Measuring Knowledge in the Agentic Enterprise: What Agentforce and the Workforce Mean for the Annual Report

Following the cloud era, enterprises got comfortable calling data "the new oil," or "the new gold," a raw asset to be extracted, refined, and mined for competitive advantage. The metaphor was directionally right about where value was migrating, but it obscured a stranger fact: companies were amassing this new form of wealth without ever putting it on a scale.

Ocean Tomo's long-running Intangible Asset Market Value study shows just how far that migration has gone. In 1975, intangible assets, patents, brands, data, software, know-how, accounted for just 17% of S&P 500 market capitalisation; tangible assets like factories and equipment made up the rest. By 2005 that had flipped to 79% intangible. By 2025, intangible assets represent roughly 92% of S&P 500 market value, with physical assets down to a residual 8%. Ocean Tomo calls this a "wholesale transformation in the nature of value creation whereby economic worth has migrated from what can be 'touched' to what can be 'thought.'"

And yet the accounting and reporting apparatus enterprises use to describe themselves to investors, regulators, and boards has barely moved. Property, plant, and equipment gets depreciated on a schedule, insured at 55% of exposure, and itemised on the balance sheet down to the forklift. According to Aon's 2026 Intangible vs. Tangible Risks Comparison Report, information assets, the modern proxy for knowledge, have now edged past PP&E in average reported value ($1.49 billion versus $1.40 billion), but only 20% of that exposure carries insurance coverage, with roughly 60% self-insured and largely undocumented. Knowledge became the majority of enterprise value before anyone built the instruments to measure, protect, or report it.

If data was the gold, AI, and specifically the current wave of enterprise AI agents, is increasingly framed as the refinery, or the fire that finally puts the gold to use. As one widely circulated extension of Clive Humby's original "data is the new oil" line puts it, if data is the new oil, generative AI is the new rocket fuel: data is inert without something to convert it into usable output, and AI is the mechanism that does the converting, at a speed and scale no prior generation of BI tooling could match.

An accelerant is a useful word for what that means in practice. It doesn't create the reserve of fuel; it burns through it faster, and it makes any miscalculation about how much fuel you actually had far more visible, far sooner. That is the mechanism by which AI adoption forces measurement where the cloud-data era never did: enterprises could stockpile unmeasured data for a decade with no consequence, because nothing was consuming it fast enough to expose the gap. AI agents consume institutional knowledge continuously and at volume, which means an enterprise now finds out, in weeks rather than years, whether the knowledge it assumed it had was ever actually there.

That framing matters because it clarifies what's actually new. Enterprises have collected data for decades. What's changed is that AI agents, deployed at scale under names like Salesforce's Agentforce and its many enterprise equivalents, are now the thing querying, synthesising, and acting on institutional knowledge in real time, often standing in for the humans who used to hold that knowledge in their heads. That shift turns "knowledge" from an abstract cultural asset into something with a measurable operational footprint: conversations resolved, decisions made, errors introduced, and, critically, knowledge lost when the humans who held it leave before the AI has actually absorbed it.

Salesforce is the most visible test case, but the same pattern is showing up across every major enterprise survey of 2026. Deloitte's State of AI in the Enterprise report found that "search and knowledge management" ranks among the generative AI use cases enterprises expect the most impact from, right alongside virtual assistants and content generation, which is another way of saying that organisations already intuit knowledge as the asset AI is meant to unlock, even without a line item for it. Yet Deloitte also found a wide gap between ambition and delivery: 74% of organisations are hoping to grow revenue through AI, versus just 20% who currently are, with the more concrete gains so far concentrated in productivity (two-thirds of respondents) and cost reduction (40%) rather than the knowledge-driven transformation many expect. Deloitte's own framing of what separates the two groups is blunt: "governance is the difference between scaling successfully and stalling out," with organisations where senior leadership directly shapes AI governance reporting significantly more business value than those that leave it to technical teams alone.

Forrester's 2026 predictions suggest that gap is about to get an accounting mechanism. The firm expects the leading human capital management platforms to add "digital employee management" capabilities this year, turning HR technology into what Forrester calls "the sophisticated employee system of record, tracking and optimizing a hybrid workforce" of humans and AI agents together. IDC's parallel forecast, that 40% of all Global 2000 job roles will involve working alongside an AI agent by the end of 2026, implies that a large share of enterprises will need exactly this kind of record within the year, whether or not their finance function is ready to use it. Put plainly: the workforce systems that already feed human-capital disclosures are being rebuilt, right now, to also track the agents doing an increasing share of the knowledge work, which means the data enterprises will eventually need for an annual report is being generated years before anyone is required to report it.

Salesforce's own experience with Agentforce offers both the optimistic case for AI as an accelerant and the cautionary tale about what happens when you accelerate before you've measured what you're running on.

In its more favourable telling, Salesforce describes what it calls "the agentic enterprise": AI agents and humans working side by side rather than AI simply displacing headcount. By the company's own account, Agentforce has handled 2.6 million customer conversations, resolving 63% of questions with satisfaction scores comparable to human agents, while sales agents booked hundreds of meetings and surfaced millions of dollars in pipeline. CEO Marc Benioff has argued that "the Agentic Enterprise is not only a technology transformation. When humans and agents work together, it radically transforms the structure of your company," pointing to entirely new roles created in the process, such as AI conversation designers and deployment strategists, alongside redeployment of support staff into higher-value work like adoption and retention.

The other half of the story is less flattering, and more instructive for anyone thinking about knowledge as an asset. In 2025, Salesforce laid off roughly 4,000 of its 9,000-person customer support organisation, betting that Agentforce could absorb the difference. The company later reversed course and began rehiring, acknowledging, in the account of outside analysts, that it had overestimated AI's problem-solving capacity while undervaluing what its own people knew. The damage wasn't just to headcount plans: the layoffs eliminated institutional knowledge that had never been captured anywhere the AI could query it, eroded internal trust, and dented customer confidence after the company had marketed AI as a wholesale replacement for human support. The lesson distilled by outside advisers is blunt: address your operational and data debt first, preserve institutional memory deliberately, and verify actual efficiency gains before you assume the AI has learned what the departing employee knew.

Put the two halves together and the pattern is clear: Agentforce-style deployments don't just consume enterprise knowledge, they expose how little of it was ever actually inventoried. The 63% resolution rate is a measurement. The 4,000 layoffs and subsequent rehiring is what happens when an enterprise tries to act on knowledge it never measured in the first place.

The clearest illustration of why this measurement gap matters didn't come from a tech company: it came from one of the firms enterprises pay to advise them on exactly these questions. In June 2026, KPMG withdrew a report titled "Total Experience: Redefining Excellence in the Age of Agentic AI," its own analysis of how organisations were adopting autonomous AI, after the AI-detection firm GPTZero found that only 5 of its 45 citations were real; the other 40 were fabricated, misattributed, or too vague to trace. The report also misstated KPMG's own survey data, citing 55% of CEOs as prioritising AI when the underlying KPMG survey had actually found 71%, and organisations named as case studies, including UBS and the NHS, disputed the claims attributed to them. A GPTZero researcher's assessment was blunt: "no human at KPMG checked the citations, the claims, or the sources before the report went out." It is hard to find a cleaner demonstration of this article's thesis: a firm published a document about the value of agentic AI without ever measuring whether the knowledge inside it was real.

The same failure mode is compounding in the courts, at a scale that's becoming its own dataset. A public tracker of AI-hallucinated legal citations counted 1,598 documented US cases by June 2026, up from roughly 200 a year earlier and growing at about eight new cases a day. The penalties are no longer symbolic: a federal magistrate in Oregon imposed roughly $109,700 in sanctions in one case over fifteen fake citations and eight fabricated quotations spread across three briefs, and a Sixth Circuit panel assessed $15,000 per attorney, plus opposing counsel's fees and costs, in another. US courts imposed more than $145,000 in AI-filing penalties in the first quarter of 2026 alone. None of this is knowledge management in the abstract: it's institutional knowledge (what a source actually says, what the record actually shows) failing to get verified before it was acted on, at a cost that currently shows up on an invoice or a sanctions order rather than a balance sheet.

If knowledge is now the dominant form of enterprise value, and AI agents are now actively consuming and producing it, you might expect annual reports to reflect that. They largely don't, for a specific and fixable reason: US disclosure rules were built around a different kind of capital.

Item 101 of Regulation S-K requires public companies to disclose human capital resources and any measures or objectives management uses to manage the business, but the rule is principles-based, with no mandated metrics. Gibson Dunn's five-year survey of S&P 100 Form 10-K filings found that human capital disclosures actually got shorter in 2025: roughly 85% of surveyed companies trimmed their disclosures, driven largely by the wholesale removal of diversity-related content, with the terms "DEI" and "DE&I" disappearing entirely from S&P 100 filings that year. Companies pivoted towards softer language about "merit," "belonging," and "engagement" instead. Notably, that same survey found no meaningful discussion of AI, knowledge management systems, or intangible intellectual assets inside human capital disclosures at all. The section of the annual report closest to "how does this company create and retain knowledge" is currently shrinking rather than expanding to meet the moment.

Risk disclosure has moved faster than value disclosure. White & Case's guidance for the 2026 annual reporting and proxy season notes that more than 85% of Fortune 100 companies already discuss AI-related risk factors, with 36% now maintaining standalone AI risk sections covering regulatory uncertainty, cybersecurity exposure, system failure, and competitive disadvantage. But the same guidance warns companies to "avoid overstating AI benefits and assess whether claims could mislead investors," a caution sharpened by an SEC enforcement action against a company that overstated its AI capabilities. In other words, enterprises are now required to tell investors how AI could hurt them, while having no comparable obligation, or even a settled methodology, to tell investors what their AI-augmented knowledge base is actually worth.

That gap is starting to close, from an unexpected direction. In its Q1 2026 exposure draft, the Financial Accounting Standards Board proposed the first US GAAP framework specifically governing AI used in material financial reporting processes, covering things like revenue recognition, credit loss allowances, goodwill impairment testing, and warranty reserve estimation. Under the proposal, companies would need to disclose, qualitatively, the nature and purpose of AI use in each material reporting process, the type of model and data inputs involved, and the human oversight applied to its outputs; quantitatively, they'd need to provide sensitivity analysis showing how AI-generated estimates respond to key assumptions, and compare those estimates to actual outcomes where available. The rule is slated to take effect for fiscal years beginning after December 15, 2026, meaning calendar-year filers would need documentation in place for first-quarter 2027 filings, with AI systems touching internal controls over financial reporting subject to the same SOX 404 testing rigour as any other material control.

This is narrower than a true "knowledge asset" disclosure regime: it governs AI used to produce numbers on the financial statements, not the broader question of what an enterprise's institutional knowledge is worth. But it establishes a precedent that matters: for the first time, regulators are requiring companies to show their work on how AI arrives at a judgement, and to compare that judgement against reality. That is, functionally, the beginning of a measurement discipline for knowledge-derived output. It is a reasonable bet that human capital and intangible-asset disclosure rules follow a similar path over the next several reporting cycles, especially as agentic AI takes on more of the judgement work that used to live exclusively in people's heads.

Enterprises don't need to wait for a mandate to start treating knowledge as a measured asset rather than an assumed one. The technical building blocks already exist, largely because they were built to make AI agents reliable rather than to satisfy accountants, but the two purposes turn out to converge.

Enterprise knowledge graphs, structured maps of entities, relationships, lineage, and ownership that an AI agent can query to understand a business, are becoming the operational substrate for this. Gartner projects that more than 50% of enterprise AI agent systems will rely on graph-based context by 2028, and for good reason: research cited by knowledge-infrastructure vendors attributes most enterprise AI agent failures to missing or stale context rather than to weak underlying models, and shows graph-grounded retrieval cutting hallucination rates by more than 40% compared with naive retrieval. The metrics that matter here look less like a traditional knowledge-management maturity model and more like an audit checklist: whether definitions and lineage stay current as systems evolve (context freshness), whether the system can reason across multiple connected entities correctly (query performance), whether governance policy is enforced before an agent acts (governance integration), and whether the same entity is reliably recognised across different systems (entity resolution accuracy).

That checklist doubles as a reasonable starting point for what a knowledge-asset disclosure might eventually look like in an annual report: how much of the organisation's operational knowledge is captured in a system an AI agent (and a new employee) can actually query, how current that knowledge is, what oversight governs it, and how its outputs compare against ground truth. Enterprises that build this inventory now get two things at once: a defensible answer when FASB-style disclosure obligations eventually broaden beyond financial-statement AI, and a hedge against the version of the Salesforce mistake where an organisation cuts the humans holding its knowledge before confirming the AI actually holds it too.

The hallucination cases above, KPMG's fabricated citations, the 1,598 sanctioned legal filings, are the visible failure mode: an AI system inventing something from nothing and getting caught. There's a second, quieter failure mode that the same measurement gap enables, and it's arguably more dangerous because it doesn't announce itself as an error.

It starts with a piece of unverified human speculation: a rumour in an internal Slack channel, a leaked draft memo, a single unconfirmed report that a CEO is stepping down. If an enterprise's AI agent ingests that claim through its knowledge graph or retrieval layer without any attached provenance, source-confidence, or timestamp, the agent has no way to distinguish "reported by one anonymous account, unconfirmed" from "verified in the company's own HR system." Ask the agent whether the CEO is leaving, and it can answer with the same flat certainty it would use for a fact drawn from a filed 8-K.

That's where the loop closes. Employees, journalists, or even other companies' procurement and research agents querying that system get the claim back not as a rumour but as institutional knowledge, sourced, apparently, from the company's own systems. Repeat that exchange a few times, agent to human, human back into a document an agent later indexes, agent to agent, and the hedge language ("sources say," "unconfirmed") tends to fall away at each hop while the confidence attached to the claim goes up. What began as one unverified claim becomes, functionally, consensus reality, not because it was confirmed, but because it was repeated by something that sounded authoritative. Worse, a later correction or denial rarely gets re-indexed with the same speed or prominence as the original claim did, so the false version can outlive the correction inside the very system built to be the source of truth.

This is a different problem from the KPMG and legal-citation cases. Those were AI inventing a source that never existed. This is AI amplifying and certifying a source that did exist, but was never verified, at the exact moment enterprises are wiring these systems into the workforce and, per Forrester and IDC's 2026 forecasts, treating agents as system-of-record contributors alongside employees. For a claim like an executive departure, the exposure isn't only reputational: premature or false personnel information can move markets, implicate Regulation FD, and create the kind of "AI capabilities" overstatement risk White & Case has already flagged regulators are watching for.

The mitigation is an extension of the same knowledge graph discipline already discussed, not a new stack. Provenance has to travel with the fact, not just the fact itself, so every claim in the graph carries its source, a confidence or verification status, and a timestamp, distinguishing "verified in HR system" from "reported once, unconfirmed." Confirmed changes must propagate, and so must retractions; a knowledge graph that adds a contradicting node without demoting or timestamping the original claim just gives the agent two facts to average between rather than one corrected one. And the same periodic ground-truth comparison FASB's exposure draft already contemplates for quantitative AI estimates, checking output against actual outcomes on a cadence, applies just as well to qualitative claims the agent is repeating as fact. An enterprise that can show it audits for this kind of drift, not just for invented citations, has a real answer when a regulator, journalist, or board member asks how it knows its AI-mediated knowledge is actually true, rather than merely repeated.

Three practical moves follow directly from where the data and the disclosure landscape both point. First, treat institutional knowledge capture as a prerequisite for any workforce restructuring tied to AI agent deployment, not a byproduct of it. Inventory what a departing team actually knows before assuming an agent has learned it, using the same context-freshness and entity-resolution discipline knowledge-graph teams already apply to production agents. Second, get ahead of FASB's 2026 exposure draft by documenting AI use in material financial reporting processes now, including model type, data inputs, human oversight, and sensitivity of AI-generated estimates to key assumptions. The SOX 404 control-testing implications alone justify starting before the fiscal-year-2026 effective date. Third, use the widening gap between shrinking human capital disclosures and rapidly growing AI risk-factor disclosures as a strategic opening: a company that can credibly describe how it measures, retains, and governs institutional knowledge, not just the risks AI poses to it, has a genuine differentiator in a reporting environment where 92% of enterprise value is already intangible and the accounting is only now starting to catch up.

Data was the gold everyone knew they were sitting on but rarely weighed. AI agents are now spending that gold at agentic speed, whether or not the balance sheet, or the annual report, has caught up to what's actually being drawn down.