Why enterprise AI isn’t delivering ROI yet – and the infrastructure layer that will change that

Each phase of B2B software – from relational databases to SaaS/cloud – had the same underlying logic: it enabled humans to collaborate by coordinating access to data. AI agents are doing more and more of the work humans used to do, and so the question is: what software are we going to build to enable AI agents to collaborate better, both with us and with each other?

Right now, there is a growing gap between models’ capability and their impact in the real world. Satya Nadella says the real arbiter of AI’s impact is seeing double digit global GDP growth. We’re not there (yet) – and we think that’s because the enabling coordination layer for AI does not exist yet to enable it to deeply diffuse into the economy.

The companies that enable AI diffusion will close the gap.

Read on for:

• What AI diffusion is all about, and why all four elements are needed to accelerate AI;
• Which companies are blazing a trail in AI diffusion; and
• How the lines between companies addressing raw data, the control layer, agents and services, and organisational capacity are blurring.

* * *

What’s AI diffusion?

AI diffusion is a vague term. We find it helpful to split it into four categories:

• Raw data: existing data is messy and siloed; decision traces are not captured.
• Control layer: you can’t unleash agents onto your data without a way to direct attention; equally, you can’t allow agents to take action in your data stack without the ability to track, merge and reverse changes.
• Agents and services: agents need to be configured and improve over time, for each problem space.
• Organisational capacity: organisations and workforces need to be capable of adopting AI.

Beneath all this, models should get better at generalising. The better they get, the easier diffusion gets. But we believe AI diffusion is nevertheless a critical infrastructure layer for AI ROI – hence we invest in all four elements.

Raw data

For agents to complete work, they need access to the right data. We see two key opportunities – and pose a question about the future of legacy systems of record.

Enterprise data management alone will be a quarter-trillion opportunity within 10 years.

Capturing decision traces: You have likely read Foundation Capital’s piece on this. In sum, most software today captures what happened: a record was created, a field was updated, an email was sent. What it does not capture is why. The reasoning behind decisions, the context that informed them, the alternatives that were considered. These decision traces are enormously valuable for training and directing AI agents, and they are almost entirely absent from existing data. The companies best placed to capture them are vertical, domain-specific applications – the ones that sit inside the workflows where decisions are actually made.

Cleaning up existing data: Enterprise data is in a bad state. CRMs are full of duplicates, broken hierarchies and conflicting records. Warehouses contain tables that no one can explain. The companies tackling this are building AI-native data cleaning and structuring tools that work continuously rather than as one-off projects. Kernel is a good example – they resolve entity identities, build corporate hierarchies and remove dead accounts at scale, keeping CRM data clean on an ongoing basis.

The question for incumbents: For incumbents sitting on decades of enterprise data, this is a genuine revenue opportunity as AI agents multiply and each one needs to query something. Salesforce has already started raising prices on apps that tap into its APIs. SAP tried to charge Celonis customers for extracting their own ERP data (and got sued for it). 2026 may see others join in. But charging for access changes what you are. If your primary value to the ecosystem is as a database that agents query, then others own the workflow — and over time, they build their own systems of record. Every specialist AI solution accumulates work artefacts, every agent platform logs decision traces, every control layer company saves context — each could become its own systems of record.

Control layer

Access to data is one thing, but unleashing agents on the entire warehouse is dangerous, costly and ineffective. To make sense of data and act responsibly, agents need a control layer, composed of:

Access to data is one thing, but unleashing agents on the entire warehouse is dangerous, costly and ineffective.

The context layer: The context layer solves the targeting problem: which parts of the warehouse matter for which type of problem? It uses metadata, query logs, human feedback and more to dynamically maintain an understanding of what data is relevant, trustworthy and current for a given task. This is difficult and important for several reasons. There is a lot of data in enterprise warehouses which is conflicting and messy. Compute costs a lot, and you don’t want to unleash agents on the warehouse itself until you are sure exactly what they need. And data lives across many systems at once – silos matter, which means the context layer won’t be built by a single data warehouse provider. Rig is a key example here, building an automation platform for warehouse data that provides context and tooling for business teams to work with complex internal data.

The governance layer: You need a way to track, merge and reverse changes made by agents in your data. As the number of agents running concurrently grows, this governance problem grows exponentially. Think of how Git solved the coordination problem for multiple people contributing code to the same project – versioning, branching, merging, conflict resolution. The governance layer will solve the equivalent problem for AI agents making concurrent changes to data. Modern Relay is building in this space – every write, whether from a human or an AI agent, produces a proposal (what changed, why, with sources) that gets reviewed and approved before it enters the source of truth. The principle applies across verticals: you need version control for agent behaviour.

Company-specific RL environments: The first generation of reinforcement learning (RL) in AI was training foundation models. The second RL generation is company-specific. Enterprises increasingly want to fine-tune or train models on their own workflows and edge cases, but they lack the infrastructure to do so safely. Mercor started by supplying RL to frontier labs, now they are using their expert network to build contained RL environments and evaluation infrastructure for enterprises. Our portfolio company Huzzle is moving into a similar space.

Agents and services

The space of all economically valuable tasks that AI could automate is being divided up by three approaches: frontier models expanding from the top, end users configuring AI tools themselves from the bottom, and specialist AI solutions filling the space in between. Different problem areas lend themselves to each. Two startup opportunities stand out.

The future of work will see human-AI collaboration, with AI augmenting workers by automating manual processes so people can spend more time on higher-value tasks.

Specialised AI solutions: As we’ve covered elsewhere, these win in problem spaces that rely on tacit knowledge – the kind experts can’t easily break down into rules or playbooks, and that doesn’t exist in documents. They handle data that cannot be shared with frontier models, or that is generated by the product itself (which overlaps with the decision traces point above). The tasks are non-routine, judgement-based and end users can’t easily define what “good” looks like in advance. And they tend to focus on high-stakes workflows in regulated environments where getting it wrong has real consequences. PlanLab is a good example – they do AI-powered project scheduling for infrastructure and construction, a domain where the knowledge of how to coordinate thousands of interdependent tasks across dozens of contractors lives in the heads of experienced planners, not in documents. A frontier model can’t do this, and an in-house configuration won’t cut it either. Popp is another – they’re deeply embedded in staffing firms and support them with agentic high-volume workforce management, where the complexity of matching, scheduling and compliance across thousands of workers requires domain-specific intelligence that improves with every placement.

AI-native services: The entire professional services world is being turned on its head by AI. Instead of selling AI tools to professionals, AI-native services companies handle the full workflow end-to-end. In-house certified professionals – chartered accountants, solicitors, etc. – scale to 50 to 100x the clients they could serve manually, with AI handling 90% of the volume and the human providing review, sign-off and accountability. The big opportunity is serving the 33M SMBs with fewer than 50 employees across the UK, EU and US, which traditionally have been systematically locked out of professional services due to their high cost. The winners will start narrow (accounting for construction, compliance for healthcare), accumulate cross-client data within a vertical that makes them better for each new customer, and expand horizontally into adjacent services. Central is an autopilot that runs payroll, rather than giving you software to figure it out yourself.

Organisational capacity

Organisations and workforces need to be capable of adopting AI. We’ve written about this at length before. Three opportunity areas stand out.

Services and implementation: Deploying AI in an enterprise requires process redesign, data plumbing, and quality assurance. Today, a lot of this is done by the big consultancies and systems integrators. But there’s an opportunity for specialist firms that focus exclusively on AI implementation and can move faster, charge less and build reusable playbooks across clients. The best of these will look more like product companies over time, packaging their implementation knowledge into software.

AI adoption and training: Most employees have figured out that ChatGPT can help with a draft email, few have actually redesigned their own workflows around AI. This is a skills gap, and it will only get wider as AI moves from chat-based tools to agentic systems. The EU AI Act’s Article 4 mandate for AI literacy makes AI training a compliance requirement. The winners here will build training that is personalised to each user’s role, tasks and IT environment, and embedded in the flow of work — rather than delivered as a standalone course.

Measuring AI ROI: Even when the training is done and the systems are deployed, organisations struggle to measure how AI is actually being used and whether it’s driving results. There is a massive opportunity to build ROI tools that track and categorise usage and make the ROI concrete and visible to leadership. Mendo is a company that spans both AI adoption and AI ROI.

Where next?

These categories sound like neat delineations but the truth is that they all bleed into each other.

• Legacy systems of record are building agent authoring tools to entice users to stay in the ecosystem;
• Companies that are building the context layer are most likely also addressing the governance layer; both will be tempted to replace the system(s) of record they overlay on;
• Specialised AI solutions are storing work artefacts and organisational memory that could make them systems of record in their own right; and
• AI Adoption tools started with in-the-flow-of-work training, but are now increasingly moving into governance and transparency for AI agents.

For founders, three things follow…

• First, pick your entry point based on where you can build a genuine advantage
• Second, design knowing your neighbours are coming. Build defensibility into your model through economic and technical moats
• Third, expand. The biggest winners in this space will own multiple elements of the AI diffusion stack.