy an agent platform | Salesforce | Actions, paid in Flex Credits | $500 per 100k credits, with Agentforce actions at 20 credits each and Agentforce Voice actions at 30, or $2 per conversation |
| Buy an agent platform | Microsoft | Copilot Credits | Copilot Studio at $200 per pack of 25,000 credits per month |
Salesforce's rate works out to $0.10 an action, and its page adds: "Unused Flex Credits do not roll over into subsequent subscription terms." Microsoft says a varying number of Copilot Credits is billed for each completed action or response, so the pack price on its own does not give a cost per run. n8n defines an execution as a single run of your entire workflow, however many steps it has. One agent run could therefore be a single execution on one platform and five actions on another, which makes units consumed per run the first number to pull from any pilot. Our per-contact cost model goes further on the token side.
What has changed for builders since 2025?
Build vs buy thinking for AI agents on a marketing platform that dates from 2025 needs one correction for 2026: the model underneath is a moving part. Anthropic lists seven models retired from the Claude API between 5 January and 5 August 2026, on five separate dates, and it deprecated Claude Sonnet 4.5 on 30 September 2026 with retirement set for 30 November. The policy promises at least 60 days' notice before a publicly released model is retired, and the page is plain about what follows: "Requests to retired models will fail."
For an in-house agent built on one of those models, its retirement date means rerunning an evaluation set against the replacement model and repairing the prompts that regress. A migration can move the budget as well. The pricing page says Claude 4.7 and later models use a newer tokenizer and adds: "This tokenizer produces approximately 30% more tokens for the same text." It also lists Claude Sonnet 5.5 at $2 and $10 per million input and output tokens against $3 and $15 for Claude Sonnet 4.6, so cost per run has to be measured again after each move. When you buy, much of that retesting can pass to the vendor, and the thing to find out is how you will be told that the model under your agents has changed.
Four tests for building in-house
A build vs buy decision framework for LLM agents run in-house can be reduced to four tests, and a team should pass all four before it builds.
| Test | Points to build | Points to buy |
|---|---|---|
| Volume | Monthly runs sit above your break-even | Runs are below it, or nobody has counted them |
| Ownership | A named engineer has standing hours for evaluations and migrations | Upkeep would land on marketing operations |
| Data | A retention or residency rule needs terms you negotiate directly with the model provider | The platform's documented retention already satisfies the rule |
| Fit | The workflow depends on your own scoring logic or data model | The agent mostly reads and writes one system the vendor already owns |
The data test deserves a closer read, because control is set feature by feature. Anthropic documents a zero data retention (ZDR) option, requested through its sales team and enabled per organization, and describes it this way: "Under a ZDR arrangement, Anthropic does not store customer prompts or responses at rest after the API response is returned." The same page marks batch processing as ineligible, with 29-day retention, and says Claude Managed Agents session transcripts persist until you delete them. Sending a job through the Batch API therefore buys the 50% token discount and gives up the zero retention terms for that data. For the fit test, see our piece on what agents can write back to your CRM.
A worked comparison at 4,000 runs a month
To put numbers on build vs buy for AI agents on a marketing platform in 2026, we priced one hypothetical team both ways. Run volume, tokens per run, actions per run, two tracing seats with no trace overage, hours, loaded hourly rates and the 24-month spread are illustrative assumptions. Only the unit prices come from the published pages above.
A B2B team wants agents for three jobs: inbound lead research briefs, pre-launch campaign checks and a weekly performance summary, 4,000 runs a month in all. Each run makes six model calls averaging 5,000 input and 700 output tokens, so 30,000 input and 4,200 output tokens, which costs $0.102 on Claude Sonnet 5.5. On a platform billed per action, the same run is assumed to consume five actions, or $0.50 at the Flex Credit rate. Licenses for the CRM under that platform are treated as already owned.
| Monthly line | Build on the API | Buy per action |
|---|---|---|
| Usage | 4,000 runs × $0.102 = $408 | 20,000 actions × $0.10 = $2,000 |
| Tracing | 2 LangSmith Plus seats = $78 | None assumed |
| Initial work, spread over 24 months | 360 hours × $100 = $36,000, or $1,500 a month | 80 hours × $75 = $6,000, or $250 a month |
| Upkeep | 30 hours × $100 = $3,000 | 10 hours × $75 = $750 |
| Total | $4,986 | $3,000 |
The $100 engineering rate is our assumption. The Bureau of Labor Statistics reports 2025 median pay of $64.44 an hour for software developers, quality assurance analysts and testers, and that figure is pay alone.
Buying wins this case by $1,986 a month. Fixed costs are $4,578 for the build and $1,000 for the platform, and each run costs $0.398 less on the API, so the lines cross at $3,578 ÷ $0.398, or about 8,990 runs a month. At 10,000 runs the build comes to $5,598 against $6,000.
Two assumptions carry most of that answer. If a run consumes three actions where we assumed five, the break-even moves out to about 18,070 runs. If upkeep takes 15 hours where we assumed 30, it drops to about 5,220. Token tuning barely registers by comparison: sending every run through the Batch API would save $204 a month, and batch jobs sit outside zero data retention. Put your own hours and rates into the AI ROI calculator before you trust ours.
How do you evaluate an orchestration platform once buying wins?
To evaluate an AI agent orchestration platform in 2026, after buy has beaten build on the numbers, read its documentation before you sit through its demo. NIST gives procurement a starting list in its Generative AI Profile (NIST AI 600-1, July 2024). Suggested action GV-6.1-009 extends vendor due diligence to intellectual property, data privacy, security and other risks. GV-6.2-007 lists what to request in vendor contracts, including notification of serious incidents and service level agreements that address incident response, response times and availability of critical support. GV-6.2-006 covers testing rollover and fallback, which the profile acknowledges may include manual processing.
Six criteria to score
The AI agent orchestration platform evaluation criteria we would score in 2026 are the six below, each tied to something a vendor either documents or leaves out. Give a row 0 when the documentation is silent, 1 when the item is disclosed with conditions such as a beta label or a higher plan, and 2 when the disclosed terms meet your requirement.
| Criterion | What the documentation should state | What we read on the pages above |
|---|---|---|
| Billing unit | The unit, its price, how many one run consumes and whether unused units expire | Salesforce states 20 Flex Credits per action and no rollover of unused credits; Microsoft says the credits billed vary with usage |
| Model lifecycle | Notice before a model is withdrawn | Anthropic commits to at least 60 days for publicly released models |
| Retention and training | Retention per feature and any training on your data | LangChain says LangSmith does not use your data to train models; Anthropic documents retention feature by feature |
| Hosting | Regions and any self-hosted option | LangSmith lists US or EU cloud, with hybrid and self-hosted options on Enterprise; n8n lists its Business plan as self-hosted |
| Portability | Open protocols for tools and exportable definitions | n8n Business includes version control using Git |
| Contract terms | Incident notification, response times, a tested fallback | NIST AI 600-1 covers these in GV-6.2-006 and GV-6.2-007 |
On portability, ask whether the platform can call tools over the Model Context Protocol, whose introduction opens: "MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems." The same page names Claude, ChatGPT, Visual Studio Code and Cursor among the applications that support it. Connectors written to an open protocol are the part of the work you keep if you switch. A 0 on any row is a question to send the vendor in writing before a pilot starts.
Run one pilot before you commit
Whichever way the tests point, a four-week pilot on a single workflow settles the inputs the example had to assume. It needs read access to the systems the agent draws on and a set of past cases with known good answers to grade against. Log runs, units and tokens per run, review minutes per output and the share of outputs a reviewer changed, then judge the route on cost per accepted output and on the pipeline measure the workflow feeds, such as time from lead to first qualified touch. Keep a person approving anything that writes to the CRM, reaches a customer or moves budget; our governance operating model sets out the permission tiers. Watch for the failures that distort a comparison: retries that multiply billed units, a model swap mid-test, prepaid credits that expire unused and a connector outage with no manual fallback.
Our guide to putting one marketing workflow on agents covers the setup. If the pilot shows volume well under your break-even and no data rule the platform fails, buy and look again at renewal. If it shows the reverse, the case for building rests on measured numbers, and scoping that build is the work our agentic AI automation team does.
Sources
- Anthropic, Pricing
- Anthropic, Model deprecations
- Anthropic, API and data retention
- Anthropic, Claude Managed Agents overview
- LangChain, LangSmith pricing
- n8n, Pricing
- Salesforce, Agentforce pricing
- Microsoft, Copilot Studio pricing
- U.S. Bureau of Labor Statistics, Software Developers, Quality Assurance Analysts, and Testers
- NIST, Generative Artificial Intelligence Profile, NIST AI 600-1
- Model Context Protocol, What is the Model Context Protocol (MCP)?