Brian Lewis Says Only About 5% of AI Startup Demos Ever Land an Enterprise Contract – finance.biggo.com

Every 11 days, somewhere in the world, a new frontier AI model is released. That cadence — faster than most companies can schedule a security review — has created a strange kind of market distortion. Startups race to demo the newest model capabilities, while the enterprises they pitch are still mid-migration on ERP projects that started five years ago. The gap between those two timelines, according to Brian Lewis, a product lead at Millennium speaking on AI Engineer, is where most AI startup deals go to die.
Lewis runs the enterprise evaluation funnel at one of the most demanding customers a startup can face. His firm screens 10 to 15 startups for every pain point it identifies, schedules two or three demo calls, lands zero to one pilot, and converts about one in four pilots into a signed contract. The math lands at a sobering figure: roughly 5% of demo calls ever produce a deal. “We have a lot of startups that tell me what enterprise ready means,” Lewis noted, “and we go through all of our requirements, and then we have a very different idea of what enterprise ready actually means.”
The pilot window itself is shrinking at a pace that should alarm any startup founder. When Lewis joined Millennium just over two years ago, pilots commonly ran six months. That became three months. Now they’re often scoped at two weeks. For a vendor, that means there is no runway for iterative discovery — the product must demonstrably work on day one, or the pilot ends.
Lewis breaks down the failure points into four buckets: efficacy, security, reliability, and legal. Efficacy — the product actually solving the stated problem — accounts for 40% of failures. That may sound surprising until he details what he sees on demo calls.
Some startups pitch vaporware that his platform team could rebuild in about six weeks. That’s not always a deal-breaker — sometimes buying is still better than building — but it forces a build-versus-buy discussion the vendor never intended to trigger. One startup proposed an upside-down pricing model: Millennium would report its LLM gateway telemetry back to the vendor, so the vendor could charge a margin on traffic that wasn’t even running through its infrastructure. That deal died quickly. Other failures are more mundane: demo-call promises with no ETA two months later, salespeople re-pitching features the customer explicitly declined, and success criteria written by the vendor rather than the buyer.

Security is where a large share of remaining deals die, and Lewis runs frontline defense on those conversations despite not being a security specialist himself. The requirements are specific and non-negotiable.
The failure modes are equally concrete. Lewis has caught vendors sending pilot data to their own cloud servers despite explicit instructions. He has seen integrations that require read-write-all scopes across every system. He has seen beta features auto-enabled with each release, with no rollout control. And he has heard the same promise repeated across weekly check-in calls — “We’ll get you the security architecture diagram next week” — until his CISO distilled the entire conversation into one question: What are you going to do if there’s a breach?
The most common answer is not an answer. “We get this response: well, we haven’t had a breach yet,” Lewis said. “With the subtext of we don’t know what we would do if we did.”

On reliability, Lewis wants a control plane that actually works: every admin setting available via API, audit logs on configuration changes, controlled rollout of updates, real SLAs, and a reachable support engineer. What he encounters instead is a litany of operational immaturity: apps shipping multiple updates per day with a “relaunch to update” button that propagates across 3,000 users with no version tracking; SSL certificates breaking in new releases with no way to diagnose because everyone is on a different version; support documentation changing without versioning, introducing terms never reviewed by legal; and core APIs down for hours during a trading day — a catastrophic event for a firm moving billions of dollars.
The legal requirements are just as strict. No training on customer data under any feature or product. Transparency on sub-processors, because fourth-party risk becomes Millennium’s risk. IP indemnification with reasonable liability caps, since the customer does not control model outputs. Lewis has caught vendors claiming ZDR in the contract while retaining data in practice — caught because the vendor referenced something it should never have seen. He has also seen vendors deliberately keeping features in “beta” status with permissive data-retention clauses buried in the fine print, a sneaky attempt to harvest enterprise data that gets rejected on sight.
Beneath the specific horror stories is a structural insight about what AI actually does to a large organization. “I really look at AI as a flashlight, not a band-aid,” Lewis said. “AI shines a light on a lot of what’s already working or not working. It can accelerate what’s working really well, and it breaks down very quickly when things don’t work well.”
That asymmetry has a number attached to it. Lewis’s estimate, admittedly unscientific but emphatic: 40% of becoming AI-native is about models and products. The other 60% is data hygiene, clean architecture, strong integration, enablement, and change management. ChatGPT has been public for only 43 months — not even four years — while enterprise architectures can be a decade old. No model can fix that gap. If anything, plugging AI into broken foundations will surface the breakage faster and at greater scale.
Shining that flashlight internally surfaced four lessons at Millennium. First, entitlements need a new paradigm. Large enterprises have over-entitled and under-entitled users, and the existing permission models collapse when agents need to exercise judgment and act quickly. “Agents inherit your foundations,” Lewis warned. “If you think about the problems that you run into when things go rogue, processes go rogue, people go rogue, agents are going to like 100X that problem.”
Second, cross-platform integration moved up the stack. AI is only as good as what it can reach, and enterprises want it everywhere. Third, centralized knowledge is critical — thinner agents require a smarter substrate, meaning all documentation and institutional knowledge must be machine-consumable. Fourth, some firms may need a separate sandboxed ecosystem for experimentation when the gap between legacy architecture and the target state is too vast to bridge directly.
Lewis closes with a checklist that doubles as an indictment of the current market. Best-in-class AI startups have security architecture that actually works, support engineers who respond, an admin API from day one, a 90-day plan that deploys into the customer’s infrastructure and cloud, and success criteria written by the customer. The worst have none of those things.
“The worst AI startups don’t have any security architecture diagrams, no path to a support engineer, no deployment control or audit logs, no ETAs,” Lewis said, “and salesmanship over solid product building.”
For buyers, the same checklist inverts into a mandate: invest in entitlements, governance, and audit logging — the boring 60% — before expecting AI to deliver value. At current model intelligence, the bottleneck is not the model. It’s the plumbing. The startups that win will be the ones treating enterprise readiness as a first-class engineering problem rather than a sales afterthought. The enterprises that win will be the ones fixing their foundations before — not after — they turn on the agents.
References:
Once added, BigGo Finance appears first in Google Search Top Stories, so you get the broadest, most up-to-the-minute, and most comprehensive global financial news first.
The news and data on this website are for reference only and do not constitute investment advice or an offer to buy or sell. Information is sourced from exchanges and public sources, and may be delayed, interrupted, or updated. While we strive for accuracy, we do not guarantee timeliness, correctness, or completeness. Content may include external links for which we are not responsible. By using this site, you agree that we and our partners are not liable for any losses. Investment carries full responsibility; please carefully assess risks and consult professionals. If there are errors in the content, please contact us for correction.

source
This is a newsfeed from leading technology publications. No additional editorial review has been performed before posting.