In the early 2000s, SAP didn’t win because the software was beautiful. It won because it shipped a pre-baked version of how a company should run (finance, procurement, HR) and then brought an army of implementers to bend it onto the customer’s actual mess. The product was the template. The money was in the deployment.
AI is heading to the same place, and fast. The model is the engine. Nobody buys an engine. They buy the thing that gets their invoices matched, their claims triaged, or their KYC queue cleared.
The labs figured this out first. OpenAI and Anthropic both launched joint ventures for enterprise AI services in May, with OpenAI’s valued at $10 billion and Anthropic’s, built with Blackstone and Hellman & Friedman, valued at $1.5 billion. Salesforce has publicly committed to building a team of 1,000 forward deployed engineers to drive Agentforce adoption. LinkedIn found forward deployed engineering roles grew 42-fold between 2023 and 2025. The companies that make the best models are building services arms. Take that seriously.
Why the model alone doesn’t sell
The demo is easy. Production is where deals die. Kyndryl found 57% of enterprises now have AI embedded in core processes, but only 11% have hit both of their top two AI objectives. Everyone has AI. Almost nobody has the result.
The gap is workflow, data access, permissions, and trust. A model doesn’t close that. A person sitting next to the customer’s ops team for six weeks does, if they have the right tooling in their bag.
The company you’d build today
Think of it as five layers.
1. Models, rented. Don’t marry one. Vercel says the average enterprise on its AI Gateway routes across 17 models. Your customers will switch models every few months. Your product should make that a non-event.
2. The harness. This is the real product. Agent runtime, tool connections, memory, evals, guardrails, identity, audit trail. Vercel is a good example of the substrate forming: its Agent Stack bundles the AI SDK, AI Gateway, Sandbox, Workflow SDK and Chat SDK, and its enterprise package puts internal agents behind the company’s own identity provider by default. Build or buy this, but own the experience.
3. Pre-configs. This is your SAP best-practice template. A packaged agent for one workflow in one industry, with the connectors, evals, and policies already wired. “Reconcile settlements.” “Triage AML alerts.” “Review contracts against our playbook.” The more specific, the faster you close. Glean started with search and Agentforce started with support tickets. Narrow wedge first, platform later.
4. Forward deployed engineers. They take the pre-config, plug it into the customer’s real systems, tune it on their data, and hand it off. AWS embeds them for about 45 days to build production-ready systems and make the customer self-sufficient. That’s the right shape: fixed scope, fast, and a clean handoff.
5. Trust. SSO, audit logs, data residency, run-in-your-own-cloud. For regulated buyers this isn’t a feature, it’s the entry ticket. Boring work, and it wins deals.
How to organize it
Founder-led sales. You sell the outcome and the risk reduction, not the model.
FDE pods that own a deployment end to end, with real authority. A solutions engineer relabeled as an FDE is just a relabel.
A product team whose job is to eat the FDEs’ work. Every deployment should leave behind a better pre-config. If it doesn’t, you’re a consultancy with a nice website.
Security and compliance early. Not after the first bank asks.
Pricing: paid, fixed-scope deployment to land, then usage or outcome-based to expand.
We’ve seen this movie at Blockdaemon
None of this is new to us. We’ve lived a version of it in digital assets.
Institutions don’t want to run their critical infrastructure in someone else’s black box. A bank touching a blockchain wants it inside its own perimeter, under its own controls, wired into its own compliance stack. We learned that selling nodes, staking, and key infrastructure to regulated buyers. AI is hitting the same wall now.
Three things carried over directly:
The control panel is the product. When infrastructure runs on-prem or in a customer’s own cloud, the buyer’s first question is “who can do what, and how do I see all of it?” For us that was a control panel: what’s running, who approved it, what the policy is. For AI it’s the same list: which agents exist, which models they call, what data they touch, what they cost, who signed off. Whoever owns that view owns the relationship.
Compliance integration decides whether you ship. In digital assets, the software is never done until it plugs into the customer’s existing compliance and reporting stack. AI is identical. Audit logs, model risk review, data residency, human-in-the-loop approvals. Nobody in a regulated shop deploys an agent that can’t produce an audit trail. This is where deals stall or close.
Pre-configured, then deployed by engineers. Same SAP pattern. A hardened, pre-configured stack, then engineers who fit it into the customer’s workflows and hand it off.
That’s why I describe Blockdaemon as the trust layer for institutional finance. The pattern is the same whether the thing being trusted is a validator or an agent.
The AI box
Here’s the opportunity. The buyers we already serve are the same buyers struggling to get AI past security review. They’ve approved us as a vendor, we know how to deploy inside their environment, and we speak compliance. So putting an AI box next to what we already sell is a natural extension.
We use Konstant at Blockdaemon for exactly this reason. Every institutional customer wants something specific, and that’s how you end up with a codebase per customer and a payroll to match. Konstant’s answer is that your company is ten to twenty parts in one box. It takes the software, the methods, and even one person’s judgment out of the stack you already run, and turns each into a named part with a defined input, output, and revision, so an agent can call it. Nothing gets rewritten or migrated. The parts run in a box you control, hosted by Konstant or in your own cloud.
Then you pick the number you want to move (cost per order, margin, tickets touched by hand) and the system tries parts in and out of the live work, measures whether each move got you closer, and rolls back what doesn’t, with a person judging every step.
If that sounds like the SAP playbook, it is. The named parts are the pre-configs, the box is the harness, and the engineers who fit it into the customer’s workflow are the FDEs. The difference is that it runs inside the customer’s perimeter, which is exactly where our buyers need it.
The market is moving the same direction. Vercel’s enterprise package lets internal apps and agents live behind a company’s identity provider by default, with the option to deploy inside a company’s own AWS account. When infrastructure players build “run it in my perimeter” into their pitch, the demand is clear.
Who to study
Palantir wrote the FDE playbook. Its AIP platform posted 133% year-over-year US commercial revenue growth in Q1 2026.
Salesforce Agentforce is the incumbent doing the SAP move in real time: template plus FDEs plus an installed base.
Sierra and Decagon show what a sharp wedge (customer service) looks like.
Glean won with search before expanding to agents.
Harvey is the vertical play: legal, deep, trusted.
OpenAI’s Deployment Company and Anthropic’s Ode are the labs becoming the implementers.
Vercel is the substrate for teams building and shipping their own agents inside the enterprise.
Where this goes wrong
You become a body shop. Margins collapse, every customer is custom, nothing compounds. The fix is the loop in layer 3: deploy, extract, productize, repeat.
The bet
The labs are betting billions that deployment is the bottleneck. They’re probably right. If you’re building in this space, decide which workflow you’ll own, then ship the template that makes the next customer’s deployment half as long as the last one.
