Research and insights
A reading of the current research on enterprise AI execution, with every figure kept at its original scope and linked to its source. Updated October 2026.
Analyst research, surveys and recent enterprise reporting keep landing on the same obstacles. Repetition across sources signals a persistent problem, not a single verified prevalence rate.
80% of McKinsey's 1,719 respondents report personal productivity improvement; 37% report any positive enterprise EBIT contribution; 6% meet the high-performer definition of at least 5% EBIT impact.2 IBM's survey of 1,500 CFOs finds only 6% consider their finance function transformation-ready.3 Gartner frames the gap as an operating-cost and measurement discipline problem rather than an adoption problem.15
What it implies: an economic owner, a baseline, and a hard separation between time released and spending eliminated, capacity redeployed or margin created.
Nearly 75% of AI high performers say they fundamentally redesigned workflows, against about 25% of other respondents. This is association, not causation.2 SAP's own account of enterprise deployments describes pilots disconnected from systems and handoffs that stay manual after the demo.8
What it implies: design one process from trigger to completed outcome, including approvals, exceptions and write-back into systems of record.
Gartner reports that organizations with successful AI initiatives invest up to four times more in data, governance and people foundations as a share of revenue, in a survey of 353 leaders. It is a correlation, not a prescribed spend.7 Providence's internal-hiring case combined existing Oracle systems rather than replacing them.9
What it implies: map sources of truth, permissions and context; validate key records; build controlled retrieval and integrations before scaling agents.
In PagerDuty's survey of 1,250 office professionals, 66% reported using AI at work despite believing it was prohibited and 34% reported entering customer data into public AI tools.5 Akamai's traffic analysis flags unmanaged personal identities in enterprise AI use.6 Practitioner legal guidance recommends governing existing use rather than assuming it can be blocked.4 NIST's Generative AI Profile provides the voluntary control framework most organizations map to.13
What it implies: inventory actual usage, approve usable tools, classify sensitive data and use cases, and monitor access, outputs and high-impact decisions.
Gartner cites 81% of CIOs expecting GenAI skills gaps to obstruct their 2025 objectives.14 Reuters reports Ropes & Gray letting more than 1,000 lawyers count up to 100 hours of AI training toward billable targets, with over 35,000 associate hours logged in 2026, a change to incentives rather than to curriculum.10
What it implies: train on real tasks with approved data and realistic failure cases, assign approvers with authority, and fix the incentives that penalize learning.
About one in five McKinsey respondents say AI operating costs constrain use.2 Google's cross-application agent announcement is notable for what it ships with: dedicated agent identities, connectors, permissions and cost controls.11 IBM and Ponemon find one in four malicious breaches studied were AI-enabled, in a sample of 602 breached organizations.12
What it implies: budgets, least-privilege identities, approval boundaries, tests, audit logs and rollback paths, and a tracked cost per successfully completed case.
Operational benchmarks we use to size a workflow before an assessment. Each is a median or an average from a named study, with the comparison group stated.
Invoices and intake
$10.89
vs $2.78
Cost to process one invoice: average organization vs top performers. Top performers also post in 3.1 days instead of 10.9 and see 9% exceptions instead of 22%.17
At 50,000 invoices a year, the gap is about $400,000 and seven calendar days of float on every payable.
Reporting and close
50%
take 6+ business days
Share of finance teams whose month-end close runs six or more business days. 27% regularly take more than seven.18
Every extra close day is a day leadership decides on last month's numbers, and a week of senior finance time spent assembling rather than analyzing.
Rekeying between systems
1 to 3%
of fields
Error rate for trained staff entering data by hand, measured per field in a controlled study.19
On a 40-field record, that is an error in roughly half of all records. Each one is found downstream, by a customer, an auditor or a colleague who has to chase it.
Support and service desks
$13.50
vs $1.84
Median cost per contact on assisted channels (phone, chat, email) vs self-service.20
Most tier-1 requests are lookups against systems you already own. Each one that reaches an agent costs seven times what it needs to.
Every figure keeps its original scope and links to its source.
80% / 37%
Personal productivity gains / any positive enterprise EBIT impact. McKinsey, 1,719 respondents.2
6%
Met McKinsey's AI high-performer definition, including at least 5% EBIT impact.2
~75% / ~25%
AI high performers / other respondents who fundamentally redesigned workflows.2
40% / 27%
Large enterprises scaling AI agents in at least one function, 2026 / 2025.2
32%
Skipped a software purchase because agentic coding could build it instead.2
62% / 6%
CFOs with expanded AI responsibility / finance judged transformation-ready. IBM, 1,500 CFOs.3
66% / 34%
Policy-defying AI use / customer data entered into public AI tools. PagerDuty, 1,250 respondents.5
Up to 4×
Foundation spending among successful AI initiatives, as a share of revenue. Gartner, 353 leaders.7
1 in 4
Malicious breaches that were AI-enabled, among 602 breached organizations. IBM/Ponemon.12
One primary business metric the process owner will defend, then five supporting measures. Definitions we use, in plain terms.
Hours recovered
Staff time no longer spent on the workflow, counted from time studies or system timestamps, and what that time was redeployed to. Capacity until finance books it.
Cost per transaction
Fully loaded cost per invoice, case, report or contact: labor, review, exceptions, licenses, inference and monitoring, divided by acceptable completions.
Error and rework rate
Exceptions, corrections and escalations per hundred items, measured at the same point in the process before and after.
Processing time
Elapsed time from the triggering event to a completed, valid outcome. Not the time to produce a draft.
Revenue impact
Margin attributable to the change through faster quotes, fewer lost cases or higher conversion, against a credible counterfactual.
Run cost and risk
Inference, licenses, integration and monitoring per period, plus unauthorized actions, sensitive-data events and audit completeness.
Two outside anchors for this. Deloitte's guidance for CFOs calls for locked pre-deployment baselines and names missing baselines as the largest source of indefensible ROI claims.30 Gartner's survey of 724 respondents finds only about a third of teams report high productivity gains, and tells CFOs to challenge assumed headcount savings until workflows are redesigned.31 The UK government experiment is a usable template for measuring hours recovered: an active user is one interaction or more, adoption is active users over licences, and time saved is a survey-band midpoint capped at an hour a day, with the self-report caveat stated.27
The value rule: hours saved are capacity, not savings. A defensible case records spending avoided, capacity demonstrably redeployed and attributable margin, then subtracts build, license, model, review and operating costs.152
Primary studies and case announcements are linked. The recent reporting window is 22 September to 13 October 2026. Older foundational evidence is labeled by date.