Free calculator · 3-year TCO, three paths · prices verified 2026-09-19
Build vs buy AI agents: 3-year TCO calculator
Build vs buy AI agents is a volume question. A vendor platform charges per conversation plus seat fees and wins at low volume. A custom agent costs cents per task in tokens but needs a build and upkeep. This calculator prices three paths over 36 months: a vendor platform, an in-house build with your own engineers, and a fixed-scope studio sprint with retained maintenance. It plots cumulative cost month by month and solves for the monthly volume where building beats buying. Published frameworks put that crossover near a million conversations a year; your numbers move it.
Table view
| Period | Vendor platform | In-house build | Studio build |
|---|---|---|---|
| M1 | 42K | 37K | 35K |
| 84K | 73K | 35K | |
| 126K | 110K | 37K | |
| 168K | 147K | 39K | |
| 211K | 183K | 41K | |
| M6 | 253K | 220K | 43K |
| 295K | 230K | 44K | |
| 337K | 241K | 46K | |
| 379K | 251K | 48K | |
| 421K | 262K | 50K | |
| 463K | 272K | 52K | |
| M12 | 505K | 283K | 54K |
| 547K | 293K | 56K | |
| 589K | 304K | 58K | |
| 632K | 314K | 60K | |
| 674K | 325K | 61K | |
| 716K | 335K | 63K | |
| M18 | 758K | 346K | 65K |
| 800K | 356K | 67K | |
| 842K | 367K | 69K | |
| 884K | 377K | 71K | |
| 926K | 388K | 73K | |
| 968K | 398K | 75K | |
| M24 | 1.0M | 408K | 77K |
| 1.1M | 419K | 78K | |
| 1.1M | 429K | 80K | |
| 1.1M | 440K | 82K | |
| 1.2M | 450K | 84K | |
| 1.2M | 461K | 86K | |
| M30 | 1.3M | 471K | 88K |
| 1.3M | 482K | 90K | |
| 1.3M | 492K | 92K | |
| 1.4M | 503K | 94K | |
| 1.4M | 513K | 95K | |
| 1.5M | 524K | 97K | |
| M36 | 1.5M | 534K | 99K |
- Built-agent run cost per month
- $1,304 (Claude Sonnet 5 tokens + $600 platform)
- Token cost per task
- $0.0176
- Monthly volume where build beats buy
- Build wins at any volume (platform fee alone exceeds the studio line)
- Benchmark crossover
- ~1M conversations/year (Digital Applied, third-party estimate)
- Benchmark 3-year multiple
- $80K build implies $230K to $320K over 3 years (TechCaffeine)
Over three years the studio build is cheapest at $99,169; the vendor platform fee alone exceeds the amortised studio line, so build wins at any volume.
Assumptions and sources (12)
| Constant | Value | Basis |
|---|---|---|
| Claude Haiku 4.5 price per MTok | $1 in / $5 out | sourcedAnthropic pricing page, fetched 2026-09-19 |
| Claude Sonnet 5 price per MTok | $2 in / $10 out (introductory price made permanent) | sourcedAnthropic pricing page, fetched 2026-09-19 |
| GPT-5.6 Terra price per MTok | $2 in / $12 out | sourcedSECONDARY: CloudZero GPT-5.6 pricing tracker (updated 2026-09-10). openai.com/api/pricing returned HTTP 403 on 2026-09-19 |
| Gemini 3.6 Flash price per MTok | $0.75 in / $3.75 out ($1.50 / $7.50 / $0.15 after 2026-12-31) | sourcedGoogle Gemini API pricing page, fetched 2026-09-19 |
| Token split | 70% input / 30% output | estimateStudio estimate |
| Built-agent platform overhead | $600/month | proxyInside the $200 to $2,500 published run-rate band (SoftTeco) |
| In-house upkeep | 0.5 FTE after first release | estimateStudio estimate |
| Studio delivery | 2 months (6 to 10 week MVP range) | sourcedRocket Farm Studios 2026 |
| Default vendor price | $0.99 per resolution | sourcedIntercom Fin benchmarks |
| Default studio price | $35,000 (your quote; not a TheoSym price) | proxyMid-market band $25K to $120K |
| Crossover benchmark | ~1M conversations/year | sourcedDigital Applied 2026 (third-party estimate) |
| 3-year multiple | $80K build implies $230K to $320K | sourcedTechCaffeine (third-party estimate) |
Key takeaways
- Buy wins at low volume. Build wins when tokens per task cost less than the vendor charges per task, at scale.
- A build quote is a third of the 3-year number. Maintenance, run cost and model changes are the rest.
- Two engineers for six months costs more than most fixed-scope sprints. In-house pays off across many agents, not one.
- Buy standard layers: model access, observability, vectors. Build where the workflow is proprietary or regulated.
How AI agent total cost of ownership is calculated
Vendor path: unit price times volume plus platform fees, every month, for 36 months. In-house path: engineers times loaded cost for the build months, then half an engineer for upkeep plus run cost. Studio path: the fixed sprint up front, then maintenance as a share of build plus run cost from month three.
Run cost is the same on both build paths: tokens per task times the blended model price at your volume, plus platform overhead. The cumulative chart shows when the lines cross. TheoSym's build packages are priced as the studio path: fixed scope, eval suite, harness, handover. A $35K default is your quote to enter, not a TheoSym price.
- Vendor = 36 x (volume x unit + fees)
- In-house = build FTE months + upkeep + run cost
- Studio = sprint + 3 x maintenance + 34 months run cost
When does building an AI agent beat buying a platform?
Published 2026 frameworks put the crossover near one million conversations a year (Digital Applied). Below it, buy: speed, integration, the vendor owns the harness. Above it, build: model cost per task is cents while vendors charge around a dollar per resolution.
The calculator solves for your crossover: the volume where the vendor's monthly bill equals the studio path amortised over 36 months. Vendor fees, your model choice and tokens per task move it by an order of magnitude. The AI Factory workproofs show what the built path looks like when it ships.
Studio build vs in-house engineers
For the first agent, a fixed-scope studio sprint is usually cheaper than two engineers for six months, and it arrives with the eval suite and harness already written. In-house wins when you have many agents to build and the evals, MCP layer and harness become shared infrastructure your team runs.
Most enterprises that scale do neither exclusively. They buy standard layers such as model access, observability and vector storage, and build where the workflow is proprietary or regulated (Lyzr CTO framework). TheoSym's AI consulting maps your workflows layer by layer before the TCO is priced.
Build vs buy AI agents in financial services
Regulated workflows change the answer. A vendor agent that cannot show a reproducible decision path, a versioned rule set and an audit trail fails the model-risk review regardless of price. That pushes credit, underwriting, disclosure review and reporting toward a built decision layer.
QGI, led by Dr. Sam Sammane, Founder & CEO, builds those layers deterministic: same inputs, same decision, every run. The TCO for that path includes evidence packaging, not just tokens. The press center carries both companies' announcements.
build vs buy AI agents: questions people ask
When does building an AI agent beat buying a platform?+
Published 2026 frameworks put the crossover near one million conversations a year: below it, buying wins on speed and integration; above it, building wins because model cost per task is cents while vendors charge around a dollar per resolution. Your resolution rate and platform overhead move that line, so the calculator solves it for you.
What is the real three-year cost of a custom AI agent?+
Guides warn that an $80,000 build implies a three-year budget of $230,000 to $320,000 once run costs, maintenance, model changes and monitoring are included. The calculator makes those lines explicit: build, annual maintenance as a percentage of build, LLM run cost at your volume and platform overhead.
Is a studio build cheaper than hiring engineers in-house?+
Often for the first agent. Two fully-loaded engineers for six months costs more than a fixed-scope studio sprint, and in-house teams typically need ongoing capacity for upkeep. In-house wins when you have many agents to build and the platform, evals and harness become shared infrastructure.
Which layers should be built versus bought?+
Most enterprises that scale do not pick one side. They buy standardised layers such as model access, observability and vector storage, and build only where the workflow is proprietary or regulated. Evaluate layer by layer and score each workflow on TCO, speed to value, governance and future-proofing.
Is the studio price in this calculator a TheoSym quote?+
No. The $35,000 default is a placeholder inside the published $25K to $120K mid-market band. Enter the quote you actually received. TheoSym scopes one workflow on the strategy call and prices a fixed sprint from that scope; the calculator only shows how any fixed price compares with vendor and in-house paths.
Next step
Turn the estimate into a scoped plan
Send the calculator result. You get one reply from a human within one business day, with an eval plan for the workflow and a real scope.
Prefer to talk? +1 657-888-0688 or contact@theosym.com