Intuit bets its second act on agentic AI, and on rebuilding the plumbing underneath

On a vendor-conference stage scheduled for June 2026, an Intuit engineer is queued up to talk about plumbing. Not the flashy kind. The kind that routes a tax calculation through a model, watches the model's output, catches the drift, and routes the next calculation better. That is the bet the company is making on its second act: not a smarter chatbot bolted onto TurboTax, but a rebuild of the substrate that runs beneath its small-business software, calibrated so that autonomous agents can be trusted to act on a bookkeeper's behalf. The wager is structural, and it is being made in public, in slides, in the language of "agentic AI" that has colonised enterprise software marketing since 2024. Whether the plumbing actually holds will determine whether Intuit's next decade looks like the last one or like something else entirely.

The argument inside the preview is straightforward on its face and radical underneath. Intuit's product suite is, in volume terms, the dominant small-business finance platform in the United States. QuickBooks alone runs the books for a meaningful slice of the country's sole proprietors and very small employers. Mailchimp adds the marketing layer; Credit Karma the consumer credit layer; the tax products the seasonal spike that justifies the rest of the year. For two decades that stack has been assembled as a collection of vertically integrated apps, each with its own data model, its own UI, its own update cadence. The agentic bet is that the next decade of small-business software will not be organised around apps at all, but around intents. A user says "close the month," and an autonomous agent moves entries across ledgers, reconciles bank feeds, files the right form, and reports back. That agent, to be trusted, needs to read every app's data model fluently and write to them safely. Hence the plumbing.

What the substrate actually is

Strip the marketing off and the engineering question is concrete. The company is investing in three layers at once. First, an orchestration layer capable of holding a long-running task across multiple products without losing state when a model hallucinates or a downstream service fails. Second, an evaluation harness, the internal equivalent of a continuous integration pipeline, that runs thousands of synthetic scenarios through the agents and scores them against accounting rules, tax code, and prior human decisions. Third, a data-fabric layer that lets a single agent see the same customer across QuickBooks, Mailchimp, and the tax products without each integration being re-negotiated by hand. None of that is on a roadmap slide in the way a new button is on a roadmap slide. It is the kind of investment that shows up in capitalisation, in the way headcount is allocated, and in how engineering organisations are restructured around platform teams rather than product teams. The preview frames this as the substance of the talk, and the substance is the story.

For competitors, the question is whether this kind of rebuild is replicable or whether Intuit's installed base gives it an unfair advantage. The installed base is real: decades of customer transactions, years of bank-feed integrations, a tax engine updated annually against a code that itself changes annually. A startup can build an agent from scratch, but it cannot build an agent that already knows what a Schedule C looks like across thirty years of IRS revisions, what a regional sales-tax rate does to a quarterly filing, or what an equipment write-off means to a sole proprietor in California versus a partnership in Maine. That institutional memory, encoded in data and in the rules that data has been forced through, is the moat the agentic bet tries to widen rather than let erode.

The marketing problem underneath

Small-business software has a marketing problem that enterprise software does not. The buyer is often a single person, working after hours, on a phone or a laptop at a kitchen table. The vocabulary of "agentic AI" is not native to that buyer. They do not buy "orchestration layers." They buy the absence of a January panic. Intuit's challenge in the second act is to rebuild a substrate that the company itself can barely explain while selling the experience of it to buyers who do not want to be educated about plumbing. The preview does not engage with this marketing problem directly, which is itself a tell. The assumption seems to be that if the agents actually work, the marketing takes care of itself, because the only thing harder than selling plumbing is selling plumbing that does not work.

The risk is that the agents do not work, or work inconsistently, and the company is forced into a cycle of patches, customer-service escalations, and trust repair that consumes the margin it invested in the rebuild. Tax software has a long and instructive history of this pattern. Every filing season produces a category of errors that only human review catches and that only retroactive refunds repair. An agent that misfiles a payroll return, or that confidently issues incorrect tax advice, is the same kind of failure multiplied by the number of intents it handles without supervision. The evaluation harness is supposed to catch that before it ships. The track record of large-language-model evaluations against domain-specific rules is mixed. Building the plumbing is one project. Proving the plumbing works against the adversarial inputs of real small businesses is another.

What the second act actually costs

The financial shape of this bet is not in the preview, but the contours can be sketched from what Intuit has said publicly in earnings and analyst sessions over the prior year. Capital expenditure at the company has stepped up materially as it builds data-centre capacity and reorganises engineering around AI-native workflows. Operating margins in the most recent reported quarters were cushioned by the Mailchimp and Credit Karma product lines, both of which carry higher gross margins than the legacy desktop software Intuit was once known for. That cushion is what funds the rebuild. If the rebuild takes longer than the cushion holds, the financial pressure arrives not as a single quarter's miss but as a multi-year compression that the market will eventually reprice. Investors who have watched Intuit for a decade know the company can absorb this kind of pressure. Investors who came in during the AI-infrastructure rally of the mid-2020s may not.

The competitive set is also wider than it was when QuickBooks first became the default. Vertical SaaS competitors now serve specific industries with deeper workflows. The new generation of AI-native accounting tools, some of them built on top of large language models and some on proprietary infrastructure, are reaching for the same agentic pitch from a younger starting point. Intuit's defence is that the substrate moat compounds: every year of bank-feed data, every tax season of edge cases, every Mailchimp campaign outcome makes the next agent better than a competitor's agent would be. The offence is the rebuild itself, because the substrate that runs the last decade's software is not the substrate that runs the next decade's.

What to watch

The first filing season after the agentic rollout is the real test. If customers using the new agents file without escalation, the plumbing holds and the second act has begun. If escalations spike, or if regulators notice patterns that suggest the agents are confidently wrong in categories that matter, the company will be forced into the kind of trust-repair cycle that erodes moats faster than it builds them. The next earnings call after that season will be more diagnostic than any vendor-conference talk. Until then, the slides say plumbing, and the plumbing, if it works, does not announce itself at all.