The returns from enterprise AI are not spread evenly. They gather in a few places, and those places share a structure.

The economics of enterprise AI are usually discussed as though value were spread evenly across the org chart, waiting for whichever team adopts a tool first. Our work points the other way. The returns gather in particular places, and the places have a shape in common, identifiable by three conditions:

1. The first is that decisions are assembled from documents and tacit judgement together, so neither a database nor a model alone can reproduce them.

2. The second is that senior bandwidth is the binding constraint, and the cycle moves only as fast as a few experienced people can attend to it.

3. The third is that a missed obligation or a piece of lost knowledge is genuinely expensive, whether the exposure is regulatory, contractual, or the quiet attrition of the people who carry the operation in their heads.

Where one condition holds, software helps at the margin. Where all three coincide, the available compression is large and durable, and most of it is still unclaimed. Four sectors meet all three.

Manufacturing: costing, evaluation, compliance

A supplier receives a document set from its customer: drawings, technical descriptions, testing requirements. Engineers turn that set into an evaluation, the evaluation into routing and process decisions, and those decisions into a cost sheet, drawing on rates held across purchase teams, supplier relationships, and memory. The reasoning that joins these steps rarely exists anywhere outside the people performing them. Process costs are estimates calibrated over years, with no formula behind them. Routing decisions are made the way they were made last time, by the person who made them last time.

Two decades of ERP did not touch this, because ERP records outcomes. The reasoning between input and outcome was never captured, so every evaluation starts close to zero and every retirement is a quiet write-off of an intangible asset. The compliance side carries the same structure with a harder edge. Global retail customers maintain supplier codes and living document sets, updates arrive continuously, and a missed change or an expired certificate can stop dispatch overnight.

When the assembly work moves to a system that organises the documents, asks for what is missing, and drafts the evaluation and the costing for senior review, three things happen at once. The cycle compresses from weeks towards days, sometimes minutes. The exposure narrows, because deadlines and obligations stop depending on anyone's memory. And a structured record of how the company evaluates, costs, and decides accrues as a consequence of the work itself, with nobody asked to document anything. The first outcome pays for the engagement, the second is insurance, and the third is the asset.

Financial services: the non-core load

In regulated distribution, the scarce resource is the relationship. A mutual fund distributor, or the client-facing team at an asset manager, spends a striking share of its hours on product education, routine portfolio queries, onboarding mechanics, and status questions. None of this needs the judgement the relationship is built on, and all of it competes with it.

The reason this persisted is regulatory, and it is legitimate. Under SEBI and AMFI frameworks, the line between describing a product and advising on it carries consequences, so generic conversational tools were correctly treated as a liability. The work stayed manual because the boundary could not be trusted to software.

That calculus changes when the boundary becomes an engineering property. A system can describe, inform, and facilitate while stopping short of advice, hold the institution's voice in every exchange, and log everything it says in a form a compliance function can inspect. Capacity per relationship manager rises. The brand stops fraying at its most numerous touchpoints. And the institution gains something it has never had, a governed record of what clients ask, where they hesitate, and how the answers land. Until now that record sat scattered across inboxes. It now belongs to the firm.

eCommerce: the purchase conversation

Baymard Institute's long-running aggregation of checkout studies puts average cart abandonment near seventy per cent. Behind the figure sits a simpler observation. Most online stores can display products, and most chat tools can answer questions, yet almost nothing on the page can conduct a purchase the way a capable shop-floor person does, holding the thread from discovery through comparison to decision and staying present after the sale.

This persisted because the conversation sat in the support budget, and support budgets exist to be minimised. The craft is also unusual. It lies where persuasion, brand, and engineering meet, and teams that hold one of the three rarely hold the others.

Treated as a sales function and engineered accordingly, the conversation recovers revenue the traffic budget has already paid for. It carries the brand's own register through every exchange. And it produces a record no analytics suite can, the questions buyers asked, the points they hesitated over, and what resolved them. This is merchandising intelligence in its rawest and most honest form.

ESG reporting: the assurance problem

BRSR is the clearest case of a deadline doing the market's persuasion. Since FY 2022-23, India's top 1,000 listed companies have filed Business Responsibility and Sustainability Reports. SEBI's July 2023 circular then introduced BRSR Core, a set of KPIs requiring independent assurance, phased from the top 150 companies in FY 2023-24 to the full top 1,000 by FY 2026-27, with value-chain disclosures phasing in alongside. The widest tier of that glidepath is now immediate.

Assurance changes the nature of the work. A disclosure can be drafted. An assured KPI has to be traceable to source records held across HR, EHS, facilities, procurement, and finance, and most of those records live in spreadsheets maintained by people doing the work in the margins of other jobs. The annual filing is an assembly problem wearing a reporting costume, and it returns every year with higher stakes.

A system that holds the framework, gathers evidence from its owners, follows up where people forget, and assembles the filing for review turns the season of panic into a process. The by-product compounds, since this year's evidence trail is next year's starting point. And the value-chain provisions steadily pull a company's suppliers into the same discipline, which is where this lane meets manufacturing.

The combination, stated plainly

It is the combination of competences that decides whether the compression is captured. The technology layer is now available to everyone. What is done with it inside a real operation is where the difference is made, sector by sector, the same architecture applied to different rooms.

Each sector above has its point solutions. Costing tools exist; compliance trackers exist; chat widgets exist; ESG platforms exist. What the three conditions actually demand is a set of competences the market almost never holds in one team.

• Operational context. Knowing where the decision chain actually runs, which is rarely where the org chart says. This is learned in plants and back offices, never in demos.

• People context. Deployments fail on adoption and on measurement long before they fail on capability. Trust has to be designed into a system from the first week. No training programme retrofits it.

• Interface discipline. Conversations and surfaces that people still choose to use in month six. This is a design problem with engineering consequences.

• Technical depth. Systems that hold in production, under audit, at the boundary regulation draws.

Any one of these can be hired. The combination is rare, because careers do not usually cross these boundaries.

Mitochondria builds this.