Most enterprise AI starts life as a tool. A system that prepares quotes. An assistant that answers customer questions. An agent that processes documents. The value is clear: the work gets done faster, more cheaply or more consistently than before.
That is where most deployments stay. MIT's 2025 research found that only 5% of custom enterprise AI tools reach production, and many of those that do are still judged on task counts rather than business results. The value that would justify wider investment is never measured, so it is never claimed.
A useful way to see the path beyond the tool is to think in stages. AI value develops in three stages: the tool, the outcome and the strategic relationship. Each stage asks a different question, uses different measures and creates a different kind of value. Organisations that understand the stages make better investment decisions. Those that do not often conclude, wrongly, that AI has underdelivered.
Why measured gains arrive late
In 1987, the economist Robert Solow observed that computers could be seen everywhere except in the productivity statistics. The paradox resolved itself a decade later, once organisations had reorganised their work around the technology.
Brynjolfsson, Rock and Syverson describe the same pattern for AI as a productivity J-curve. General-purpose technologies need large investments in intangibles: new processes, new data, new skills and new ways of organising work. Those investments are rarely counted as assets, so measured productivity can even dip early on. The gains arrive later and are larger once the intangible investment pays off.
The practical lesson is that the first months of an AI deployment are not a fair test of its value. They lay the foundations. Judging a system only by its first-quarter task counts is like judging a factory by its construction costs.
The tool stage: necessary, but limited
At the tool stage, a system is judged by operational measures: items processed, time taken, error rate. Those confirm that the system works. They do not show whether it creates value.
Tool-stage thinking has three limits. Value is measured as cost saved on work already being done, which caps it at efficiency. Tools are compared on price and features, which keeps the conversation at procurement. And a tool that handles a thousand transactions has done a thousand transactions: it may have improved its own performance, but the organisation has not necessarily learned anything about its operation, its customers or its opportunities.
There is nothing wrong with starting here. A visible, well-run tool builds trust. The mistake is stopping here.
The outcome stage: measuring what matters
The shift happens when the question changes from "what does the system do?" to "what results does it produce?" Kaplan and Norton's balanced scorecard offers a helpful distinction between lagging measures, which report results, and leading measures, which predict them. Operational measures such as turnaround time are leading indicators. Outcome measures such as revenue or audit findings are lagging ones. A good measurement plan needs both and needs to show how they connect.
What this looks like by workflow:
Quoting: conversion from quote to order, how response time relates to win rate, and revenue from enquiries that previously never received a quote.
Compliance evidence: audits passed without findings, expiries caught in time and hours returned to the team.
Sustainability disclosures: reporting cycle time, assurance queries per disclosure and cost per disclosure.
Customer conversations: resolution without escalation, satisfaction and repeat purchase.
Property operations: requests closed within service levels, repeat faults and tenant renewals.
Outcome measures are harder to set up, because they often need tracking that didn’t exist before the system arrived. A manufacturer that has never connected quotes to orders cannot measure conversion until it does. That work is worth doing. It shows where the next gain lies, and it is part of the intangible investment the J-curve describes.
The strategic stage: insight about the business
Once outcomes are measured, a third kind of value appears. Michael Porter's value chain reminds us that every activity sits within a larger sequence. A quoting system touches one link: lead generation, qualification, quoting, follow-up, negotiation, order capture and fulfilment all sit around it. The data a quoting system creates speaks to all of them.
A system that has handled thousands of cases holds a record of how the business performs: which leads convert, which products sell at which margins, where processes stall, which exceptions recur. That record raises questions well beyond the system's own workflow. How should lead qualification change, given what conversion data shows? What follow-up cadence works for which quote types? How should the product catalogue be structured to quote faster and communicate more clearly?
These questions are about the organisation's operation, not the AI. They arise from the visibility the system creates, and they need an understanding of the work that comes from being inside it. At this stage, the relationship with a provider changes. The conversation moves from features to priorities: which workflow to tackle next, what the patterns suggest and how governance should evolve. The organisation keeps ownership of its data, configuration, and decision memory throughout. It gains a partner that understands its operation.
Why many AI relationships stay at the tool stage
Tools are easier to buy. "A quoting system" is a clear deliverable. "Better lead-to-revenue performance" is harder to scope, price and approve.
Measurement infrastructure is missing. Outcomes cannot be shown until systems are connected and processes are standardised enough to track.
Trust takes time. An organisation will not invite a provider into operational strategy after one project. Trust grows through delivered commitments and visible understanding of the business.
Incentives may not align. A provider paid only for usage has little reason to help the client need less of it. Outcomes need a model where the provider benefits when the client succeeds.
Questions to ask any AI provider
Does the provider start with questions about your operation, or with a product demonstration?
Will it help establish outcome measures and baselines before go-live, not only operational ones?
Who owns the data, configuration and decision memory the system builds?
Is its commercial model tied to your outcomes, or only to usage?
Does the relationship include regular reviews of what the data shows, beyond support tickets?
Outcome measures to agree before go-live
One business outcome per workflow, such as conversion, audit findings or resolution rate, with a baseline.
The leading indicators that predict it, such as turnaround time or first-time-right rate.
How both will be tracked, including any new data the system needs to record.
When they will be reviewed: monthly at first, then quarterly, with the J-curve in mind.
Who owns them, on the business side, not only in IT.
What would trigger a change to the system, the workflow or the scope.
How Mitochondria's agentic AI systems are measured
The approach. Every deployment begins by mapping how the work runs, which also shows what can and cannot be measured today. Outcome measures, leading indicators and baselines are agreed before go-live. Each unit records the data needed to track them, in the organisation's own environment.
What it enables. Results are reviewed against business outcomes from the first months, with realistic expectations of when gains will show. Decision memory and the record of every case reveal patterns across the wider value chain, so the next priority is chosen on evidence.
The impact. AI that is justified by what it produces for the business, and a growing understanding of the operation that belongs to the organisation.
Read more about Mitochondria's approach.
Sources
Brynjolfsson, E., Rock, D. and Syverson, C. (2021). The productivity J-curve: how intangibles complement general purpose technologies. American Economic Journal: Macroeconomics, 13(1).
Solow, R. M. (1987). We'd better watch out. New York Times Book Review, 12 July.
Kaplan, R. S. and Norton, D. P. (1996). The Balanced Scorecard: Translating Strategy into Action.
Porter, M. E. (1985). Competitive Advantage: Creating and Sustaining Superior Performance.
—
Mitochondria builds agentic AI for evidence-led core operations. Research-led and model-agnostic. ISO/IEC 27001:2022 certified, with controls aligned to the GDPR, UK GDPR, India's DPDP Act and Australia's Privacy Act. Amsterdam and Pune. Request a working session →