AI can now write code, read contracts and answer specialist questions, yet companies still struggle to turn those abilities into operating results. A new survey puts that gap in plain numbers.
01What happened
On October 6, 2026, the enterprise software company Infor published the second edition of its Enterprise AI Adoption Impact Index. YouGov ran the survey in August 2026 across seven markets and 2,111 business decision-makers. 68% said off-the-shelf AI does not adequately address their industry’s needs, a majority in six of the seven markets. Among manufacturing respondents the figure was 73%; in distribution, 76%.1
The same day, Infor launched an expanded industry AI offering that bundles industry-specific AI agents, process mining, automation and implementation services.1
These are Infor’s own survey and products. How well the products work still has to be tested against customer results.
02The real question
That 68% of businesses are unhappy with general-purpose AI is not the point. The real question is this: if AI is this capable, why do companies still find it so hard to turn it into operating results?
We think the answer lies less in the model than in the space between the model and how a company actually runs.
03What we think
Faster people do not add up to a better company
Individual work does get faster. A study in The Quarterly Journal of Economics followed 5,172 customer-support agents: with a generative AI assistant, the issues they resolved per hour rose 15% on average, with the largest gains among the least experienced.2
Yet in McKinsey’s August 2026 survey, 80% of respondents said AI had improved their own productivity, while only 37% said AI had contributed positively to their organization’s EBIT, even in part.3
The two numbers cannot simply be subtracted, but they point to a gap: we have started to solve how to make one person faster, not how to make a whole company run better. Another McKinsey article in September calls the gap a “coordination tax”: the total cost a company pays for handing work from one step to the next through people.4 If AI only speeds up each step and leaves the handoffs between them alone, the time it saves is easily eaten by waiting, checking and rework.
What general AI lacks is not industry knowledge but operating context
“AI doesn’t understand our industry” is only half right. Today’s models know a great deal about industries: accounting principles, regulations, processes. But knowing how an industry usually works is not the same as knowing what is happening inside one company right now.
Take an accounting firm chasing a client for documents. An AI can write a polite reminder in seconds. To own the chasing, though, it would need to know whether the client has sent anything, whether by email or by LINE, whether the accountant has confirmed it, whether a manager approved an exception to the deadline, and when the last reminder went out. That information is spread across systems, files, chats and people’s memory, and some of it was never written down at all.
What a general AI already knows
- The regulations and filing deadlines
- How to write a polite reminder
- How firms usually divide the work
What it would need to know to own the chasing
- Where things stand
- Has the client sent the documents? By email or LINE?
- Has the accountant confirmed them? When did the last reminder go out?
- The rules that apply
- Has a manager approved an exception to the deadline?
- What has to be true before the case moves on?
- Who may do what
- Who can confirm the documents are complete?
- Who can approve an exception?
- What happens after
- Which state gets updated?
- What record is left behind, to look up later?
What a company needs is not just knowledge but operating context: what the AI must know about this company, right now, before it can act. Where things actually stand, which rules apply, who may do what, and which state to update and what record to leave once the work is done. Without that, AI remains an adviser standing outside the company.
Large enterprise software is moving the same way
Infor is not alone. SAP grounds its Joule Agents in a knowledge graph of the company’s processes, data and how they relate.5 Palantir’s Ontology represents an operation as data, logic and actions.6 In May 2026 ServiceNow expanded AI Control Tower to monitor and govern the AI running across a company’s systems.7
These products differ in architecture, customers and maturity, and they are not the same thing. But they are all filling the same gap: letting AI know where the company stands, act within its rules and permissions, and leave a record of every step that someone can oversee. In other words, the large vendors are building operating context too. The part outside the model is becoming where enterprise software’s value lies.
The hardest step: much of the work was never recorded
Infor’s offering includes process mining. Process mining rebuilds the order in which work actually happened from the event records systems leave behind, and an event log needs at least three fields: a case ID, an activity name and a timestamp.8
If most of the work happens in LINE chats, spreadsheets, phone calls and hallway handoffs, process mining has nothing to rebuild those events from, and neither does AI. The problem is not only that the data is messy; a company’s real work may never have been expressed as information a system can read. A document that exists has not necessarily been checked. A payment on record has not necessarily been matched. An email sent has not necessarily moved the case forward.
04Why it matters
Finding where AI belongs is itself a capability. In a 2026 field experiment by INSEAD and Harvard Business School, 515 high-growth startups were shown how other companies had used AI to reorganize their work. Compared with the control group, they found 44% more uses for AI, completed 12% more tasks and earned 1.9 times the revenue, with the gains concentrated in the strongest outcomes.9 In McKinsey’s survey, nearly three quarters of AI high performers said they had fundamentally redesigned workflows because of AI, against about a quarter of everyone else.3
Both studies have limits: one samples high-growth startups, the other shows correlation, not cause. But they point the same way. The question for a company bringing in AI should move from “which tasks can AI automate?” to “which work should exist at all, which handoffs can go, which decisions belong to the system and which stay with people?”
Industry-specific does not have to mean building everything from scratch for every company, either. Companies in one industry share many rules, forms and data formats, which can be built into a product; what differs from one company to the next, such as permissions, who answers for what and the exceptions, should be configurable rather than redeveloped on every rollout.
How AI is measured has to change too. Logins, words generated and the share of staff using agents are intermediate signs. What matters is the end-to-end result: how much shorter the path from a client’s request to delivery has become, how many needless handoffs are gone, whether errors and rework have fallen, and whether, when something goes wrong, you can trace who did what. The US NIST AI Risk Management Framework likewise asks organizations to document roles and responsibilities and to define how people oversee AI.10
05What we do about it
Models will keep improving and intelligence will keep getting cheaper. A company’s real processes, data, permissions and responsibilities will not become clear on their own because a model was upgraded. Whoever sees them first, writes them down and builds them into a system gives AI somewhere to work, and having done it once makes the next company faster.
Augvant builds operating software for process-heavy small businesses, starting with accounting firms in Taiwan. This is how we work.
First, make the work leave a record. Inside a firm, our first step is to make the operation visible: how data flows, where money moves, which deadlines are coming due, where work is stuck. Without that, any automation or AI that follows is guessing.
Keep “sent” and “done” apart. An email sent, a payment received and a document arrived each have to map to a clear state and a person who answers for it, so the system knows where a case really stands.
Find what repeats, in one real firm. We find the operating components that recur in one real business and build them into the product; each firm’s own permissions and exceptions are left to configuration, not a rewrite.
AI is a means, not the goal. Every agent has a role, a scope and someone who answers for it, and what falls outside the scope goes to a person. What cannot hold up on a real firm’s floor, we do not ship.
Rather than asking “what other AI tools could we bring in?”, the better first question may be this: if we redesigned how this company runs today, what would we stop doing the old way altogether?
Based on public research and product information as of October 9, 2026. Passages marked “Inference” and “Our view” are Augvant’s analysis; product results announced by vendors are not independently verified.
Sources
-
Infor, “Infor Unveils Its Expanded Infor Industry AI™ as New Global Research Finds Off-the-Shelf AI Falls Short for Two Out of Three Businesses,” PR Newswire, October 6, 2026. https://www.prnewswire.com/news-releases/infor-unveils-its-expanded-infor-industry-ai-as-new-global-research-finds-off-the-shelf-ai-falls-short-for-two-out-of-three-businesses-302899181.html ↩ ↩2
-
Erik Brynjolfsson, Danielle Li and Lindsey Raymond, “Generative AI at Work,” The Quarterly Journal of Economics 140(2): 889–942, 2025. https://academic.oup.com/qje/article/140/2/889/7990658 ↩
-
McKinsey & Company, “The state of AI in 2026: On the road to ROI,” August 25, 2026. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai ↩ ↩2
-
McKinsey & Company, “Cutting the ‘coordination tax’: How agentic AI can reshape workflows,” September 18, 2026. https://www.mckinsey.com/industries/industrials/our-insights/cutting-the-coordination-tax-how-agentic-ai-can-reshape-workflows ↩
-
SAP, “Joule Agents: How SAP Uniquely Delivers AI Agents That Truly Mean Business,” SAP News Center, February 2025. https://news.sap.com/2025/02/joule-sap-uniquely-delivers-ai-agents/ ↩
-
Palantir, “Platform overview,” Foundry documentation. https://www.palantir.com/docs/foundry/platform-overview/overview/ ↩
-
ServiceNow, “ServiceNow expands AI Control Tower to discover, observe, govern, secure, and measure AI deployed across any system in the enterprise,” May 5, 2026. https://newsroom.servicenow.com/press-releases/details/2026/ServiceNow-expands-AI-Control-Tower-to-discover-observe-govern-secure-and-measure-AI-deployed-across-any-system-in-the-enterprise/default.aspx ↩
-
Wil van der Aalst, “Event Logs: What kind of data does process mining require?,” Eindhoven University of Technology, 2016. https://www.processmining.org/old-version/files/event_logs_the_input_for_process_mining.pdf ↩
-
Hyunjin Kim, Rembrand Koning and Dahyeon Kim, “Mapping AI into Production: A Field Experiment on Firm Performance,” INSEAD Working Paper 2026/20/STR, 2026. https://www.insead.edu/faculty-research/publications/working-papers/mapping-ai-production-a-field-experiment-firm ↩
-
National Institute of Standards and Technology, “Artificial Intelligence Risk Management Framework (AI RMF 1.0),” NIST AI 100-1, January 2023. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf ↩