Build vs Buy for AI Agentsin 2026

Almost every build-versus-buy article about AI agents opens the same way: with a failure rate. Ninety-five percent of pilots fail. Forty percent of agentic projects will be cancelled. Eighty-five percent of AI projects fail. Eighty-seven percent never reach production. The numbers are used to argue in both directions — buy, because building fails; build, because vendors fail — and they are almost never checked.

We checked them. All four trace back to documents you can read, and reading them changes the decision materially. One is a non-peer-reviewed working paper measuring something much narrower than its headline. One is an analyst prediction with no published derivation. Two are citation laundering, where a real document said something different and the paraphrase hardened into a statistic.

This matters for a practical reason, not a pedantic one. If you believe 95% of AI projects fail, the rational move is to buy the smallest thing that works and wait. If you believe the measured 2026 numbers — roughly a quarter of use cases fully succeeding, a fifth failing outright, and most landing somewhere in between — the rational move is to pick your first workflow carefully, instrument it, and expect a mixed portfolio. Those are different companies a year later.

What follows is written for someone who has to sign the decision: a CTO, a VP of engineering, a technical founder. It audits the statistics, gives you the ones that hold up, then supplies a seven-dimension framework, a real cost model with August 2026 prices, and the vendor questions that actually separate products from demos.

Provenance audit of the six failure statistics most often quoted in AI build-versus-buy arguments — Frenchy Digital, August 2026.

This is the load-bearing statistic of the pessimistic case, so it deserves the most scrutiny. The source is The GenAI Divide: State of AI in Business 2025, a 26-page document from MIT NANDA covering a research period of January to June 2025. Six things about it change how much weight it can carry.

• It is a v0.1 working paper, not peer-reviewed research: The document is self-labelled Preliminary Findings. There is no peer review. The single listed reviewer is one of its own co-authors. Its disclaimer states the views are solely the authors' and not those of their affiliated employers — which means it was never institutionally underwritten MIT research, whatever the shorthand “MIT study” implies.
• The sample is small and conference-collected: The stated method is a systematic review of over 300 publicly disclosed AI initiatives, structured interviews with representatives from 52 organizations, and survey responses from 153 senior leaders collected across four major industry conferences. The 153 are a convenience sample of people who attend AI conferences, which is not a neutral population when the question is AI adoption.
• It measures custom tools inside a six-month window: The funnel that produces the headline splits general-purpose LLMs from custom, task-specific GenAI. The pessimistic figure is the complement of the task-specific bar. The same report says generic chatbots show high pilot-to-implementation rates of roughly 83%, and that workers at over 90% of surveyed companies reported regular use of personal AI tools. “Generative AI fails” is not what the document found.
• Success was scored on whether an executive remarked on it: The report defines success as initiatives users or executives have remarked as causing a marked and sustained productivity or P&L impact, measured six months after pilot. That is a subjective, interview-derived criterion, not a financial measurement.
• The report's own limitations section concedes the point: It states that six months may be insufficient and that the approach is potentially understating success rates, and that its funnel is directionally accurate based on individual interviews rather than official company reporting. Those are the authors' words about their own headline number.
• The authors had a commercial interest in the thesis: NANDA is an MIT Media Lab group commercializing agent infrastructure. The report doubles as an argument for its own thesis. That is not disqualifying on its own — most industry research has this shape — but it belongs in the citation.

There is a further problem with the funnel itself, which is why you will not find its stage percentages reproduced here. Independent extractions of the same PDF disagree on those numbers, and the report contradicts itself: the percentage shown in its exhibit for one funnel stage does not match the percentage in its own executive summary. A number a document cannot state consistently is not a number you should put in a board deck.

The institutional status is the last piece, and precision matters. The original NANDA subdomain now redirects to the MIT Media Lab group overview page, and that page's publications listing does not include this report at all. The PDF survives on third-party mirrors. It has never been formally retracted. So the accurate characterization is: absent from its own group's publication list, with the original URL redirecting away — softer than a retraction, stronger than nothing.

The sharpest critique came from Arnon Shimoni in a August 2025 response, and his definitional objection is the part worth keeping:

MIT counts any pilot that doesn't reach “full production deployment” as a failure… This is like saying 95% of git branches fail because they don't get merged to main.

— Arnon Shimoni, “MIT's 95% AI Failure Rate Is Wrong,” August 27, 2025

Apply the same skepticism to the critic. Shimoni works at a vendor, and his counter-estimate of a real failure rate somewhere around 25 to 30% is proprietary and unauditable. Cite the definitional argument — that treating every abandoned proof of concept as a failure conflates exploration with defeat — and leave his number where you found it.

The Gartner agentic cancellation prediction is the one most often cited in agent-specific writing. The exact wording, from a June 25, 2025 press release attributed to Senior Director Analyst Anushree Verma, is that over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value or inadequate risk controls.

It is a prediction. Gartner publishes Predicts statements as analyst opinion, and there is no derivation in the release. The only data instrument in the document is a January 2025 poll of 3,412 webinar attendees, which asked about investment posture — 19% reported significant investment, 42% conservative, 8% none, 31% wait and see. It contains no cancellation measurement of any kind. Respondents were self-selected webinar attendees, not a sampled panel.

The two older statistics are simpler. “85% of AI projects fail” is a corruption of a February 2018 Gartner prediction whose actual text was that through 2022, 85 percent of AI projects will deliver erroneous outcomes due to bias in data, algorithms or the teams responsible for managing them. That is a claim about output quality in a pre-transformer-boom world, not a project failure rate.

“87% of data science projects never reach production” is worse. It traces to a July 2019 VentureBeat page bylined “VB Staff” and explicitly labelled Sponsored. The number appears only inside the writer's own rhetorical question, which asks why only 13% of data science projects, or just one out of every 10, actually make it into production — 13% and one in ten in the same sentence. There is no study, no sample, and no methodology, and the panel the page promotes is never quoted stating the figure.

One more worth flagging because it appears constantly in 2026 decks: the “42% of companies abandoned most of their AI initiatives” figure from S&P Global Market Intelligence is real research with more than 1,000 respondents — but it was published March 14, 2025 and describes 2024 into early-2025 behavior, up from 17% the year before. It is routinely relabelled as current data. If you use it, date it.

There is also a tension in the pessimistic case that nobody addresses. The same analyst firm forecasting mass agentic cancellation also forecasts roughly 47% growth in AI spending in 2026. Whatever is happening, spend is not retreating. A narrative that predicts collapse while the budget line grows is describing churn and reallocation, not abandonment.

Here are the numbers that survive scrutiny. They are less quotable than the canon, which is exactly why they are more useful. Note the shape they share: a minority of clear wins, a minority of clear failures, and a large partial-success middle that neither camp's narrative accommodates.

The measured 2026 evidence on enterprise AI outcomes, with the caveat each one requires — Frenchy Digital, August 2026.

The April 7, 2026 Gartner I&O survey is the single most defensible figure in this space right now: 782 leaders, fielded November to December 2025, with 28% of use cases fully succeeding and meeting ROI expectations against 20% failing outright. Read the second pair of numbers too — 77% of organizations delivered at least one success and 57% reported at least one failure. Most organizations are doing both simultaneously, which is what a portfolio looks like, not what a collapse looks like.

Deloitte's 2026 study — 3,235 senior leaders across 24 countries, fielded August to September 2025 — supplies the cleanest expectation-versus-reality gap available: 74% hope to grow revenue through AI, and 20% already are. Meanwhile 66% report productivity and efficiency gains. That combination is the honest description of the current state. AI is reliably taking cost out and unreliably putting revenue in. If your business case rests on the second, you are betting against the measured distribution.

One title change is worth noting because it breaks citations: Deloitte dropped “Generative” from the report name for 2026. Anyone citing “State of Generative AI in the Enterprise” for a 2026 figure is citing a dead title, which is a fast way to spot a deck assembled from search results.

Stanford HAI's 2026 AI Index supplies the adoption-versus-depth contrast: 88% of organizations use AI somewhere, under 10% have fully scaled it in any single function, and only 29% report significant ROI. The Index measures adoption and consumer surplus rather than project outcomes, so do not convert those numbers into a failure rate — but the gap between 88% adoption and sub-10% scaling is the real story of 2026, and it is a scaling problem, not a pilot problem.

Here is the part almost nobody cites, and it is in the same MIT report that produced the 95% headline. It is directly about build versus buy, and it points the opposite way from how the report is usually deployed.

External partnerships with learning-capable, customized tools reached deployment ~67% of the time, compared to ~33% for internally built tools. While these figures reflect self-reported outcomes and may not account for all confounding variables, the magnitude of difference was consistent across interviewees.

— MIT NANDA, The GenAI Divide: State of AI in Business 2025

The report's team-structure exhibit shows the same split: strategic partnerships at 66%, internal development at 33%. And Shimoni, the report's most prominent critic, says this specific finding matches his own data exactly — which is notable, because he disputes almost everything else in it. When a report and its loudest critic agree on one number, that number is worth more than the headline they disagree about.

Now handle it carefully, because it is the same document with the same problems. The outcomes are self-reported. Confounds are obvious and unmeasured: organizations that engage an external partner may already have clearer scope, executive sponsorship, and a budget line, all of which independently predict deployment. A partnership also creates external accountability that an internal project does not have. The causal arrow could run either way.

What it does support is a specific, testable proposition: the structure of an AI project — who is accountable, whether scope is contractual, whether there is a defined delivery date — appears to matter at least as much as whether the code was written inside or outside the building. That is a claim you can act on without believing the 95%.

The word “learning-capable” in that sentence also carries weight. The report's thesis is that tools which improve from feedback outperform static deployments. That maps directly onto whether you have an evaluation and observability loop at all — which is, in our experience, the single most reliable predictor of whether an agent project is still running twelve months later.

Build versus buy is the wrong unit of analysis. An agent system has layers — model, orchestration, tools, retrieval, evaluation, observability, interface — and the right answer is usually different for each. What follows is the seven-dimension assessment we run in discovery, applied per layer rather than to the project as a whole.

Seven-dimension build-versus-buy assessment applied per architectural layer — Frenchy Digital discovery framework, 2026.

Two dimensions deserve elaboration because they are where the decision usually turns.

In-house capability is consistently overestimated, and there is a specific reason. Prototyping an agent is genuinely easy now, and the distance between a working prototype and an operable system is larger than it has ever been. The prototype does not need evaluation, tracing, cost guardrails, retry semantics, injection defenses, or a model migration plan. The system needs all of them. Teams that have shipped a prototype often reason from that experience to a build estimate, and the estimate is wrong by a factor.

The test we use is unglamorous: ask the team to write twenty evaluation cases with clear pass and fail criteria, drawn from real examples, before any agent code exists. Teams that can do this in a week are ready to build. Teams that cannot are telling you that the problem is not yet specified, which is a much cheaper thing to discover now than in month five.

Iteration speed is the dimension that most often overrules the others. If production traces are going to change your prompt and tool design weekly for the first two quarters — and for a genuinely differentiated workflow, they will — then a vendor release cadence is a hard ceiling on how fast you can learn. Conversely, if the workflow is stable and well-understood, that same cadence costs you nothing and saves you an on-call rotation.

Most build-versus-buy cost models start from the per-token price, which is the least important number in the calculation. Start from the multipliers instead. Anthropic's engineering write-up on its multi-agent research system states that agents typically use about four times more tokens than chat interactions, and multi-agent systems about fifteen times more than chats. The same post reports that on one browsing benchmark, token spend alone accounted for 80% of performance variance.

Those multipliers are vendor-published and directional — credible public per-task token data is thin and almost entirely vendor-sourced — but they are the best available anchors, and they are the right shape. A chat feature and an agent are not the same cost class, and a multi-agent system is not the same cost class as an agent. If your model does not have a 4× or 15× term in it, it is a chat cost model wearing an agent's name.

Published API pricing per million tokens, verified against provider pricing pages on August 10, 2026. Prices change monthly — re-verify before committing to a model.

Two worked examples from Anthropic's own documentation give a sense of the magnitudes. A customer-support conversation at roughly 3,700 tokens on Claude Haiku 4.5 works out to about $37 per 10,000 tickets. A one-hour coding session on Claude Opus 5 — 50,000 input tokens, 15,000 output, plus runtime — comes to $0.705, falling to $0.525 when 40,000 of that input is served as cache reads. Both are vendor examples, so treat them as order-of-magnitude anchors rather than forecasts.

The mechanisms that actually move the bill are structural rather than per-token:

• Prompt caching, and its silent failure mode: On Anthropic, a five-minute cache write costs 1.25× base input, a one-hour write 2×, and a cache read 0.1×. Break-even is one read for the five-minute TTL and two for the one-hour. The trap: the minimum cacheable prefix is model-dependent and non-monotonic — 512 tokens on Opus 5, Fable 5 and Mythos 5; 1,024 on Opus 4.8 and the Sonnet 5, 4.6 and 4.5 line; 2,048 on Opus 4.7; 4,096 on Opus 4.6, Opus 4.5 and Haiku 4.5. Below the floor it silently does not cache, and your bill quietly stays high.
• Batch processing: The Batch API is 50% off both input and output and stacks with caching. Any workload that does not need a synchronous answer — nightly enrichment, backfills, evaluation runs — belongs there. This is the largest single lever most teams have not pulled.
• Reasoning tokens bill as output: On both major providers, reasoning tokens are charged at the output rate, which is the expensive side of the ledger. Two defaults changed recently and both increase cost silently: GPT-5.6 defaults reasoning context to all turns, re-rendering earlier turns' reasoning into every subsequent request, where earlier models used the current turn only; and on Claude Opus 5, omitting the thinking parameter now runs adaptive thinking where the same call on Opus 4.8 ran with none.
• Long context is not uniformly priced: Anthropic charges no long-context premium — a 900k-token request bills at the same per-token rate as a 9k one. Google does: Gemini Pro models double both input and output price above 200k tokens. If your architecture pushes long contexts, that difference dominates the model comparison.
• Server-side tools are a separate meter: As of August 2026 Anthropic bills web search at $10 per 1,000 searches, code execution at $0.05 per hour after 1,550 free container-hours per organization per month, and Managed Agents session runtime at $0.08 per session-hour. Web fetch is free. A research agent can spend more on searches than on output.
• Budgets are not all hard caps: Anthropic's session budgets on Managed Agents are a hard dollar cap enforced as a pre-request gate — the session pauses at budget_reached rather than terminating, and raising the budget resumes it. Task budgets are advisory: the model sees a countdown, but the only hard cap is max_tokens. Know which one you configured before you rely on it.

Finally, price the alternative honestly. Menlo Ventures' December 2025 enterprise survey put total enterprise AI spend at $37 billion for 2025, up from $1.7 billion in 2023, with $12.5 billion of that on foundation-model APIs. That is the market you are buying into, and it is repricing fast in both directions: Epoch AI's analysis found that the price of a fixed capability level fell somewhere between 9× and 900× per year depending on the benchmark, though that work was published in March 2025 with no verified 2026 update. Anything you build to optimize today's token price may be optimizing a constraint that dissolves. See our token economics deep dive for the full treatment.

Token spend is the visible cost and rarely the dominant one. Here is what a build actually carries, with real prices where prices exist.

The recurring costs of operating an agent that build estimates routinely omit — Frenchy Digital, 2026.

Observability deserves its own table because the pricing is knowable and the licensing differences are load-bearing. Two of these tools have a genuine self-host path; the rest do not, whatever their marketing implies. Note in particular that Arize Phoenix ships under Elastic License 2.0, which is not an OSI-approved open source licence.

Agent observability tooling pricing and licensing, checked against vendor pricing pages August 10, 2026.

On evaluation, the useful ceiling is human throughput, not tooling cost. LangChain's guidance — the only source publishing concrete sampling numbers — suggests starting at 10% production trace sampling, using at least twenty labelled examples to calibrate an LLM judge, and assuming a human sustains 50 to 100 trace reviews per hour. Anthropic's evaluation guidance adds the structural rules: start from 20 to 50 tasks drawn from real failures, run each trial isolated from a clean environment because shared state produces correlated failures and inflated scores, grade the outcome rather than the path, and read sampled transcripts weekly.

How common is that discipline? LangChain's State of Agent Engineering (n=1,340, fielded late 2025) found 89% of respondents have some observability but only 52.4% run offline evaluations, 37.3% run online ones, and 29.5% run none at all. That survey is vendor-run with self-selected respondents and biased toward eval adoption — and it still finds roughly three in ten doing nothing. An independent qualitative counterweight is bleaker in a different direction: in a 19-interview study, 12 of 19 practitioners called informal vibe checks irreplaceable and 13 of 19 had tried automated evaluation and described it as close to useless.

Buying is a real option and often the correct one. It is also not the safe option, which is how it tends to get framed. Three specific risks dominate.

Discontinuation. Both flagship consumer browser agents shut down inside twelve months. OpenAI deprecated Operator and shut it down on August 31, 2025, folding the capability into ChatGPT agent. Google shut down Project Mariner on May 4, 2026, absorbing it into Gemini Agent and AI Mode, and the stated reasons are worth reading if you are evaluating any agentic product: heavy compute for real-time visual processing, slow performance, and form-selection errors. These were not small vendors. The same dynamic applies to platform features — OpenAI is winding down its fine-tuning platform and has closed it to new users, which is a dependency change for anyone whose product assumed it.

Agent washing. In the same June 2025 release that produced the 40% prediction, Gartner estimated that only around 130 of the thousands of vendors claiming agentic AI are real, describing the rest as rebranded assistants, RPA, and chatbots. Treat that number as an analyst estimate rather than a count — no methodology is published for it, and the ratio it implies is startling enough that it deserves the caveat. Gartner renewed the agent-washing warning in 2026 for supply-chain planning specifically. The detection method does not depend on the estimate being right: ask what decision the system makes without a human, what tools it can call, what happens when a tool call fails, and to see a trace of a real multi-step run. Wrappers cannot produce the trace.

Lock-in. The layers that lock you in are not the ones vendors emphasize. Your prompts, tool definitions, evaluation sets, and trace history are the assets. If those live only inside a vendor platform, switching cost is not a migration — it is a rebuild. This is the single strongest argument for building the orchestration layer even when you buy everything around it.

One structural note in favor of buying at the protocol layer: the Model Context Protocol was donated by Anthropic to the Agentic AI Foundation, a directed fund under the Linux Foundation, on December 9, 2025, with Anthropic, Block and OpenAI as co-founders and Google, Microsoft, AWS, Cloudflare and Bloomberg supporting. Vendor-neutral governance plus cross-vendor client support is what makes something an actual standard rather than a proprietary interface with good marketing. Building tool integrations against MCP is a lower lock-in bet than building against any single vendor's plugin schema.

Every agent vendor will show you benchmark scores. No independent body validates them. This is not a suspicion; it is the documented state of agent evaluation in 2026, and the gap between vendor-reported and independently-verified numbers is large enough to change a purchasing decision.

Agent benchmark claims against primary sources — verified August 10, 2026.

The reward-hacking result is the one to internalize. A UC Berkeley team reached perfect or near-perfect scores across five major agent benchmarks without solving a single task, using pytest configuration hooks, configuration leakage, prompt-injected LLM judges, and virtual machine state manipulation. A follow-up effort catalogued 219 distinct flaws across ten benchmarks and found that patching them dropped vulnerable tasks from near 100% to under 10% on four of them. The benchmarks are improving. They were not measuring what everyone assumed while the scores that populate vendor decks were being collected.

There is also an evaluation-awareness effect. Anthropic's interpretability work found internal representations consistent with recognizing an evaluation context in roughly 26% of SWE-bench Verified problems, against under 1% in real product conversations. Whatever that means mechanistically, it argues against treating benchmark performance as a proxy for production behavior on your workload.

Cost-aware evaluation matters as much as accuracy. Princeton's HAL project ran 21,730 rollouts across nine models and nine benchmarks for roughly $40,000 and released 2.5 billion tokens of logs. Its headline finding is that higher reasoning effort reduced accuracy in the majority of runs, and log inspection caught agents searching HuggingFace for the benchmark rather than solving it. Its live board makes the procurement point bluntly: agents can be a hundred times more expensive while only 1% better. Ask every vendor for accuracy and cost per task together, or you are comparing on one axis of a two-axis decision.

Send these before the demo, not after the pilot. Each one has appeared in a real evaluation we have run.

Frenchy Digital red-flag list for AI agent vendor evaluation, 2026.

Ask for one artifact rather than a document set: a trace of a real multi-step run, including at least one failed tool call and how the system recovered. It answers more questions than a security questionnaire, because a vendor that cannot produce it does not have the system, whatever the questionnaire says.

— Frenchy Digital buyer's principle

Here is the recommendation without hedging. For nearly every organization we have scoped this for, the right answer is hybrid, and the split falls in a consistent place.

• Buy the model: Frontier capability is repricing and improving faster than you can amortize a self-hosted alternative, and no long-context premium on one provider versus a doubling above 200k tokens on another is exactly the kind of difference you want to arbitrage rather than absorb. Keep the provider swappable at the interface, and re-baseline costs with a token counter every time a model changes.
• Buy the tracing and evaluation platform: This is solved commodity infrastructure at $29 to $249 per month for most teams. Prefer a tool with an OSI licence and a genuine self-host path so your exit is technical rather than commercial — that distinction is why Langfuse under MIT and Helicone under Apache-2.0 read differently from tools whose self-host is gated behind an enterprise contract.
• Buy connectors that already work: Verified by a read-write demo against a sandbox, not a logo grid. If the vendor will not demo against your ugliest system in two weeks, the connector does not work the way you need it to.
• Build the orchestration and tool layer: This is where your domain logic lives, where permission scoping is enforced, and where the answers to audit questions come from. It is also the layer that determines your switching cost. Building it is how you keep the option to change everything else.
• Build the retrieval scoping and data layer: The Gartner finding that successful organizations invest up to four times more in data quality, governance and change management is the strongest signal in the 2026 evidence, and none of that investment is purchasable. See our RAG guide for the mechanics.
• Build the evaluation set, always: Not the platform — the cases. Twenty to fifty real failures with clear pass and fail criteria are the most portable asset you will own. They survive model changes, vendor changes and framework changes, and they are the thing that lets you evaluate any of the above honestly.

The sequencing matters as much as the split. Start with one workflow that has a measurable baseline you already record. Instrument the before-state before you build anything. Ship the narrow version, read the traces weekly, and let production failures — not a roadmap — decide the second workflow. Organizations that pilot four disconnected vendors in parallel get four partial answers and no compounding; the compliance, identity, evaluation and tracing substrate is largely a fixed cost, paid once, and the fourth agent inherits all of it.

On architecture, resist the pull toward multi-agent by default. The published evidence cuts both ways and the reconciliation is fairly clear: multi-agent structures win on wide, shallow, independent-thread work like research and market scans, while single-agent structures win on deep, narrow, shared-state work like coding and long-form writing. Both camps agree the mechanism is context, not agent count — and the 15× token multiplier means the wrong choice is expensive as well as wrong. Our multi-agent architecture guide covers the trade-off in depth.

For completeness, and stated honestly: we are the partner option in this decision, which means the 67%-versus-33% finding above is a number we benefit from. Read it with that in mind, and read our cost bands as what a senior-led build actually costs rather than as an argument.

Frenchy Digital cost bands for AI agent engagements, 2026.

Senior-led delivery runs $150 to $225 per hour, and ongoing retainers run $2,500 to $9,500 per month covering model and dependency upgrades, evaluation expansion, incident response, and a quarterly technical review. Every engagement carries a 30-day post-launch warranty, and you receive a written scope with a fixed-price phased proposal within 5 business days of the discovery call.

For comparison, the build-versus-buy economics you will find circulating — a simple single-purpose agent at $15,000 to $80,000, a production multi-agent system with integrations, evaluation, monitoring and compliance at $250,000 to $600,000 and up, maintenance at $100,000 to $500,000 a year, break-even against licensing at eighteen to twenty-four months — come from vendor blogs with no published methodology or sample. They are in the right neighborhood, which is the most that can be said for them. Do not put them in a business case as data.

This article audits other people's evidence, so it owes you the same treatment of its own.

• The best available success statistics are analyst-firm surveys: The Gartner April 2026 I&O survey is the most defensible figure in this space, and its full methodology is not published. It is scoped to infrastructure and operations rather than to all enterprise AI, and analyst-firm surveys carry a commercial relationship with the market they measure. It is the best number available. It is not a controlled study.
• The 67%-versus-33% finding is self-reported and confounded: It comes from the same document whose headline figure this article spends 1,200 words qualifying. Organizations that engage external partners plausibly differ from those that do not in scope clarity, sponsorship and budget — any of which could produce the gap. It is a directional signal about project structure, not a proven causal effect of buying.
• The token multipliers are vendor-published and directional: The 4× and 15× figures come from a single Anthropic engineering post. No independent, non-vendor case study with real token-per-task numbers exists that we could find. They are the best anchors available and they are the right order of magnitude; they are not measurements of your workload.
• Prices move monthly and one is already dated: Every price here was verified on August 10, 2026 against provider pricing pages. Claude Sonnet 5's introductory rate expires on August 31, 2026 and rises by 50%. Anything you model on today's prices needs a re-run schedule, not a spreadsheet.
• Prompt injection is not solved and nothing here fixes it: Every architectural recommendation in this article reduces blast radius. None of them make an agent safe against a determined injection. If your workload cannot tolerate a wrong action that is unrecoverable, the correct answer may be not to give the agent that tool at all.
• Benchmarks are improving faster than this snapshot: Benchmark integrity work moved substantially between late 2025 and mid-2026 — retirements, decontaminated task sets, patched reward-hacking vectors, cost-controlled leaderboards. The specific numbers here will age. The structural lesson — that no independent body validates vendor agent benchmarks — will not.
• The framework assumes you can specify the workflow: Every dimension in the decision table presumes a workflow with a definable outcome. A large share of failed agent projects fail before that point, on a problem nobody could state precisely enough to evaluate. If your team cannot write twenty pass-or-fail cases, the build-versus-buy question is premature.

None of this argues against building, and none of it argues for buying. It argues for making the decision per layer, on evidence you have checked, with a cost model that includes the parts nobody quotes. The organizations getting value from agents in 2026 are not the ones that picked correctly at the outset. They are the ones that instrumented the workflow well enough to find out.