Enterprise teams
People who research how to put AI to work inside their organisation.
what infrastructure is needed to support enterprise ai across multiple business units?
The attention gap
Who’s getting disrupted first?
Teams building AI products ask what’s next. Enterprise teams ask how to keep the lights on.
See where the questions leadmarkets or business models
Enterprise teams
6in 1,000Teams building AI products
228in 1,000Each dot = a question.
11,337 enterprise questions. 7,871 questions from teams building AI products.
A record of questions; a company’s strategy is more than its searches.
Read the method and limits ↗
Who asks
Teams building AI products
People who build AI products.
which class of business model is most resilient to build in the ai era?
Large companies
What 240 of the world’s largest companies say in public about AI.
“the first AI-enabled CGM system that can predict glucose levels up to two hours ahead”
Summary
Key findings
We placed six months of enterprise questions about AI, and the questions teams building AI products ask each other, on one strategy map. Enterprise teams aim AI at their own operations, data and control. Teams building AI products aim it at new products and customers. A company that asks only about its own operations has a blind spot: it may not see a new product coming for its customers.
- Teams building AI products ask about new products, new markets, new revenue streams. Enterprise teams almost never do.
- Enterprise teams ask 64 questions about running the business for each one about something new (1.5%). Teams building AI products split the two about evenly.
- Enterprise teams name the workflow and rarely the money.
- Risk is read from the inside. One enterprise question in five (~20%) is about governing their own AI; almost none ask about a rival or a new entrant.
- The customer journey already has new stops. Agents priced per conversation, answers without a click and AI shopping assistants are on offer today.
The strategic contrast
Builders plan what to launch. Enterprises plan how to run.
Nearly 1 in 4 builder questions is about something new. Most enterprise questions are about running operations.
Enterprise teams
Teams building AI products
Enterprise teams Teams building AI productsLit = this group asks about it
Read upwards: capabilities support processes, customer outcomes and financial results.
Teams building AI products ask each other what to make with AI, and how to make money from it:
do you see proprietary datasets holding up as a moat for ai based applications?
why llm providers sell access instead of consulting services?
is ai resulting in net revenue gain for non-tech companies?
Nearly one question in four from teams building AI products asks about new products, new markets or new business models, every year from 2023 to 2026.
Enterprise teams: 64 questions about running the business for every one about something new.
Teams building AI products: about one to one.
On money, builders ask about growth nearly 7 times as often as about cost.
Customer support agents
what operating model do teams need for production ai support agents?
top ai agent to replace call centers?
Sales agents
what is the roi of using ai sales agents for mid-market and enterprise teams?
will b2b companies welcome an ai voice bot qualifying their leads?
Business software
which generative ai tool is best for automating internal business workflows?
is replacing an enterprise product with llms a realistic strategy?
The enterprise team asks how to run the agent or the tool. The team building AI products asks what it can replace, and may sell it to the enterprise team’s own customers or rivals.
A few already ask which businesses AI will replace:
will ai kill software businesses?
if ai has 'eaten software', does it eat 'business' next?
what economic reasoning is there that bigcos have an advantage post-ai?
These threat questions are rare, about one in forty (~2.5%), and almost absent among enterprise questions.
Inside the enterprise
Enterprise questions stop at the workflow
Enterprise teams ask how to run today’s business with AI. They rarely ask about financial returns.
On the strategy map, three boxes hold four in five (80%) enterprise mentions: Data and platforms, Govern AI risk and Automate operations. The Customer and Financial layers stay nearly empty.
Where the enterprise questions go
Box weight shows each box’s share of mentions.
Fewer mentions More mentions
Enterprise teams
64 : 1
Questions about running the business for every question about something new.
A question about both is left out.
Enterprise teams name the workflow and rarely the money.
A question can mention several objectives.
Only one question about return on investment in thirty (~3%) mentions the customer.
The typical enterprise question is about the workflow:
how does an enterprise decide which workflows to automate with ai?
what's the difference between automating a process and redesigning it for ai?
Enterprise questions on each layer of the map, out of 11,337. One question can sit on more than one layer.
Financial
163questions1%what is the roi of enterprise ai cost reduction?
how do i measure ai-attributed revenue?
how can enterprise ai show measurable revenue lift in my business unit within 12 months?
Customer
303questions3%what roi can i expect from deploying ai agents in my customer service operations?
how does ai agent performance directly impact csat and nps scores?
what kpis should i use to track the rate of ai misinformation about my brand over time?
Process
3,084questions27%how can we maintain automation standards while accommodating unique business unit requirements?
how to evaluate approval workflows and approval speeds in a platform?
how do agencies scale transportation ai from corridor pilots to network-wide operational programs?
People and platforms
5,972questions53%what infrastructure is needed to support enterprise ai across multiple business units?
we don't know what ai components are running in our environment, how do we inventory them?
how do you reskill workforce for generative ai?
Operations outnumbers lower cost about 15 to 1. When the money does appear, it is a saving:
what is the expected cost reduction over 5 years with agentic ai intervention?
Three in four questions about lower cost also sit in the Automate operations box, and almost none touch a customer. Return is read as a saving:
what metrics prove enterprise ai cost reduction?
what hard and soft cost savings should be included in an ai finance automation roi model?
Only one question about return on investment in thirty reaches the customer. New products, markets and business models appear in fewer than one enterprise question in a hundred, in each of the six months.
Deloitte’s State of AI in the Enterprise 2026 finds the same order: two in three leaders said AI had delivered efficiency, one in five higher revenue.
The blind spot
Your risk register faces inwards
Enterprise teams ask a great deal about the risks of their own AI. They almost never ask how a competitor could use AI to take their customers.
Governing own AI
22%Enterprise questions on the NIST GOVERN function.
External threat
8Questions about a rival, new entrant or business model under attack.
Give competitive risk its own place.
NIST AI RMF is a framework for the harm risks of AI systems. Competitive disruption sits outside that remit. A strategy discussion also needs to ask how competitors could use AI to win the company’s customers.
These counts have different bases. They are shown as separate measures, not a ratio. Our coding covered 21 governance and risk frameworks, with limits set out in the notes.
Risk faces inwards
Enterprise mentions
Govern AI risk2,661
Protect revenue at risk8
One in five enterprise questions about AI is about governing it, in the sense of the GOVERN function of the NIST AI Risk Management Framework. Questions about a rival, a new entrant or a business model under attack number eight in all our enterprise data. The governance questions look inwards:
how do companies assign accountability for ai systems and agents?
why do ai governance programs fail to scale?
how to report ai risk to the board best practices
Questions about return on investment name governance far less often than chance predicts. When the two meet, enterprise teams ask what governance costs, and rarely what it protects.
We read 21 governance and risk frameworks, from the NIST AI Risk Management Framework to the OWASP Top 10 for large language models. They are built to manage the harm an AI system can do. None has a clause for a rival or a new entrant that replaces your business.
On the map, the governance box draws a quarter of enterprise mentions. The box for revenue at risk draws almost none.
The NIST framework has four functions. GOVERN sets the rules and names the owners. MAP describes the system, its use and its risks. MEASURE tests and tracks those risks. MANAGE acts on them. Each function has numbered clauses, 72 in all.
Four functions, 72 clauses
The NIST AI Risk Management Framework has four functions. This section is about GOVERN.
Select a function to see what it covers.
GOVERN
Sets the rules and names the owners.
19 clauses
MAP
Describes the system, its use and its risks.
18 clauses
MEASURE
Tests and tracks those risks.
22 clauses
MANAGE
Acts on them.
13 clauses
What GOVERN covers
- Policy
- Roles
- Training
- Culture
- Suppliers
What MAP covers
- Context of use
- What kind of system it is
- Benefits and costs
- Third-party parts
- Impact on people
What MEASURE covers
- Methods and metrics
- Trust tests
- Tracking over time
- Feedback on the measures
What MANAGE covers
- Rank and respond
- Benefits against harms
- Third-party risks
- Response and recovery
Figure 4b shows where enterprise questions land on the 72 clauses: most on context, benefits and the system’s own task. About one enterprise question in two touches MAP (56%), and about one in five touches GOVERN (22%).
Figure 4b
Where enterprise questions land
Where enterprise questions land on the NIST AI Risk Management Framework, clause by clause. Each square is one clause, with a short name. The shade shows the share of enterprise questions, in four steps from under 1% to 5% or more. Hover or tap a square to read the clause in plain words.
GOVERN 22%Sets the rules and names the owners: policy, roles, training, culture and suppliers.
MAP 56%Describes the system: what it is for, where it is used, who it affects and what could go wrong.
MEASURE 29%Tests and tracks the risks: accuracy, safety, security, privacy, fairness and more.
MANAGE 23%Acts on the risks: decide, fix, monitor, switch off and report incidents.
Signals from the market
The customer’s path gets new stops
Three AI services now stand between a company and its customer. Each one takes a step that used to belong to the company.
Figure 5 · The customer’s path gets new stops
Conversation
The agent prices the conversation.
$2
Salesforce sells its Agentforce agents at $2 per conversation, as read on its pricing page on 11 October 2026. The conversation is the moment a company learns what its customer wants, and the vendor now prices it.
Source: Salesforce · pricing page · read 11 October 2026
Answer
The answer replaces the visit.
68.01%
From January to April 2026, 68.01% of US Google searches ended without a click. The company’s own page gets fewer visits.
Source: SparkToro with Similarweb data, 9 June 2026.
Purchase
The assistant chooses; the merchant still takes the payment.
Inside ChatGPT
In September 2025 OpenAI and Stripe let US buyers pay for Etsy items inside ChatGPT. By March 2026 OpenAI had moved checkout back to the merchant's own site; Walmart said purchases inside the chat converted at a third of its own site's rate. The assistant still does the research and the comparison.
Sources: Stripe newsroom, 29 September 2025 · OpenAI merchant FAQ · Search Engine Land, March 2026
The shift is early.
We found no study yet that shows a company losing customers to these services.
In 2001 Michael Porter wrote: "Many have argued that the Internet renders strategy obsolete. In reality, the opposite is true." Two of his five forces are in play on the customer’s path: a substitute sells the conversation itself, and a new intermediary stands between the company and its customer.
Two forces in play
Two of Porter’s five forces are in play on the customer’s path.
New entrants
Suppliers
Rivals
New intermediary
An AI assistant stands between you and your customer.
Buyers
Substitutes
A vendor sells the conversation itself.
$2 per conversation
The shift is early. In the Reuters Institute Digital News Report 2026, "only 1% say AI is their main source of news". For retailers, Bain finds AI brings "still less than 1% of total traffic". In Bain’s view, retailers that lose the customer connection "will become little more than a fulfillment pipe".
One enterprise question already looks at the new stop:
what metrics help explain why one brand consistently wins ai recommendations over another?
Our forecast · A hypothesis
A view forward: are you automating instead of innovating?
For every question about something new, enterprise teams ask dozens about operations. This is our forecast of what comes next.
In 1994 N. Venkatraman described five levels of IT-enabled change, from local automation to redefining a company’s business scope, and advised each company to "first determine the level at which benefits are in line with the costs", then to move up as competition demands.
In 2018 Erik Brynjolfsson, Daniel Rock and Chad Syverson argued that AI and other general purpose technologies "enable and require significant complementary investments", "including co-invention of new processes, products, business models and human capital". Paul David (1990), Carlota Perez (2002) and Clayton Christensen (1995, 1997) tell the same story.
Nicholas Carr took the other side in 2003: "IT is best viewed (and managed) as a commodity". If he is right about AI, today’s enterprise teams ask the right questions.
Where the weight moves
Then and Next are hypotheses. Now is drawn from the questions.
The trigger is the market signal in section 5. In our reading of Christensen’s cases, the threat from new entrants arrived before incumbents asked about it. Teams building AI products already ask what to build in nearly one question in four.
Take it to the team
What to do
on Monday
Two playbooks. Defend the core business that AI is changing. Grow beyond it with what AI makes possible.
Enterprise questions are about running the business, so the first playbook defends it. That is where the blind spot is. But large companies also grow with AI. Their public statements show four kinds of move, and the second playbook follows them.
Method and limits
How we read the questions
A record of questions, with a contrast set. Descriptive shares and ratios; no test verdict for the contrast.
How we read the questions
Enterprise teams
11,337
Questions typed into our site, which publishes research on enterprise AI, March to August 2026.
Teams building AI products
7,871
Questions about AI asked in a public forum, 2023 to 2026.
Large companies
240
Public statements about AI by companies in the Forbes Global 2000, 2026 edition. Technology and internet companies are left out. Section 7 draws on them.
Three sources
Enterprise teams type questions into our site. Teams building AI products ask each other on Hacker News. Large companies speak about AI in public statements.
Placed on the map
Each question lands on the strategy map boxes it is about. One question can light several.
Rules set first
We wrote each rule before we saw any result.
Result
Each rule ran once. Where we revised a rule, we show both results.
The four layers of the balanced scorecard strategy map, with 28 objectives. A question can sit on several objectives. 7,471 enterprise questions and 5,183 questions from teams building AI products sit on at least one. Kaplan and Norton call the people and platforms perspective learning and growth.
A record of questions; a company’s strategy is more than its searches. Section 8 sets out the limits.
Limits
Who we see.
Questions on our own site, March to August 2026, shaped by the pages we publish. A question is one person’s search.
The contrast set.
The questions from teams building AI products are Ask HN posts on Hacker News that mention AI: a contrast set, never a sample of companies.
The coder.
One coder, a language model, placed every question from both sources on the map. A question counts on an objective only when the match is clear. We checked the coder on too few questions about new products and threats to rate it there.
No significance test for the contrast.
The tests we set in advance were stopped by their own checks. We report shares and ratios, and claim no test result.
One run.
Each rule ran once. We revised the rule for operations and new products after its first result. We show both results and read the revised one as an early signal.
The forecast.
Section 6 and figure 6 are our hypothesis; only the middle map comes from the data. Six months cannot show a strategy changing.
The pairs.
Both sides of each pair in section 3 are quoted as asked; the pairing by subject is ours. That the team building AI products may sell to the enterprise team’s customers or rivals is our reading of three examples.
The frameworks.
We read 21 governance and risk frameworks, the paywalled ISO texts from public summaries. Our library omits enterprise risk frameworks such as COSO ERM (2017) and ISO 31000, which give business context a place.
Outside numbers.
Each figure in section 5 was read on the publisher’s own page on the date given.
The company examples.
The four moves in section 7 come from an early reading of public company statements. We show the kinds of move, not how common they are. The company report will test that.
09 · Sources
Sources
Our data
- Enterprise teams: 11,337 questions typed into the search on our site, March to August 2026.
- Teams building AI products: 7,871 Ask HN posts on Hacker News that mention AI, 2023 to 2026.
- Large companies: public statements about AI by 240 companies in the Forbes Global 2000, 2026 edition. Technology and internet companies are left out.
- Shares in section 2 count mentions on the map. One question can make several mentions.
Frameworks
- Robert Kaplan and David Norton, Strategy Maps, 2004.
- NIST, AI Risk Management Framework 1.0, 2023.
- OWASP, Top 10 for Large Language Model Applications.
- ISO/IEC 42001, ISO/IEC 23894 and ISO/IEC 27001.
- European Union, AI Act, 2024.
Market signals
- Deloitte, State of AI in the Enterprise 2026. 3,235 leaders, August to September 2025.
- Salesforce, Agentforce pricing page, read 11 October 2026.
- Chartbeat, The 2026 Publisher Playbook, read 14 October 2026.
- SparkToro with Similarweb data, 9 June 2026.
- Stripe newsroom, 29 September 2025.
- OpenAI, merchant FAQ for shopping in ChatGPT.
- Search Engine Land, March 2026.
- Reuters Institute, Digital News Report 2026, 16 June 2026.
- Bain, Agentic AI in Retail, 14 May 2026. Bain, Agentic AI Commerce, 2 March 2026.
Company statements
Strategy and history
- Michael Porter, Strategy and the Internet, Harvard Business Review, 2001.
- N. Venkatraman, IT-Enabled Business Transformation, MIT Sloan Management Review, 1994.
- Erik Brynjolfsson, Daniel Rock and Chad Syverson, The Productivity J-Curve, NBER Working Paper 25148, 2018.
- Nicholas Carr, IT Doesn't Matter, Harvard Business Review, 2003.
- Paul David, The Dynamo and the Computer, 1990.
- Carlota Perez, Technological Revolutions and Financial Capital, 2002.
- Joseph Bower and Clayton Christensen, Disruptive Technologies, 1995. Clayton Christensen, The Innovator's Dilemma, 1997.