If you want to know when corporate America will actually run on AI agents — not demos, not pilots, but end-to-end automation across sales, finance and HR — Aghi Marietti has an answer that will disappoint a lot of people: about five years.
Marietti is co-founder and CEO of Kong, an infrastructure company whose AI gateway sits between several hundred large enterprises and the model providers they call. That position gives him something rare — a live ledger of which models companies are actually paying for, and how much. Speaking on the MTS podcast, he laid out what that ledger shows, and why the industry’s favorite narrative about AI agents is running well ahead of the reality inside large companies.
The headline numbers are not what the discourse would lead you to expect.
Open-weight models are winning the loud part of the market, not the big one
The debate over open-weight models — freely downloadable systems anyone can run themselves — tends to be framed as a takeover. Marietti says the data splits sharply by customer size, and the two stories get conflated constantly.
“There is open weight adoption in the last 9 months but it’s still actually minimal — if you take the largest enterprise cohort it’s like under 5%,” he said. “But then you have data like SMB routers or longtail providers where you have a spike of open weight model even 20-30%.”
The gap isn’t about model quality. It’s about governance. Large enterprises carry security, compliance and regulatory obligations that make a model you can’t audit or contract with a hard sell, no matter how cheap the tokens are. The long tail has no such burden — and behaves exactly as economics would predict.
“The long tail will always go for the cheapest intelligence to get stuff done, but in large, large, large enterprise most of the spend is still to the usual suspects, whether it’s Anthropic or OpenAI,” Marietti said.
He frames the long tail as the loud part of the market: millions of developers generating visible debate and benchmarks, while enterprises are “much more quiet by design” yet consume enormous volumes of tokens. Kong, he said, proxies a run rate approaching one quadrillion tokens.
SegmentOpen-weight share of token trafficWhyLargest enterprise cohortUnder 5%Security, compliance, governance aversionMidsize enterprise”Much more” than largest cohort (no figure given)Less governedLong tail (SMBs, routers, small providers)20–30%”Always go for the cheapest intelligence to get stuff done”
The average big company is running 20 to 30 models, not three
The assumption most people carry — that an enterprise picks a model vendor the way it picks a database vendor — turns out to be wrong. Among Kong’s global-2000 customers, 61% of active behavior involves running multiple models at once. Marietti says the common guess of four or five models is far off; some companies run hundreds, though often different versions from the same provider. The typical count is 20 to 30 model variants.
Within that traffic, one model stands out in a way that undercuts the idea that spend simply follows usage.
ModelShare of tokensShare of spendClaude Sonnet 549%32%
Sonnet 5 handles nearly half the tokens but captures under a third of the money — a direct signature of tiered routing, where cheap capacity does the volume and expensive capacity does the hard work. Marietti adds that GPT-4.1 and 5.1 remain front-running, and that Claude Opus “has been showing up more and more since August.”
That raises a question he leaves open: will enterprises eventually become “mini labs” of their own, post-training models on proprietary data? He compares it to companies that had no engineering function 20 years ago and later internalized it. The incentive is obvious — a company sitting on unique data may not want to hand it to a model provider that could one day compete with it.
The agent blocker isn’t intelligence — it’s wiring
Here is the argument that ties the whole episode together. Marietti contends that model capability stopped being the constraint on agentic AI some time ago. The constraint is connectivity.
He cites a Cursor study from roughly a month before the interview finding that 21% of inference cost goes to tool calling — the process by which an agent decides which tool or MCP server to invoke. A fifth of the bill is spent just figuring out where to go.
“Finding those workflows and automating them is not an intelligence problem, it’s a connectivity problem,” he said.
His illustration is the departmental silo. HR builds an agent on top of Workday and declares victory. Finance does the same on NetSuite. Talent acquisition uses whatever its vendor shipped. None of it adds up to automation across sales, marketing and finance.
“HR builds an agent on top of Workday and they say hey, we’re doing agents. No, you’re just adding agents on top of Workday… you’re not really automating an enterprise yet,” he said.
What’s missing is a mapped, discoverable nervous system — unified data storage, APIs and programmable endpoints — plus the access controls to keep it safe, since an engineer-built agent with payroll access could do real damage.
“The biggest problem is the connectivity. This nervous system in enterprise is not mapped out, it’s not accessible, it’s not discoverable, it’s not findable,” he said.
That is where the five-year timeline comes from. Small businesses and mid-market companies move faster because they have fewer systems and fewer people to answer to. A highly governed large company takes far longer — Marietti’s estimate is roughly five years before anyone can point to a genuinely agentic Fortune 500 enterprise.
Cheaper tokens won’t shrink the AI bill — they’ll move it down a ladder
When Anthropic cuts prices on new models, the intuitive expectation is that enterprise AI spending falls. Marietti thinks the opposite: cheaper tokens create more demand, and the bill keeps climbing. His model of how enterprises consume intelligence is a ladder.
At the top sits frontier intelligence reserved for mission-critical work — his example is a pull request merged into production code under a 99.99% service-level agreement. Below it, cheaper models handle everyday tasks. He predicts roughly 80% of enterprise intelligence consumption will land on the cheaper side, with the remaining 20% going to frontier models — some of which will spend their time reviewing the cheaper models’ output.
That structure turns model selection into a genuine strategic decision along three axes — speed, quality and cost — rather than the default-to-the-smartest-model behavior enterprises showed a year earlier. It’s also why model routers are proliferating.
Routing gets commoditized. The intelligence on top of it doesn’t.
Marietti draws a distinction that matters for anyone trying to figure out where value accrues. Statistical routing — optimizing purely for cost, speed or quality — is already a commodity and will become more of one. What retains value is the layer above: failover when an inference provider goes down, and routing that improves through accumulated context rather than static optimization.
He expects routers to get “smarter” by embedding their own models, and flags adjacent efficiency techniques such as prompt compression as part of how enterprises will chase more intelligence per token.
Three moats that survive even if the code layer goes to zero
If AI writes the software, what’s left to invest in? Marietti’s framework has three parts.
First, brand. His analogy is Coca-Cola: anyone can copy the taste, but rewiring the positioning already installed in a billion people’s heads is extraordinarily hard. He argues OpenAI and Anthropic have claimed that position — Anthropic through Claude Code on both the developer-harness and consumer sides — and that Gemini and others face “fatigue in catching up.”
Second, network effects: a marketplace or social network that resists displacement.
Third, proprietary data no one else can train on — the same thread that connects to his open question about enterprises becoming mini labs.
He also points out where intelligence is thinnest: coding has plenty, but highly available systems and infrastructure don’t, because there’s no training data for it — only static code and syntax.
The surprising prediction: closed frontier models keep outrunning the skeptics
If the open-weight argument and the cheap-token argument point one way, Marietti’s one-year forecast points the other. Closed frontier models, he says, will still drive and pump a lot of revenue, and their growth rate will be phenomenal. He expects to look back and be startled at how fast they grew, even with strong open-weight alternatives on the market.
He concedes open-weight will grow too — he just rejects the idea that it takes everything. The reasoning loops back to his moat framework: brand positioning is sticky, and enterprises buy the positioned solution they have used for a decade, as long as the provider keeps delivering. He also expects model providers to move up the stack and build moats around their models for specific vertical applications.
The tension to watch is the one Marietti himself leaves open. All of these figures come from Kong’s view of its own proxied traffic, which skews by construction toward large, governed customers — precisely the segment least likely to adopt open-weight models. If enterprises with proprietary data eventually start post-training their own models at scale, that would erode the brand moat he describes from the inside, and his gateway data would be the first place it showed up.
For investors, the practical implications run in three directions. Connectivity and routing layers — API registries, programmable endpoints, intelligent gateways — look like the durable chokepoints, because that is where the bottleneck actually sits. Single-vendor model bets carry more risk than the “one winner takes all” framing suggests, since the average big enterprise is already running two to three dozen model variants. And the AI spending bill is likely to rise even as the price per token falls, because cheaper intelligence on the bottom 80% of the ladder will unlock volume that the expensive top tier never could.