JournalEssay
The future
of enterprise
When everyone can buy intelligence, what makes a company worth more?
As models improve and agents take on more of the work, companies will have to be more deliberate about what makes them different: the customers they reach, the judgement they exercise and what they learn by doing the work.
- People answer for the result.
- Agents and models do the work, and keep changing.
- What the company keeps: its customers, its responsibilities and what it has learned.
Speculative scenario · an invented company, not a customer result
One night at a pump maker, a few years from now.
22:00Work proceeds
The night shift runs the lines. In the planning office, agent teams work through tomorrow’s schedule on three different models, and the three people who answer for it have gone home.
03:40A condition changes
A supplier reports that a seal compound will be six weeks late. Of the five revised plans the planning team drafts, the cheapest uses a substitute the plant has never qualified, and no agent is allowed to approve that.
07:30A person decides
The quality engineer rules out the substitute for this order and starts a qualification trial. The plant manager approves the next-cheapest plan and gives the customer a new date.
Weeks laterThe next job starts ahead
A different part runs short, and the planning team starts from what that night taught: check that a substitute is qualified for the customer’s use before costing it. It brings two workable plans and nothing to escalate.
Five things were at work that night. The essay takes them in turn, then tries them on eight industries.
01 · Intelligence: General intelligence becomes easier to buy.
Knowledge is becoming easier to access. Advantage is harder to keep.
That night. The models that drafted those plans are available to every competitor.
A competitor can ask the same model to analyse a contract, write a proposal or build a feature. That general capability keeps improving, and the cost of reaching any given level of performance can fall sharply.
Our bet is that this continues. Frontier models will still matter, particularly where mistakes are costly or the work is unfamiliar. But access to a capable model explains less and less about why one company outperforms another.
A company’s own experience is not available to its competitors in the same way. A model does not arrive knowing why your last implementation failed, which promise matters to a particular customer, or when the standard process stops working. The company has to supply that knowledge itself.
02 · Model choice: The enterprise will have a model portfolio.
The best model for the job will keep changing.
That night. The work ran on three different models, and the mix will change.
Some work will justify the strongest available model. Repeated, well-understood tasks may run on smaller or open-weight models. Sensitive work may need a particular deployment boundary. A company’s requirements will differ by task, and those choices will change as models improve.
Open weights widen the options for hosting, adaptation and supplier choice, subject to their licences and the practical cost of operating them. They do not make inference free, and they do not make every model equally capable.
That gives an enterprise a reason to carry its working knowledge from one model to the next. It still has to test how each new model uses that knowledge.
A majority of tokens. A minority of spending.
Vercel AI Gateway, August 2026. One platform, not the whole AI market. E2
In the same month, Anthropic accounted for 64% of the gateway’s estimated spending. Open-weight token volume coexists with substantial spending on proprietary models.
Method note
Token volume includes input, output, reasoning and cache-related tokens. Spending is valued at published list prices, not actual invoices. This report broadened its open-weight definition, so earlier reports’ historical values are not comparable and have not been combined with it. E2
“Open-weight” is not automatically “open-source”, and open weights hosted by a third party are not self-hosted inference. E3
Source detail · architecture guidance
BCG recommends a common enterprise architecture and control plane while preserving flexibility across models and providers. That is architectural guidance, not evidence that every respondent already operates a multi-model stack. E7
BCG, The Formula for Agentic AI Value, September 2026, pp.17–18. Authors’ recommendation, distinct from the survey’s measurements. No open-weight adoption percentage is inferred.
03 · Agent work: Work scales beyond the people doing it.
Imagine a company running 99 agents for every person.
That night. Agents did the work, and three people answered for it.
Illustration · drawn to scale
99% agents. 1% humans.
A thought experiment about the mix of agents and people. Not a staffing forecast, a share of working hours, or a target for every industry.
Far more work could run in parallel: researching a market, reconciling an account, preparing a release, checking a supplier or assembling a case. Small teams could attempt work that previously required a much larger organisation.
The difficult part would move. Someone still has to choose the objective, decide what may be delegated and handle situations the standard process cannot resolve. Customers still need a person or institution that can answer for a consequential decision.
The operating unit could become a team accountable for an outcome, with agents working inside clear limits. That changes management, training and software purchasing. Physical work and professional duties remain, and so does the need to develop the next generation of experts.
The activity can multiply, but someone still has to be responsible for it.
Illustration · the same night, as one workflow
| Stage | Execution | Responsibility |
|---|---|---|
| Supplier notice | Set the objective: plant manager | |
| Investigate | ||
| Propose plans | Authorise the scope: operations lead | |
| Verify | Flagged exception · handle the exception: quality engineer | |
| Release | Answer for the result: plant manager |
Routine work continues within authorised limits.
Read this diagram
In the Execution view, small strokes beneath the five stages (supplier notice, investigate, propose plans, verify, release) stand for units of illustrative activity. In the Responsibility view, four annotations appear: the plant manager set the objective before the notice arrived; the operations lead authorised the scope around the execution stages in advance; the quality engineer handles exceptions flagged at verification; and the plant manager answers for the result after release. Routine work continues within authorised limits.
Count accepted outcomes, not agents.
A cheap model call can still lead to expensive work. Include failed attempts, tools, integration, expert review and remediation when judging the result. Released capacity is useful, but it is not automatically cash saved. Competition may pass the benefit to customers rather than leave it in the company’s margin.
Cost per accepted outcome equals total workflow cost divided by accepted outcomes
A measurement principle, not a savings calculator. Define quality, risk and acceptance before comparing costs.
Source detail · job redesign
Deloitte reports that 84% of surveyed organisations had not redesigned jobs around AI capabilities. Changes to jobs are a different measure from changes to particular workflows. E9
Deloitte, State of AI in the Enterprise, January 2026, p.13 · fieldwork August–September 2025 · overall N 3,235 leaders at AI-active organisations. Exact questionnaire wording and question-specific N not supplied beside this finding. The apparent contrast with McKinsey’s workflow figure is not a contradiction or a trend; the measures differ.
04 · Advantage: A company has to earn what makes it different.
What does the next piece of work inherit from the last?
That night. The plant learned something, and the next shortage began with it.
A company can generate more output while repeating the same mistakes. A project finishes, an experienced person repairs the difficult parts, and the next team starts with the final document but none of the judgement behind it.
Useful learning comes from the whole attempt: successful outcomes, failed approaches, expert corrections and changes in the environment. Taken together, several pieces of evidence can reveal a pattern that nobody saw in any one of them. That pattern is a candidate explanation until later work supports it.
This becomes an advantage when the experience is difficult to recreate, the resulting knowledge improves important work, and the organisation can keep it current. A generic prompt or an obsolete workaround may be useful without being defensible.
What remains worth building?
- Access
- Customer relationships, distribution and the right to participate in valuable work.
- Scarce assets
- Proprietary data rights, physical capacity, capital, licences and specialist expertise.
- Working knowledge
- An evidence-backed understanding of what works in this environment.
- Reliable execution
- The ability to deliver an acceptable result and recover when something fails.
- Institutional trust
- People who will rely on the organisation and a credible way to challenge it.
Candidate sources of advantage, not guaranteed moats. Organisational learning can strengthen the others; it does not replace them.
The difference between efficiency and advantage
Producing work more efficiently can be worth having on its own. It does not tell us who keeps the benefit, whether customers will pay more, or whether a competitor can reproduce it.
Illustration · what the plant learned that night
The old guidance says to take the cheapest substitute that meets the drawing.
After that night, a look across past shortages shows something else. Some cheap substitutes sat for weeks waiting for a customer’s approval; some dearer ones shipped at once because the plant had already qualified them.
The team tests a more specific lesson: check that a substitute is qualified for the customer’s use before costing it.
Illustration · how a lesson is formed and kept current
Work evidence
Shortage records, quality holds and engineers’ decisions from earlier disruptions.
Candidate explanation
Unit price may be a poor guide to cost; whether a material is already qualified for the customer’s use explains more of it.
A rejected candidate
Never substitute materials on a drinking-water order.Rejected: qualified substitutes had shipped on such orders without a problem, so the evidence did not support a blanket rule.Tested within scope
Applied by the planning team to new shortages, with the outcomes recorded against the lesson.
Revised when conditions change
A qualification trial changes which substitutes are acceptable. Any guidance that rested on the earlier ruling is reassessed, not assumed.
Illustration · the same plant, a year on
The replacement test
Replace the models doing the work, then change the conditions, and see what the company still has.
Execution
| Task | Now | A year on |
|---|---|---|
| Simulate revised schedules | Model AFrontier model, bought as a service | Model DNewer frontier model, bought as a service |
| Reconcile supplier records | Model BRetired | Model ENewer open-weight model, run on siteModel B · Retired |
| Read drawings and qualification records | Model COpen-weight model, run on site | Model ENewer open-weight model, run on site |
What the company keeps
Customers
A water utility’s standing order, and the delivery date promised to it.
Responsibilities
The plant manager answers for the schedule.
The quality engineer qualifies any change of material. Now reviewing lesson 2.
Operating knowledge
Check that a substitute is qualified for the customer’s use before costing it.
Current guidance, subject to new evidence.
The substitute compound is not qualified. Leave it out of revised schedules.
- Basis
- The quality engineer’s decision that night.
- Applies within
- The planning team’s work at this plant.
- Accountable role
- Quality engineer.
Current guidance, subject to new evidence.
Review required. The qualification trial has reported. This lesson is not presented as current until the quality engineer has reassessed it.
The substitute compound is qualified for industrial pumps. Leave it out of schedules for drinking-water pumps.
- Basis
- The qualification trial and the quality engineer’s review.
Revised after review. Current guidance, subject to new evidence. Earlier wording kept:
The substitute compound is not qualified. Leave it out of revised schedules.
Read this diagram
Three tasks run on three models. Changing the models replaces Model A with Model D, retires Model B and moves its task to Model E, which also replaces Model C. The new models still need evaluation. Beneath them, the customers, the responsibilities and two lessons stay where they are.
Changing the conditions brings the result of the qualification trial. The second lesson becomes “Review required”, goes to the quality engineer and is not presented as current until it has been reassessed. Reviewing it narrows the lesson and keeps the earlier wording. Retaining a lesson is not the same as keeping it true. Reset restores the first illustration; it is not a reassessment.
Existing systems record customers, money, assets and transactions. Agent-operated companies also need an account, kept current, of how work should be done, why the guidance exists and where it applies. This could become a new system of record for operating knowledge, alongside the records that still have to remain authoritative.
Source detail · memory as a control
BCG includes memory and context handling among six controls for agentic AI. The report also warns that memory can carry errors or sensitive information across workflows. Our argument is that retained experience needs evidence, permission and revision. E7
BCG, The Formula for Agentic AI Value, September 2026, pp.14–15. BCG’s control framework and author analysis, followed by our interpretation. Not an empirical test of organisational memory or an endorsement of memco.
05 · Control: Delegation needs an owner.
The more you delegate, the clearer the boundaries need to be.
That night. No agent was allowed to approve the substitute.
Knowing how to do something does not grant an agent permission to do it. A lesson can be well supported and still belong to another customer, another jurisdiction or an expired policy.
A useful agent needs an identity, a mandate and limits. The organisation needs evidence of what happened, a way to stop or reverse an action where possible, and someone with the authority to resolve a dispute. Learning should improve the process without silently rewriting company policy.
Control also means being able to change suppliers. Data location matters, but so do access, retention, encryption keys, tool connections and the ability to keep operating or leave. Downloading model weights settles only part of that question.
A smaller supervisory team still needs a path for people to learn the work. Apprenticeships, supervised cases and practical experience matter when tomorrow’s reviewers will have done less of the routine work themselves.
Two findings from BCG’s 2026 enterprise survey. E7
- 42%expected agents to act autonomouslyExpected by 2030
- 5%reported having all six agent-governance controls deployed enterprise-wideReported at the time of the survey
Complete implementation of six controls is a stricter test than having some governance. These are different measures, not a measured shortfall within the same subgroup.
BCG Applied AI Index, 2026 · overall sample of 1,330 senior executives.
Illustration · the same plant in cross-section, not an architecture diagram
policy and permission constrain ↓
03:40 · the unqualified substitute stopped here and went to a person
reads and writes within permission ↕︎knowledge informs execution ↑
Read this diagram
Purpose and authority sit above everything and constrain what agents, models and tools may do. Those run inside a permission boundary drawn around one programme at the plant; a separate customer boundary is never crossed by a connection. On the night of the opening scene, the unqualified substitute stopped at the permission boundary and went to a person. Beneath the two boundaries sit authoritative business records and maintained operating knowledge: records are read and written within permission, and knowledge informs execution without granting it. An assurance line alongside checks outcomes and can stop an action.
- What is the agent trying to achieve?
- What is it allowed to access and change?
- What evidence supports the guidance it uses?
- Who can intervene and answer for the outcome?
Reversible routine actions can run inside an authorised envelope; high-consequence actions need proportionate controls and real escalation capacity. Sovereignty is assessed across several dimensions in the European Commission’s 2026 sovereign cloud framework, which is a procurement framework, not a local-model test or a universal legal certificate. E4
Source detail · governance maturity
In Deloitte’s survey, 21% of respondents said their organisations had a mature model for governing autonomous agents. E9
Deloitte, State of AI in the Enterprise, January 2026, p.20 · fieldwork August–September 2025 · overall N 3,235. Question-specific N and the operational definition of maturity are not disclosed beside the finding; this is a self-assessment. It is not comparable with BCG’s 5%: “mature model” and “all six controls enterprise-wide” are different tests in different samples.
Industry atlas
The same forces create different companies.
A bank, a hospital and a manufacturer may use similar models. They will not have the same assets, responsibilities or limits. These scenes explore what could change in each.
Speculative scenarios · not customer results. Each ends with one public signal, which shows a direction being tried and does not validate the scene.
Software and digital products
A release is ready before the product team arrives. The tests pass. The unresolved decision is whether the change serves the customer.
- Cheaper to produce
- A tested, staged change, with its investigation, specification and rollback plan.
- Still scarce
- Distribution, customer relationships and engineers who have learned through real design, debugging and incident work. Quality and demand decide whether faster delivery means better products or only more output.
- What the next attempt knows
- How this system fails in practice: which changes caused incidents and which fixes held.
- A person still answers for
- Whether this release serves the customer. Security and reliability teams can still stop it on their own authority.
- Public signal
- GitHub’s May 2025 coding-agent launch put issue-to-PR work inside a human-controlled review workflow. I1 · Human-reviewed coding workflow.
Source detail · building instead of buying
Among respondents whose organisations regularly used AI in at least one business function, 32% said they had decided against buying one or more additional software products or features in favour of building the functionality in-house with AI coding tools. E8
McKinsey, The state of AI in 2026, Exhibit 4, p.9 · fieldwork May–June 2026 · overall survey N 1,719; question-specific N not disclosed. A forgone addition is not the same as replaced software, displaced spending or a successfully maintained deployment.
Professional services, legal and accounting
The transaction changes overnight. By morning, the evidence room contains revised schedules and draft advice. The partner sees the unresolved conflicts.
- Cheaper to produce
- Research, reconciliation, obligation checks and tracked drafts on authorised evidence. When revised advice is ready by morning, time spent is a weaker basis for the fee.
- Still scarce
- Trust, a professional’s signature and independent assurance. The offer can shift to fixed-scope or recurring services; that does not guarantee outcome pricing, and the client may keep part of the saving.
- What the next attempt knows
- Which methods and precedents worked on earlier matters. Anything specific to one client stays inside that client’s boundary.
- A person still answers for
- Whether the advice can be signed. Assurance stays independent of the work it assesses, and juniors still learn on supervised client work.
- Public signal
- Thomson Reuters announced CoCounsel Legal and CoCounsel Tax in August 2025, combining professional content with assisted workflows. I2 · Product announcement.
Banking and insurance
Routine cases continue through the night. The morning queue contains disputed coverage, unusual exposures and customers who need a different decision.
- Cheaper to produce
- Handling a routine case, from gathering evidence to carrying out the authorised steps. The measure is accepted resolutions and customer outcomes, not cleared queues alone.
- Still scarce
- Capital, distribution and trusted, regulated relationships. Lower handling cost leaves credit risk, catastrophe risk and the cost of capital where they were.
- What the next attempt knows
- Which exceptions the institution has made before, from its own permissioned transaction and loss history.
- A person still answers for
- The morning queue: lending and coverage exceptions and contested outcomes, within a risk appetite people have set. Customers have a meaningful route of appeal, and physical inspection stays where it is needed.
- Public signal
- BNY reported deployed digital employees with credentials and supervisors in October 2025. I3 · Operator-reported deployment.
Healthcare and life sciences
The discharge plan is assembled. At the bedside, a clinician hears something that changes it.
- Cheaper to produce
- The administration around care: records assembled, documentation prepared, follow-up coordinated and evidence organised for review.
- Still scarce
- Trusted care relationships and the capacity to deliver care. Funding rules and clinical safety decide whether released capacity reaches patients.
- What the next attempt knows
- Validated protocols and consented outcome evidence. Patient records are not unrestricted training material.
- A person still answers for
- Whether the plan changes: diagnosis, treatment, consent and safeguarding. Nurses, carers and laboratory staff remain visible, and physical care and experiments are not assumed automated.
- Public signal
- Microsoft’s October 2025 Dragon Copilot nursing announcement required nurse review and approval of generated documentation. I4 · Nursing documentation with required review and approval.
Manufacturing and industrial operations
A component shortage produces several revised production plans. The cheapest option uses a material the plant has never qualified. The night this essay opened with ↑
- Cheaper to produce
- Plans: demand and supply assessed, alternatives simulated, production scheduled and quality or maintenance signals interpreted.
- Still scarce
- Qualified processes, equipment and supplier relationships. Reliable throughput matters more than generated plans, and software cannot remove a physical bottleneck by reasoning about it.
- What the next attempt knows
- What the plant has qualified, what it refused and why.
- A person still answers for
- Whether to change the operating envelope. Engineers and technicians commission, inspect and repair, and independent safety interlocks stay outside discretionary model control.
- Public signal
- Siemens announced an industrial agent architecture in May 2025, alongside products at different stages of availability. I5 · Architecture announcement; availability varies by component.
Retail and logistics
A delayed shipment changes the promise shown to shoppers before they need to complain. The network prepares alternatives.
- Cheaper to produce
- Forecasts, inventory allocation, delivery routes, customer updates and permitted returns and settlements. Price competition may pass the saving to buyers.
- Still scarce
- Fulfilment density, access to inventory, supplier terms and accurate stock data. If customers’ agents become an important route to demand, machine-readable offers and reliable fulfilment are worth more.
- What the next attempt knows
- How earlier disruptions were handled, and which promises held.
- A person still answers for
- Which promise to make when the disruption is unfamiliar, what a vulnerable customer is owed and anything that touches safety. Physical picking, driving and maintenance are separate automation questions.
- Public signal
- Walmart described a four-super-agent framework in July 2025, combining live capabilities with planned expansion. I6 · Partly live, partly planned.
Government and public services
A resident applies once. The evidence is assembled across services, and a caseworker can explain and change a decision when the standard rule does not fit.
- Cheaper to produce
- Accessible intake, lawful evidence retrieval, completeness checks and draft decisions. The measure is lawful service quality, access and time to resolution, not maximum throughput or denial rates.
- Still scarce
- Institutional capability, trustworthy registers and public legitimacy. This is public value, not a moat built by trapping citizens or hoarding their information.
- What the next attempt knows
- Where the standard rule did not fit and how lawful discretion was used, within the rules on residents’ data.
- A person still answers for
- The decision when the rule does not fit: policy, lawful discretion, safeguarding and appeals. Non-digital access stays open, records stay portable and the supplier can be replaced.
- Public signal
- The UK’s May 2025 Humphrey/Consult trial analysed consultation responses, with experts also reviewing every response. I7 · Trial with every response also human-reviewed.
Energy and infrastructure
The storm has not arrived. A plan for a vulnerable part of the network is ready for the controller and field crews.
- Cheaper to produce
- Preparation: forecasts and telemetry combined, contingencies modelled, maintenance and switching plans drafted and resources coordinated.
- Still scarce
- Network assets, transformers, connection rights, reliable telemetry and physical resilience. Value depends on reliability, avoided losses and regulation.
- What the next attempt knows
- What earlier restorations taught, alongside local engineering knowledge.
- A person still answers for
- Whether to act on the plan: operating limits, emergency priorities and safety-critical authorisation. Qualified crews do the site work, and engineered protection systems remain independent.
- Public signal
- Schneider Electric announced One Digital Grid Platform in March 2025 to integrate planning, operations and flexibility. I8 · Platform announcement.
Where to begin
Start with the next piece of work.
Choose a recurring workflow where experience should make a difference. Decide what counts as an acceptable outcome, then give the next person or agent the relevant lessons from earlier work, within the right permissions.
Compare the result with a credible baseline: the same tools, good documentation and competent people. Count review, rework and maintenance as well as speed. Change the guidance when the evidence changes.
- What valuable work could we delegate?
- Which decisions and responsibilities must stay explicit?
- What experience would improve the next attempt?
- Who is allowed to reuse it?
- How will we know the result is better?
Published by memco. We build shared, governed learning for agents across coding and knowledge work. This essay explores the wider enterprise changes behind that work; its scenarios are possibilities, not product claims.
memco is shared memory for AI agents.
What one agent learns doing the work, every agent on your team can use.
Sources
Evidence and assumptions.
The essay uses three kinds of statement and labels the boundaries between them.
- Observation: a dated claim tied to a public source and its actual denominator.
- Editorial bet: the essay’s interpretation of what those changes may mean.
- Scenario: an invented future scene, labelled before it is read.
The enterprise figures come from published surveys by BCG, McKinsey and Deloitte. Their samples and questions differ. Most results describe what respondents report or expect; they are not independent audits of business performance. We distinguish those findings from the publishers’ recommendations and our own argument.
Individual access to AI can save time without changing how an organisation works. The transition described here takes more than handing out tools. That reading is supported by a 66-firm field experiment E6; it is an interpretation, not proof of any particular architecture.
Sources checked 3 October 2026. Announcements describe the capabilities stated at the time; they are not assumed to be the latest feature set, independently verified results or customer outcomes.
Evidence
- E1
Epoch AI — The plunging price of thoughtObservation · study
Approximately 47% quarterly decline in the cost of reaching a fixed level of benchmark performance, across five main benchmarks since 2023.
A fitted frontier estimate, not an average customer’s bill or a promise about business productivity.
- E2
Vercel — AI Gateway production index, September 2026Observation · measured usage
56% open-weight share of token volume; 14% open-weight share of estimated spending; Anthropic 64% of estimated spending. Complements (44%, 86%) are calculated from the reported shares.
One platform, not the whole market. Token volume includes input, output, reasoning and cache-related tokens; spending is valued at published list prices, not invoices. The report broadened its open-weight definition, so earlier values are not comparable.
- E3
Open Source Initiative — Open Source AI Definition 1.0Framework
Downloadable weights alone do not establish Open Source AI status; the actual model licence has to be checked.
A definition, not a licence review of any particular model.
- E4
European Commission — Sovereign cloud framework explainedFramework
Sovereignty is assessed across several dimensions, not as a local-model boolean.
A procurement framework, not a universal legal certificate.
- E5
Brynjolfsson, Li and Raymond — Generative AI at work, Quarterly Journal of EconomicsObservation · study
Supporting note: about a 15% average increase in issues resolved per hour among human support workers given an AI assistant, with heterogeneous effects across workers.
Human workers with assistance, not autonomous agents; one deployment, not a universal gain or a memco result.
- E6
Dillon et al. — NBER working paper 33795Observation · study
Counterweight: individual AI access produced time savings without detected broader shifts in task quantity or composition.
Does not show that AI has no value; it shows that distributing tools to individuals is not, by itself, organisational change.
- E7
BCG — The formula for agentic AI value / Applied AI IndexObservation · survey
Almost half of the sample classified as scaling (41%) or future-built (7.5%), the groups BCG describes as generating value (p.8, Exhibit 2). 42% expect agents to act autonomously by 2030; 5% report all six agent-governance controls deployed enterprise-wide at the time of the survey (p.15, Exhibit 9). Architecture guidance pp.17–18; memory as a control pp.14–15.
A publisher classification and respondents’ expectations, not audited ROI. Respondents report on the business areas they know best, so not every answer is company-wide. More than half of participating companies had revenue above $5bn. Expectation and current self-report are different measures and are not subtracted. The report carries reproduction restrictions: brief attributed paraphrase and links only.
BCG’s six agent-governance controls
- Memory and context handling
- Agent scope and oversight
- Evaluation and rollback
- Tool and API integration
- Cross-functional ownership
- Security, audit and cost guardrails
The publisher’s categories, not a product-feature comparison or a certification scheme. The report’s 32% “relative importance” for memory and context is a model-based importance measure and is not shown as adoption, value share or uplift.
↩ 01 · Intelligence↩ 02 · architecture note↩ 04 · memory note↩ 05 · Control
- E8
McKinsey — The state of AI in 2026: On the road to ROIObservation · survey
Nearly three-quarters of AI high performers versus about one-quarter of other respondents report fundamentally redesigning workflows because of AI (pp.18–19, Exhibit 11; n = 92 and 1,429). Secondary: 32% of respondents at organisations regularly using AI decided against buying additional software in favour of building with AI coding tools (p.9, Exhibit 4; question N not disclosed).
Association and reported decisions, not causal ROI or displaced spend. The full questionnaire stem is not published; the measure label is used instead. Fractions are kept approximate.
- E9
Deloitte — State of AI in the Enterprise: The untapped edgeObservation · survey
66% report improved efficiency and productivity today and 20% report increased revenue (p.10, Figure 2; question N 3,235). Published question: “With regards to benefits from your AI efforts: Which benefits are you achieving today? Which benefits do you hope to achieve?” Only the “achieving today” values are charted. Secondary: 84% had not redesigned jobs around AI (p.13); 21% report a mature model for governing autonomous agents (p.20).
Self-report at AI-active organisations, not a population census or independent audit. Benefit categories overlap and are reported incidence, not gain magnitude. Question-specific N for the secondary findings is not disclosed.
↩ 04 · Advantage↩ 03 · job redesign note↩ 05 · governance note
Industry signals
- I1
GitHub — Meet the new coding agentAnnouncement
Bounded coding-agent launch with a human-controlled pull-request workflow.
A product announcement, not a measured outcome.
- I2
Thomson Reuters — CoCounsel Legal and CoCounsel Tax launchAnnouncement
Legal and tax product announcement combining professional content with assisted workflows.
A product announcement, not a measured outcome.
- I3
BNY — Unlocking the potential of an enterprise AI platformOperator report
Operator-reported digital employees with credentials and supervisors.
Reported by the operator; not independently verified.
- I4
Microsoft — Dragon Copilot extended to nurses and partnersAnnouncement
Nursing documentation with required review and approval, not autonomous clinical judgement.
A product announcement, not a clinical outcome.
- I5
Siemens — AI agents for industrial automationAnnouncement
Industrial architecture announcement with available, preview and future elements distinguished.
Availability varies by component; not a measured outcome.
- I6
Walmart — All in on agentsAnnouncement
Four-super-agent framework, partly live and partly planned.
A framework description, not a measured outcome.
- I7
UK DSIT — Humphrey reviews consultation responses for the first timeAnnouncement · trial
Live consultation-analysis trial; every response also human-reviewed.
A trial, not a routine service; projected savings are the publisher’s.
- I8
Schneider Electric — One Digital Grid PlatformAnnouncement
Announced grid platform integrating planning, operations and flexibility.
No claim of autonomous switching or achieved savings.
Contents