Intuit — a company with one of the largest AI engineering budgets in enterprise software — built an AI agent architecture, threw it away, built a second one, and threw that away too. All inside roughly four months. And when its VP of AI stood on stage at VB Transform 2026, VentureBeat's enterprise-AI conference held this week in Menlo Park, the framing wasn't confession. It was advice.
VentureBeat's report from the session captures Intuit VP of AI Nhung Ho calling those two scrapped architectures the fast path — not the detour, not the tuition, the actual fastest route to a working system. That inversion should land hard for every mid-market business currently frozen in month nine of an AI platform evaluation, waiting to pick the stack they'll never have to change.
Here's the uncomfortable math: if a company with Intuit's budget and talent can't design a durable AI architecture on the whiteboard, your 18-month “get it right the first time” plan isn't caution. It's a slower way to be wrong.
Key Takeaways
- Intuit rebuilt its AI agent architecture twice in about four months — from a fleet of specialist agents, to a central orchestration layer, to a skills-and-tools system — and its VP of AI calls that churn the fast path, per VentureBeat.
- The second rebuild took 60 days, with a first working version in under 20 — evidence that rebuilds are cheaper than the sunk-cost instinct assumes.
- Meta's infrastructure VP told the same conference the industry has “maybe 20 months” to rebuild for AI agents. If the giants plan for architectural churn, permanence is the wrong thing to buy.
- The hardest part of Intuit's rebuild wasn't technical — it was convincing leadership and the engineers who built the original system that scrapping recent work was right.
- For mid-market buyers, the decisive vendor question changes from “is this the best stack?” to “what does it cost to walk away or re-platform?”
- We recommend a deploy-observe-rebuild loop: one AI Employee on one workflow, with the expectation of a version 2 inside a quarter — not an 18-month platform project.
What AI Architecture Did Intuit Scrap — and What Replaced It?
The sequence, as VentureBeat reports it, went like this. Intuit started with a fleet of specialist agents — separate AI workers, each built for its own task. That gave way to a central orchestration layer: one coordinating brain routing work across the agents. Then the orchestrator itself was abandoned when it started failing under its own complexity, replaced by a leaner skills-and-tools system.
The reason the orchestration layer died is the most instructive part. Agents in that system passed results to each other in natural language, and every handoff shed context the next agent needed to act correctly. “If you have 10 agents and they all are passing to each other, every time that pass happens, error compounds,” Ho said. That's not a bug you patch. It's a property of the architecture — which is exactly why patching wasn't the answer. We walked through this same fork in the road in our rebuild-or-patch decision framework: when the failure mode is structural, patches buy you weeks and cost you quarters.
Two numbers from the session deserve to be pinned to the wall of every planning meeting. The full second rebuild took 60 days. The first working version took under 20. Not eighteen months. Not a committee. Sixty days to fully re-platform the AI architecture of one of the biggest fintech companies on the planet — because by the second rebuild, the team knew precisely what the system needed to do, and knowledge like that only comes from running something real in production.
Intuit had already signaled this posture publicly — VentureBeat's conference preview framed the session around how the company rebuilt its AI infrastructure for fast and complex tasks. The output of all that churn is deliberately modest, too: the clearest customer-facing result is a feature that lets a live AI conversation pull in a human, currently in early testing with about 1% of Intuit's customers. Two full architectural rebuilds, and the flagship feature ships to 1% first. That's what disciplined iteration actually looks like.
Why Is Throwing Away AI Architecture the Fast Path?
Because in 2026, the ground is moving faster than any architecture can anticipate — and the companies closest to the ground say so out loud.
At the same conference, Meta VP of Engineering Barak Yagour put a clock on it. “We spent 20 years building infrastructure for humans. We have maybe 20 months to rebuild the whole thing for a world where humans and agents co-create at scale,” he told VB Transform, per VentureBeat. Inside Meta, agentic queries grew 30x in six months, and Yagour described three infrastructure assumptions breaking at once — capacity, identity, and velocity. His capacity example is worth quoting in full: “One engineer used to mean one unit of load. Now one engineer spawns 10 agents, each spawning subagents. Your 1,000-person org can generate the load of 100,000 users practically overnight.”
Read those two talks together and a pattern emerges. Meta says the infrastructure everyone built for humans has a 20-month shelf life. Intuit says the correct response is to rebuild in 60-day strokes rather than defend any single design. Even at the very top of the market this is genuinely hard — SiliconANGLE reported in early July that Mark Zuckerberg himself acknowledged Meta's agentic AI efforts are progressing slower than he'd hoped. Nobody — nobody — has this figured out to the point of permanence. The giants aren't buying permanent. They're buying rebuildable, and budgeting for the rebuild up front.
Here's the contrast in one table:
| The 18-month platform project | The deploy-observe-rebuild loop | |
|---|---|---|
| First working system | After the plan is “final” | Weeks in (Intuit: under 20 days on rebuild #2) |
| Learning source | Vendor decks, committee reviews | Production behavior, real workflows |
| Response to a flawed design | Defend it (sunk cost) | Replace it (Intuit: 60 days, twice) |
| Architecture assumption | Pick right once | Every version is provisional |
| Risk profile | One large, late, unfixable bet | Many small, early, correctable bets |
And notice what VentureBeat says was the hardest part of Intuit's journey. Not the engineering — the persuasion. Ho said the tougher problem was internal: convincing both leadership and the engineers who built the original agents that scrapping recent work was the right call. That's sunk-cost gravity, and it pulls on a 40-person Fort Wayne firm exactly as hard as it pulls on a Silicon Valley giant. The organizations that win in the next 20 months won't be the ones that never rebuild. They'll be the ones whose culture doesn't treat a rebuild as an admission of failure.
What Does “Buying Rebuildable” Mean for a Mid-Market Business?
You are probably not writing your own orchestration layer. You're buying one — inside whatever AI platform, agent product, or AI Employee service you sign up for. Which means Intuit's lesson reaches you as a procurement lesson: since even the best-resourced builders expect to re-architect, the durability of any one architecture is no longer the thing to underwrite. The exit is.
In our experience, that changes the vendor conversation in three concrete ways:
- Ask the walk-away question directly. “If we leave in 12 months, what do we take with us, in what format, at what cost?” A vendor who can't answer crisply has answered. We covered the operational version of this in the multi-model hedge and business-continuity plan — the cost of walking away is a number you should know before you sign, not during the outage.
- Separate the layer that persists from the layers that churn. Models churn. Agent frameworks churn — Intuit just churned through two in four months. What persists is your data, your workflows, your policies, and your audit trail. That persistent layer is what we've called the agent control plane, and it's the actual buying decision in 2026.
- Price iteration, not installation. An implementation quote that's all setup and no revision cadence is a quote for version 1 only — and version 1 is, on the evidence of this week, a draft.
The institutional-memory point deserves its own sentence, because it's the part mid-market buyers most often get wrong. Intuit could rebuild in 60 days because the knowledge survived the architecture: what the system needed to do, which workflows mattered, where the failure modes lived. When Cloud Radix deploys AI Employees, that's the design principle we hold hardest — the memory, procedures, and guardrails live in an operating layer that outlasts any individual architecture, so a re-architecture swaps the engine without wiping the driver's knowledge of the roads.
How Do You Run a Deploy-Observe-Rebuild Loop Without Chaos?
“Architecture is disposable” can curdle into “nothing is ever finished” if you skip the discipline that makes Intuit's version work. Three guardrails keep the loop honest:
Scope each version to one workflow. Intuit's rebuilds converged because production told them what mattered. You get that signal fastest from a single, measurable workflow — intake calls, document review, lead follow-up — not from a company-wide rollout you'll be afraid to touch.
Keep humans in the loop as the constant. It's notable that across all of Intuit's architectural churn, the through-line was human involvement — VentureBeat separately reported that Intuit's AI agents reached 85% repeat usage, crediting the practice of keeping humans involved. The architecture is the variable; human oversight is the invariant.
Write down what each version taught you. A rebuild without a written postmortem is just churn. The continuously-improving AI Employee model treats every version's logs and lessons as inputs to the next one — that's the difference between iterating and thrashing.
And be honest about the trade-off: a deploy-observe-rebuild loop costs more attention than a set-and-forget purchase. Someone has to watch the system, read the logs, and call the rebuild. If nobody on your team (or your vendor's) owns that, the 18-month project will at least fail slowly enough to feel comfortable. That's the real choice on the table.
One more distinction worth drawing, because the loop gets caricatured both ways: deploy-observe-rebuild is not “move fast and break things.” Intuit's rebuilds happened behind careful production controls — remember, the flagship customer-facing feature is rolling out to roughly 1% of customers first. The method is aggressive about architecture and conservative about exposure. You iterate hard on how the system is built while keeping a tight cap on how much of your business each version can touch until it has earned more. Get that pairing backwards — cautious architecture, careless exposure — and you have the worst of both worlds, which is roughly where most stalled AI initiatives end up.
What Does the Intuit Method Look Like in Fort Wayne?
Here's the liberating part for Northeast Indiana: Intuit's method doesn't require Intuit's budget. It requires Intuit's posture.
A DeKalb County manufacturer doesn't need a two-year AI platform initiative to get what Intuit got — it needs one AI Employee on one workflow (say, quote-request intake or supplier-document review), instrumented well enough that eight weeks of production behavior tells you what version 2 should be. A Fort Wayne professional-services firm evaluating AI vendors right now can borrow the whole playbook with one added line on the RFP: “Describe what it costs us to re-platform or leave.” The vendors who flinch at that question are selling permanence in a market where Meta's own infrastructure VP gives permanence 20 months.
The regional advantage runs in your favor, too. A 30-person Auburn firm can complete a deploy-observe-rebuild cycle faster than Intuit can schedule the meetings to discuss one — Ho's hardest problem was organizational persuasion, and you have a fraction of the organization to persuade. Small is fast, if you stop treating version 1 as a monument.
Stop Evaluating. Start Versioning.
If your AI plan currently has more months of evaluation than weeks of production behind it, invert it. Cloud Radix's AI consulting practice builds exactly the loop this article describes: deploy one AI Employee on one workflow, instrument everything, review what production says, and re-architect without losing institutional memory — because the memory layer is built to outlive the architecture. Version 2 is in the plan from day one. Talk to us about which workflow should be your version 1.
Frequently Asked Questions
Q1.Did Intuit's AI architecture fail because it was rebuilt twice?
No — the opposite, by Intuit's own account. VP of AI Nhung Ho told VB Transform 2026 that scrapping the architecture twice in four months was the fast path to a working system. Each version generated the production knowledge that made the next one better, and the second full rebuild took only 60 days with a first working version in under 20.
Q2.Why did Intuit abandon its agent orchestration layer?
Because errors compounded across agent handoffs. The orchestrated agents passed results to each other in natural language, and each pass lost context the next agent needed. As Ho put it, with 10 agents passing work around, “every time that pass happens, error compounds.” That's a structural property, so Intuit replaced the architecture with a skills-and-tools system rather than patching it.
Q3.How long should a mid-market business expect an AI architecture to last?
Plan for meaningful revision within a quarter or two, not permanence. Meta's infrastructure VP told VB Transform 2026 the industry has “maybe 20 months” to rebuild infrastructure for AI agents, and Intuit re-architected twice in four months. In our experience, a version 2 inside the first quarter is a sign the deployment is healthy — it means production is teaching you something.
Q4.What is the sunk-cost trap in AI projects?
It's the instinct to keep defending an architecture because of what it already cost, rather than what it can still deliver. Notably, Intuit's VP said the hardest part of the rebuilds wasn't technical — it was convincing leadership and the original engineers that scrapping recent work was right. Budgeting for rebuilds up front is the practical antidote.
Q5.What should I ask an AI vendor before signing in 2026?
Lead with the walk-away question: what does it cost — in dollars, data, and downtime — to re-platform or leave? Then ask which layer persists across their architecture changes (your data, workflows, policies, and audit trail should), and what their revision cadence looks like after go-live. Vendors selling a permanent architecture are selling against the direction the entire industry, from Intuit to Meta, says things are moving.
Q6.Can a Fort Wayne business use the deploy-observe-rebuild approach without Intuit's budget?
Yes. The method scales down well: one AI Employee, one workflow, real logging, and a scheduled review that decides whether version 2 is a tune-up or a re-architecture. A 30-person Fort Wayne or Auburn firm actually iterates faster than an enterprise, because the hardest step — organizational buy-in for change — involves fewer people. The budget question isn't “can we afford to rebuild?” but “which single workflow do we instrument first?”
Sources & Further Reading
- VentureBeat: venturebeat.com/orchestration/intuit-scrapped-its-own-ai-agent-architecture-twice-in-four-months — Intuit scrapped its own AI agent architecture twice in four months. At VB Transform 2026, its AI VP called that the fast path (2026-07-17).
- VentureBeat: venturebeat.com/data/we-have-maybe-20-months-to-rebuild-for-ai-agents — “We have maybe 20 months” to rebuild for AI agents, Meta's infrastructure VP tells VB Transform 2026 (2026-07-15).
- VentureBeat: venturebeat.com/orchestration/intuits-ai-agents-hit-85-repeat-usage — Intuit's AI agents hit 85% repeat usage. The secret was keeping humans involved (2026).
- VentureBeat: venturebeat.com/orchestration/intuit-will-show-off-how-it-rebuilt-its-ai-infrastructure — Intuit will show off how it rebuilt its AI infrastructure to support fast and complex tasks at VB Transform 2026 (2026).
- SiliconANGLE: siliconangle.com/2026/07/02/mark-zuckerberg-says-metas-agentic-ai-efforts-arent-progressing — Mark Zuckerberg says Meta's agentic AI efforts aren't progressing as fast as he had hoped (2026-07-02).
- VentureBeat: venturebeat.com/vbtransform2026 — VB Transform 2026 conference hub (2026-07).
Ready to Deploy Version 1?
One AI Employee, one workflow, instrumented from day one — with version 2 already in the plan. We will help you pick the workflow where production can start teaching you fastest.
Schedule a Free ConsultationNo contracts. No pressure. Just an honest conversation about what would help your business.
