Good Timing, Bad Foundation: Why Enterprise AI Keeps Stalling at the Same Wall

In 2013, I co-founded Nimeyo and spent the next five years building an enterprise knowledge graph that surfaced contextual information for sales, customer support, and R&D teams in their own tooling surfaces like Outlook, Gmail, and Salesforce.

The customers were real. MobileIron, Western Digital, Synopsys, Cisco, Okta. These were not pilots with friendly evaluators. These were enterprise deployments with procurement cycles, security reviews, and IT sign-offs.

The technology was new and rooted in applying a company's own data assets — email archives, CRM records, ticketing systems, and internal wikis — to build the knowledge graph using Natural Language Understanding (NLU). Although we demonstrated real success with some customers, we couldn't achieve the sales velocity to break through.

A decade later, a company building on a similar premise reached a $7 billion valuation. The pattern is worth examining closely, because the lesson isn't that I was early. It's that the obstacle I ran into has two distinct parts, and only one of them has gotten easier since 2013.

Three constituencies, three definitions of success

Every enterprise deployment I ran involved three constituencies with three different definitions of success. Users wanted qPod to make their day faster: less searching, less recreating work that already existed somewhere, more time on the actual job. IT and security wanted to know the data was governed correctly, access was scoped properly, and nothing introduced new risk. Finance and procurement wanted a cost case that held up against a budget cycle measured in quarters.

These aren't different levels of enthusiasm for the same goal. They're different goals. A tool that delights a salesperson by surfacing account history instantly can simultaneously look, to a security team, like a new surface that pulls sensitive customer data across systems that were never designed to share it.

A rollout that IT is comfortable governing slowly and carefully can look, to the budget owner, like a cost center with no clear return inside the fiscal year. I watched deals move forward enthusiastically at the user level and stall completely at the procurement level, not because anyone was wrong, but because the people signing off were optimizing for something the user-facing pitch never addressed.

This is why "find an executive champion" was never a sufficient fix. A champion can accelerate one constituency's priorities. They can't dissolve the structural tension between velocity, governance, and cost. That tension exists inside the org chart, independent of who's selling the product or how good it is.

What a market wave does that one vendor can't

In 2013, Nimeyo had the data integrations, the access control logic, and the in-workflow surfaces that solved the technical half of this. What I didn't have, and couldn't manufacture, was a way to get users, IT, and finance to converge on the same urgency at the same time. That kind of convergence doesn't come from a better pitch deck. It comes from an external force big enough that every constituency feels pressure to move together.

That's what the current AI wave is providing, broadly and for the first time since Nimeyo was operating. Boards are asking CIOs about AI strategy. Users are arriving at work already using AI tools at home and expecting parity.

Finance teams are benchmarking AI spend against competitors instead of asking whether to spend at all. The misalignment between user velocity, governance, and cost hasn't disappeared. But for the first time, all three constituencies are being pushed toward the same table by something none of them control individually.

That pressure exists now. It didn't then. This is the part of the Nimeyo problem that timing has helped.

The unsolved half

None of that touches the other half of what made Nimeyo's deployments stall, and this half hasn't moved.

Access control across enterprise systems is still inconsistent. The same person's permissions look different in the CRM, the file share, and the messaging platform, and reconciling that across systems was hard in 2013 and remains hard now.

Communication is still fragmented across tools that don't share a common record: a decision made over email, refined in a Slack thread, and finalized verbally in a meeting produces no single artifact anyone can reliably retrieve later. The tools enterprises run on still don't expose a shared semantic layer. Each treats "customer," "deal," or "ticket" as its own internal concept with no common definition another system can act on.

This is not the same problem as the misalignment between users, IT, and finance. Better market timing does nothing for it. A CIO under pressure to deploy AI faster doesn't thereby fix the RBAC inconsistency between Salesforce and SharePoint.

What's worth sitting with is that this substrate problem existed in 2013 too. It wasn't solved then either. But it was manageable, because the primary consumer of enterprise knowledge was human, and humans are remarkably good at compensating for organizational ambiguity. A salesperson knew which document to trust despite the folder being a mess. A support engineer understood that "account" in the ticketing system and "customer" in the CRM referred to the same entity even though no system enforced that mapping. An account manager read between the lines of a stale wiki page and applied judgment about what was still current. That compensation was invisible infrastructure. Nobody budgeted for it. Nobody called it a skill. It just happened, at scale, every day, inside every enterprise that functioned well enough to stay in business.

What changes when the consumer is no longer human

Agents don't compensate. They retrieve what the index returns and act on it. An inconsistent permission boundary isn't navigated through context and judgment — it's either honored or it isn't. A stale document sitting next to a current one isn't deprioritized through intuition — it's retrieved with equal weight and potentially acted upon. A concept that means different things across systems isn't unified through experience — it's treated as whatever the nearest embedding happens to represent.

This means the substrate problem that humans quietly absorbed for decades is now fully exposed. The organizational ambiguity that functioned as a manageable background condition becomes a hard failure mode the moment an agent is the one doing the consuming.

Agent-to-agent protocols and emerging interoperability standards could eventually force the semantic consistency this has always needed. If systems need to expose declarative interfaces — something close to "list everyone with access to X" — for agents to act reliably, that pressure might do for technical interoperability what the AI wave is doing for organizational urgency. That's a real possibility. It is not a guarantee, and it is not imminent. Agents work well against clean declarative abstractions. Most enterprises don't have those abstractions, and building them requires exactly the kind of unglamorous, cross-functional organizational investment that has historically been the first thing cut when a budget cycle tightens.

The market conditions that stalled Nimeyo have shifted. The information substrate that caused the stalls has not. And the shift from human to agent as the primary consumer of enterprise knowledge means the cost of that unchanged substrate is compounding in ways that weren't visible when humans were quietly doing the compensating.

That's the part that won't age. Specific protocols will change. Vendor ecosystems will consolidate. But any organization deploying agents on top of ambiguous, fragmented, ungoverned knowledge is now asking machines to do something humans were only barely managing themselves.