No Process Context, No Enterprise Context, No Enterprise AI
By Nanda Gottumukkala, Co-Founder of xflow and AIMG expert Network Member
Why the context layer is only as good as the process underneath it
Ask five teams in one bank “What is our exposure to Acme Shipping?” and you get five confident answers: £500m, £410m, £385m, £305m and a question mark. Same client, same day, a £195m spread. None of them is wrong.
The example comes from xflow’s paper Context is not a thing. It’s a point of view. The Relationship Manager is sizing headroom, Credit Risk is stress-testing, Regulatory Reporting is applying a formula, Finance is matching the balance sheet. Group Risk cannot answer, because Acme also borrows through Singapore under a different name. Ask an AI agent and it will find a field called exposure and answer fluently, over the wrong facts.
My argument is blunt: without process context there is no enterprise context, and without enterprise context there is no enterprise AI.
More context is not the answer
AIMG’s Special Report The Context Layer sizes this market at $8.3bn in 2026, rising to $24.5bn by 2030. The demand is real: fewer than one in ten AI proofs-of-concept reach production, and AIMG finds that context, not model quality, is the binding constraint.
But more context is the wrong cure. AIMG cites research showing that irrelevant context beyond a point worsens reasoning; xflow puts it more sharply: a bigger window just gives the agent more candidate numbers. Context is a selection problem, and you cannot select until you know what each number is for.
Whose context?
xflow’s framing resolves this into two classes of context, and I think it is the most useful simplification in the current debate:
- Consumer intent: the question a number answers, who is asking it and for which decision.
- Producer intent: how the number is produced, its steps, rules and assumptions, and how entities are resolved.
Held together, they form process context. Separated, they decay. IT freezes consumer intent into rules (“sum approved limits where status is active”) and the why is lost. It merges five requirements into one optimised engine, and the producer logic scatters across parameters, spreadsheets and the heads of the people who built it. Humans patch the gap with tribal knowledge. Agents have none.
Entity context is not process context
This is the distinction that matters most, and the market routinely blurs it.
Entity context tells you what things are: who Acme Shipping is, how its London and Singapore records relate, which facilities it holds. It lives in master data, ontologies and knowledge graphs.
Process context tells you how a figure comes to be, and for whom: which facilities count, what share of the undrawn £180m is included, whether £15m of provisions comes off, in what order, to answer which question.
The two are complementary, not interchangeable. A perfect ontology of Acme would still let an agent hand Finance’s £305m to a Relationship Manager. And entity resolution is itself a process step: xflow’s rule is to resolve identity before you calculate. Ontology tells the agent what it is looking at; process context tells it what to do with it.
The chain, validated
Process → enterprise context. AIMG says a context layer must carry knowledge, expertise (“how the work is actually done”) and norms, and collapses without any one of them. Expertise is process context. Yet none of AIMG’s five supply segments, from meaning layers to orchestration, is labelled process. It falls between lineage and ontology.
Enterprise context → enterprise AI. Supervisors now ask, in xflow’s words, “show me how this result was produced.” The EU AI Act and ISO 42001 push toward a runtime record of what an agent did and why, and AIMG finds governance budgets now funding the layer. BFSI leads that spend because regulated institutions must produce the audit trail.
What this means
Boards should treat process logic that is documented but not provable as a governance risk: key-person dependency today, algorithmic liability tomorrow. Buyers should start with one consumer intent, name metrics by intent (“exposure – available headroom”), and stop reconciling differences that exist by design. Vendors should note AIMG’s view that intent and behavioural context are the most under-supplied classes; process context is where both get captured. Investors should, as AIMG advises, underwrite the compounding asset: an executable record of how regulated figures are produced grows more valuable every month, and a late entrant cannot buy it.
One question
Pick one number your organisation reports. Can you say which question it answers, and show how it was produced, without calling three people? If not, a bigger model or window will not help. Next time someone says “give the agent more context”, ask: whose?
Without process, context is a guess. Without context, enterprise AI is a guess made at scale.
Sources: https://xflow.solutions/insights and https://aimarkets.group/market-reports/the-context-layer/