Where AI earns
Most companies have begun to deploy AI within their organisations. Few, however, can show what it has paid back. We believe the returns from AI depend on how well companies can put their proprietary context to work.
By James Hancock and Jan-Willem Van Lancker
for the Vantage team
Summary
- 01AI adoption within companies has not yet translated into AI returns.80% of AI users say their productivity has improved, but only 37% of companies report any EBIT impact, and just 6% attribute 5% or more of their EBIT to AI.
- 02Most companies automate tasks, but the returns lie in end-to-end workflows.AI assistants and point solutions speed up single process steps. Automating end-to-end workflows that cut across teams and source systems is what changes the P&L.
- 03Reaching those workflows requires giving AI your company’s context.Agents running whole workflows need to know how your company works. General AI models don’t, so they need a context layer: a record of how you work that AI can read.
- 04Most of your company’s context has never been written down.Context is in your systems (CRM, ERP, etc.), your documents, and most importantly, in your people’s heads. To capture what they know, you have to ask the people who do the work.
- 05Your context must stay yours, current, and seen only by those who need it.When context is managed this way, it becomes a proprietary asset of the business that is uniquely yours, and outlasts any AI model, software vendor, or person.
- 06With a context layer in place, AI agents and systems can run whole workflows.Building on a context layer makes every following AI deployment easier and faster, because each one reuses the learnings and context the previous ones captured.
Companies are starting to deploy AI: assistants, pilots, and automations. However, they are not seeing the return in the P&L.
Source: McKinsey, The state of AI in 2026: On the road to ROI, August 2026
The models behind AI have dramatically improved in recent years. A few years ago, most people first interacted with AI as a chatbot that answered questions. Today, AI agents are starting to take on real work, whether that is in coding, market research, customer service, or other activities.
Despite the improvements in intelligence, AI adoption inside companies has moved more slowly. Many companies have rolled out an AI assistant to their employees (such as ChatGPT, Claude, or Copilot), set up an AI champions programme to get initial traction, and started their first pilots. These are the right first steps in the AI deployment journey.
People using AI tools say that they work, with 80% of AI users stating that it has improved their individual productivity. However, the effect on the company is much smaller, with only 37% reporting any impact on EBIT from AI, and just 6% attributing 5% or more of their EBIT to AI and calling the impact significant.
AI adoption within companies is underway. The harder question is how companies can turn AI spend into P&L impact.
Most companies are focusing ‘below the line’, where returns are lower. The larger returns lie ‘above the line’, in end-to-end workflows.
When we look at where AI can add value and automate work within a company, we think of two types of complexity. The first is how complex the workflow is: from one task (e.g., a simple email) to a workflow that crosses many teams (e.g., order-to-cash). The second is how many information sources you need to access (e.g., from one place, such as Outlook, to many source systems to complete an end-to-end process).
Most companies are early in their AI deployment journey and focusing on ‘below the line’ automation. They are providing AI assistants to their employees, buying specific point solutions, or waiting for incumbent software vendors to add AI features on top of their existing offering. These tools help, but each only improves a step of the process that runs the same way, and therefore benefits rarely reach the P&L.
Let’s consider a sales quotation. AI can draft one in minutes, but it then waits for approvals from Finance, Operations, and the Sales Manager. The individual task is faster, but the process takes just as long.
We believe the larger returns from AI exist in ‘above the line’ use cases, with automation of end-to-end workflows redesigned around AI agents. However, jumping from AI assistant roll-out to agent-orchestrated workflows is inherently difficult due to impacts on people, processes, and technology.
Realising ROI from AI means ‘crossing the line’: redesigning end-to-end workflows around AI agents. That needs AI specific to your company.
Automating workflows requires AI that knows specifically how your company works. A context layer gives it that knowledge.
Most companies start their AI deployment journey in ‘single-player’ mode. Each person uses AI for their own steps, provides the knowledge it needs, and decides when to call it. Moving above the line means ‘multiplayer’ AI: agents that run the workflow end-to-end, not left to individuals who happen to be using them. This requires more centralised AI coordination by the company, including what knowledge is given to AI agents.
General AI models know almost everything about the world, but almost nothing about how your business actually runs. To take on end-to-end work, you need to provide them with context: proprietary knowledge on who you are, what you do, and how you do it. Your company has built this knowledge over years of operating, and it is your unique advantage.
Ask an AI assistant to prepare a quotation for a long-standing client, and it will price the order like any other. Nothing it can read says that this client’s six delivery sites and hand-relabelled pallets cost more than the standard list price covers. The quotation looks right, but every order loses margin.
We believe the return from AI comes from giving end-to-end workflows to AI agents. That is only possible if the agents have context on how your company runs. A shared context layer (e.g., “Company Brain”) is what enables this and is the first step needed to cross the line.
AI models can only provide publicly available general intelligence. Company-specific intelligence is needed for real work to be done.
Context is an encoding of how your business works, drawn from your systems, your documents, and your people’s heads.
Context is the encoding of your company: who you are, what you do, and how you do it. This proprietary context can then exist in a layer that AI can read. Creating a context layer means sourcing information from three locations.
The first is your systems: your CRM, ERP, HRIS, and more. These structured records capture the company’s operations, and are continually updated. These are the systems of record and should stay as such.
The second is your documents: policies, contracts, process notes, emails, written conversations. They hold more of how a company works, but are scattered across locations and can quickly go stale. These are not necessarily the best source of truth.
The last one is the most important: inside your people’s heads. This is where the true workflow exists, where the process exceptions, judgement calls, and reasons behind past decisions live. It has never been written down or documented, because the people who know it never needed to. And knowledge held by a few people is fragile: when they move on, the company loses it.
Go back to the quotation. The CRM shows the client’s details, and the ERP shows their past orders and six delivery addresses. Neither shows that the warehouse relabels every pallet by hand for them, or what that costs. Only the logistics lead knows, because the warehouse has always done it that way, and nobody in Sales asked. Companies hold hundreds of facts like this, one for each client, product, and exception, and few of them are ever written down.
Capturing this context means asking the people who hold it: the employees on the front line as well as the management team. What they describe becomes the map of how the company really operates, and is the first version of the context layer.
AI can only safely take on work that your company has written down in a form it can read.
The systems around your context can be rented, but your context layer itself needs to remain solely yours.
Once AI has access to a context layer it can read, three questions decide whether it keeps its value and compounds it.
Who owns it, and who powers it? We believe in context sovereignty: your context belongs to you, in a form you can move, separate from the technology that manages it and reads it. The platform that manages it and the models that read it can be rented, and swapped out as better ones arrive, but the context itself should remain solely yours. Owned this way, context becomes an asset of the business: it stays as people, vendors, and models change. It should live in one shared place, because separate context for every tool means rebuilding the same knowledge and copies drifting.
How does it stay current? Your company changes every day, and its context has to keep up. Each year some knowledge is added and approved, some is reviewed and carried forward, and some is retired. That is as much a people question as an engineering one: who suggests a change, who approves it, and how often the whole is reviewed.
Who can see what? Context describes how your company runs, but not everyone should see all of it. Procurement may not need HR’s context, and HR may not need Procurement’s. The same applies to AI: each assistant, agent, and system should reach only what its task needs.
Context that stays yours, stays current, and reaches only those who need it becomes more valuable over time.
What context enables: AI agents & systems built on top, where every deployment becomes easier and faster.
Once AI knows who you are, what you do, and how you do it, the work it can take on changes.
Workflows can be redesigned around AI agents, with people spending their time on judgement, on customers, and on the difficult cases. In the quotation, an agent checks the price and what this client costs to serve, and prepares the quotation. The sales manager approves only the exceptions, and the salespeople spend their time with clients.
Custom AI systems become possible too: software built for your organisation alone, starting from what your business knows. They can take on steps that no off-the-shelf tool could, because no off-the-shelf tool knows your company.
We believe the return compounds. Each workflow built on your context makes the next one faster and cheaper to build. The pricing rules and cost to serve behind the quotation serve order-to-cash next, and the client history serves customer service after that.
That is where AI earns: in whole workflows, built on context you own.