{"slug": "for-enterprise-ai-adoption-is-so-last-year", "title": "For Enterprise, AI Adoption Is So Last Year", "summary": "Enterprise AI coding adoption has shifted from experimentation to infrastructure, with cost control and usage visibility becoming top priorities. At Gartner's summit, attendees focused on managing spiraling AI costs, as usage-based billing and agentic workflows drive token consumption. Gartner predicts AI coding costs will surpass an average developer's salary by 2028.", "body_md": "# For Enterprise, AI Adoption Is So Last Year\n\n### Focus is on visibility and control\n\nAt Gartner’s Application Innovation & Business Solutions Summit, which the Kilo team attended in early June, the conversations around AI coding had changed. Six months ago, most enterprise conversations were all about adoption: *How do we get developers to try this? Are coding agents actually helpful?*\n\nThose conversations are still happening. Some teams are just getting started, and there are still people who do not yet have the [vocabulary for coding agents, model routing, or token economics](https://blog.kilo.ai/i/203225763/the-subsidies-ended).\n\nBut instead of “How do we start?”, the most common question was “How do we keep AI costs from spiraling?”\n\n**Cost concerns**\n\nEnterprise AI coding adoption is moving from experimentation to infrastructure, and that changes things. When a developer tries an AI coding tool for a side project, the question is whether it feels useful. When an enterprise rolls it out across teams, the questions are much more operational:\n\nWho is using it?\n\nWhich models are they using?\n\nWhich projects are spending the most, and how much value do they produce?\n\nWhere are tokens being wasted?\n\nWhat happens when your usage bill increases by 3–10x in a single month?\n\nThese core questions guided the conversations we kept having, and they weren’t just hypothetical. One developer we talked to burned through roughly 8% of their monthly Copilot allotment in two hours. Another spent more than $6 on a single change request and called consumption “impossible to predict.”\n\nGitHub [moved to usage-based billing](https://github.blog/news-insights/company-news/github-copilot-is-moving-to-usage-based-billing/) on June 1, 2026. Premium requests were replaced by GitHub AI Credits, and usage has since been calculated from input, output, and cached tokens. Everything agentic — chat, agent mode, multi-step sessions, tool calls — became metered.\n\nThe token bill was always there, buried under flat-rate plans and “premium request” abstractions. Now it’s on the invoice.\n\nUber was the canary in the coal mine. They burned through their entire 2026 AI coding tools budget by April and [capped employee spending at $1,500 a month](https://techcrunch.com/2026/06/02/uber-caps-employee-ai-spending-after-blowing-through-budget-in-four-months/). They weren’t on Copilot. They were running Claude Code and Cursor. The dynamic is not vendor-specific. Agentic workflows burn tokens faster than flat per-seat budgets were built to absorb.\n\nGartner just published their [prediction](https://www.gartner.com/en/newsroom/press-releases/2026-06-24-gartner-predicts-ai-coding-costs-will-surpass-average-developer-salary-by-2028-as-token-consumption-surges): AI coding costs will surpass an average developer’s salary by 2028. [We saw this coming](https://blog.kilo.ai/p/we-predicted-the-100kyr-per-dev-ai) a year ago. It wasn’t a difficult projection to map: frontier model pricing has been visible through APIs for a long time. What has shifted recently is that the hidden **subsidy layer is collapsing**.\n\n**Visibility is now a product requirement**\n\nWe heard bravely honest feedback from multiple teams: “We have no idea what our usage looks like.”\n\nSome had adopted AI tools aggressively. Some had power users burning through long autonomous sessions. Answering more specific questions is proving harder:\n\nIs usage concentrated in a few engineers?\n\nAre the expensive runs happening in production-critical work, exploratory work, or avoidable loops?\n\nAre non-engineers starting to use coding agents too?\n\nDo model choices differ by team?\n\nWhat does a normal week of usage even look like?\n\nThat last question is important. You cannot optimize what you cannot see.\n\nThis is why usage analytics has moved from “nice to have” to table stakes. Teams do not only need a monthly invoice. They need [per-user, per-project, per-model visibility](https://app.kilo.ai/organizations/). They need to understand spikes. They need to see how much work is moving through agents and whether that work is landing in production.\n\nBut visibility alone is not enough. The dashboard that shows you a $3,000 bill after the month is over helps you understand what happened. It does not help you prevent it. What teams actually need is real-time governance including burn rate during a session, alerts when a run spikes, and honest usage data that separates what you metered precisely from what you can only estimate.\n\n**Model freedom is no longer theoretical**\n\nEnterprise buyers are getting much more serious about model flexibility. Anthropic, OpenAI, Google, and other frontier labs are still producing extremely strong models. For hard work, critical code paths, long-context reasoning, and high-stakes review, frontier models will remain important.\n\nBut enterprises do not want to be locked into one provider. Sometimes the reason is cost. [The same task can cost $1.30 on a frontier model and $0.07 on a capable open-weight model](https://x.com/kilocode/status/2063259733025321029). Sometimes it’s availability. Anthropic’s Claude Fable 5 was the most capable coding model on the market, and then, days after launch, a US export-control directive forced Anthropic to pull it for everyone, including paying enterprise customers.\n\nSometimes it’s procurement or data residency. An [X user publicly confirmed](https://x.com/BrianRoemmele/status/2067602185798680587) this month that a Fortune 500 client is moving half their coding workload to open-source GLM 5.2, abandoning Anthropic entirely. If tests go well, the plan is to move everything to a local model.\n\nThat is where model freedom becomes less of a developer preference and more of a risk management strategy. Being able to move between Claude, OpenAI, Gemini, open-weight models, local models, hosted models, and bring-your-own-key (BYOK) setups is necessary.\n\nThrough Kilo Gateway, teams can access hundreds of models via a single API, switching providers or models by changing a provider/model identifier, not by rewriting their integration. It means you can change models without redesigning your engineering workflows.\n\nFor teams ready to build a routing strategy, we put together an [AI Model Selection Strategy Guide](https://41av0y.share-na2.hsforms.com/2cfb1IKMZRUmDOcPAAjwxqg) that walks through the framework: how to route by quality, cost, and risk, and what policy guardrails look like in practice.\n\n**Open-weight models are entering the enterprise conversation**\n\nThe open-weight model conversation also felt more serious.\n\nNot everyone can use every model. Some organizations need models hosted in North America or Europe. Some want local options. Some want open-weight models but still need enterprise controls. That creates a real market for strong open-weight models produced and hosted in regions that enterprise buyers can approve.\n\nIt also creates a new kind of evaluation problem.\n\nThe best model is not always the biggest model. [The cheapest model is not always the best deal](https://arxiv.org/abs/2603.23971). A model that performs well on short tasks may be the wrong choice for long agentic loops. A model that is good at code review may not be the right daily driver for feature work.\n\nAnd generic benchmarks are not cutting it. On SWE-bench Verified, the top six models are separated by 1.3 points, making them look nearly interchangeable. On Terminal Bench 2.0, running through real agent harnesses with real terminal work, that same set of models shows a 20-plus-point spread. GPT-5.5 hits 74.1% completion rate. Kimi K2.6 lands at 54.4%. But this misses out on usage analytics that help you understand what happens in your actual workflow.\n\nThe enterprise buyer needs this clarity. And [KiloBench](https://kilo.ai/kilobench), which runs each model through Kilo’s actual agent harness on Terminal Bench 2.0 and reports true cost and accuracy, including cost per attempt, is how we’ve been closing that gap.\n\n**The real optimization is routing**\n\nSome tasks deserve frontier models. Some tasks are perfectly fine on cheaper open-weight models. Some tasks should start cheap and escalate only when needed. Some tasks should use deterministic tools so the model does less guessing. This is where automatic model routing becomes much more interesting to enterprises.\n\nThey do not want every engineer to manually calculate token economics before every prompt. They want a system that can automatically use strong models when necessary and cheaper models when sufficient.\n\nThat is exactly what [Auto Efficient](https://blog.kilo.ai/p/auto-efficient) does. A lightweight yet best-in-classs classifier reads each session in real time, determines the task type and difficulty, and routes it dynamically to the benchmark-proven best model for that job. You get the right model at every step of the workflow, backed by [KiloBench data](https://kilo.ai/leaderboard) rather than guesswork.\n\nFor enterprise teams, Auto Efficient can serve as the default once policy constraints are defined: which models are whitelisted, whether prompts are retained, and whether routing should be limited to providers that never use customer data for training. The router respects those boundaries and optimizes within them.\n\nAfter all, tokens are the meter, not the value. The useful question is not “How many tokens did this session burn?” but “What did this session produce?” A developer who spends 50,000 tokens on a change that ships is leveraging them to better use than one who spends 50,000 tokens on a change that gets thrown away.\n\nThat shift from “use less AI” to “get more useful work per token” is where the next phase of optimization lives.\n\n**The **[recognition](http://businesswire.com/news/home/20260624426904/en/Kilo-Code-Named-in-the-2026-Gartner®-Coolest-Vendor-Innovations-in-AI-Coding-Agents) confirms the direction\n\n[recognition](http://businesswire.com/news/home/20260624426904/en/Kilo-Code-Named-in-the-2026-Gartner®-Coolest-Vendor-Innovations-in-AI-Coding-Agents)confirms the direction\n\nKilo Code [was named](http://businesswire.com/news/home/20260624426904/en/Kilo-Code-Named-in-the-2026-Gartner®-Coolest-Vendor-Innovations-in-AI-Coding-Agents) one of the **Coolest Vendor Innovations in AI Coding Agents** in Gartner’s May 2026 report.¹\n\nAs Gartner put it: “In this emerging AI coding agents market, features that simply help developers write code are no longer innovative. Instead, the coolest innovations are those that enable developers to orchestrate, supervise, and govern AI coding agent behavior.”\n\nThat is exactly the shift we saw at the Summit. Watch our roundup discussion [here](https://youtu.be/Ek52Zr1N6dw).\n\nToday, more than 3 million developers use Kilo Code worldwide, and through Kilo Gateway, teams can access over 500 frontier and open-weight models through a single platform. The bet has always been that no single AI model will win software engineering and that the teams who can move fluidly between models, with full visibility into performance and cost, will be the ones who ship the best work.\n\n*¹ Gartner, Coolest Vendor Innovations in AI Coding Agents, 26 May 2026. Gartner does not endorse any company, vendor, product or service depicted in its publications.*", "url": "https://wpnews.pro/news/for-enterprise-ai-adoption-is-so-last-year", "canonical_source": "https://blog.kilo.ai/p/gartner-summit-roundup", "published_at": "2026-06-30 12:49:59+00:00", "updated_at": "2026-06-30 12:56:44.233519+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-tools", "ai-infrastructure", "ai-policy", "developer-tools"], "entities": ["Gartner", "GitHub", "Uber", "Claude Code", "Cursor", "Kilo"], "alternates": {"html": "https://wpnews.pro/news/for-enterprise-ai-adoption-is-so-last-year", "markdown": "https://wpnews.pro/news/for-enterprise-ai-adoption-is-so-last-year.md", "text": "https://wpnews.pro/news/for-enterprise-ai-adoption-is-so-last-year.txt", "jsonld": "https://wpnews.pro/news/for-enterprise-ai-adoption-is-so-last-year.jsonld"}}