Strategic Signal

AI assistants are reaching business software. Who keeps the economics?

Strategic Signal · November 2026 Market Outlook · Month-to-date evidence through November 19

Author and editor: Mike Ye. AI research analyst: Ella.

AI assistants are becoming another way to use business software. In our November sample, they can retrieve company information, initiate approved changes and return work to the systems that govern it. The strategic question is who gets paid when the conversation moves into the assistant but the underlying records, business rules and permissions remain elsewhere.

Our assessment: these arrangements create a negotiation over the customer relationship. An assistant can become the place a user starts work. A software provider can remain necessary to authorize, interpret and complete it. The November evidence establishes that separation of roles; it does not yet establish which side captures more revenue.

What the November sample shows

The frozen November 19 registry contains 296 arrangements, including 291 reviewed records and five provisional legacy records. Of the reviewed records, 44 carry announcement dates from November 1 through November 19.

Feature in the November announcement cohortRecords, out of 44
Retrieval access43
Professional workflow access32
Action execution27
Connector mechanism33
Strict data-rights qualification0

These categories overlap. Professional workflow access is a separate facet from permission to execute an action. Strict qualification requires evidence of separately material proprietary data rights under our existing test; a connector does not automatically meet it.

This is a selected research sample. The zero in the last row does not establish that no qualifying data-rights transactions occurred in November.

1. The assistant can gain the interaction while the provider retains control

LeanData made its combined orchestration and scheduling connection available to Premium customers on November 18. Claude and ChatGPT can initiate supported work, while LeanData's existing routing rules and permissions govern the actions. Record: SS-AI-2026-F6948782.

ThoughtSpot connects Claude Data to governed business definitions and lets users turn an artifact into a ThoughtSpot Liveboard. Record: SS-AI-2026-9FBB5BAF.

The distinction matters. The assistant supplies a conversational entry point; the provider supplies rules, context or the destination for completed work. Our interpretation is that providers retaining an indispensable role in completing a task have a stronger basis for bargaining than providers supplying an interchangeable information feed. Pricing and retention evidence must establish whether that advantage becomes financial.

2. Permission to act is becoming a product boundary

Infor's November 12 announcement describes approved changes to manufacturing records from Claude. Availability is limited to selected Velocity Suite customers, with a separate Claude license and administrator activation. Existing Infor permissions apply, and each change requires user approval. Record: SS-AI-2026-BDD980ED.

Wolters Kluwer's TyMetrix 360° connection supports permission-checked legal operations, including invoice and matter actions. Record: SS-AI-2026-41085518.

These examples extend October's execution thesis: the commercially relevant asset can include the authority and business logic needed to change an operational record. A model's ability to propose an action is different from the customer's permission to carry it out.

For buyers assessing software businesses, the diligence question is concrete: which steps still require this provider if the user starts the task in an outside assistant? That identifies a dependency to investigate; it does not, by itself, prove a durable moat.

3. Several assistants can reach the same asset

Impartner's November 17 release supports Claude, ChatGPT, Copilot Studio, Gemini Enterprise and Claude Code. Retrieval follows the user's Impartner permissions, with sensitive fields withheld. Record: SS-AI-2026-348B3C5B.

Together with LeanData's support for both Claude and ChatGPT, this shows why an integration should not automatically be treated as exclusive distribution. Providers can make the same underlying system available through multiple assistant interfaces.

Our inference is that providers supporting several independently operated assistants keep their options open, weakening any one assistant's leverage over them. An assistant could gain leverage if customers concentrate their work in it. Switching costs, usage concentration and contract terms will decide the outcome; compatibility announcements cannot show market share or dependence.

4. Access is observable; its financial return is still harder to measure

Thirty-seven of the 44 November records contain at least one capability classified as enabled; 13 contain at least one classified as intended. These groups overlap because one arrangement can contain both available and planned functions. None has a capability classified as demonstrated in this frozen snapshot.

That is a statement about our recorded evidence classifications. It is not proof that customers have obtained no benefits, and it is not a count of independently measured economic outcomes. An available connector, an active customer, an operational improvement and a financial return are four different observations.

We should follow known arrangements through those steps. The most useful next disclosures would identify the workflow used, a quantitative result, its measurement period and a basis for attributing the result to that workflow.

Pricing evidence: what the Tempus numbers do and don't show

No comparable data-learning-license valuation can be derived from the November cohort. Recent evidence outside that cohort illustrates why scope matters.

Tempus's October 23 SEC filing describes two arrangements. Bristol Myers Squibb pays $137.5 million, invoiced quarterly through 2031, for oncology records, Lens software and services, without allocating the amount among them. Pathos AI licenses up to 500,000 neurology records, targeted at 100,000 annually for five years, with rights to sublicense the data and agentic tool to Bristol Myers Squibb. Fees total up to $55 million: $100 per delivered record, up to $50 million, plus $5 million for updates. Fees depend on delivery; contractual amounts are not verified cash receipts. Records: SS-AI-2026-1114B90D and SS-AI-2026-9E43D368.

The filing attributes the Tempus–Pathos related-person relationship to overlapping executive leadership and equity ownership; a special committee of disinterested directors approved the neurology agreement and side letter. That relationship, delivery terms and bundled agent access make $100 a transaction-specific reference, not a market clinical-data price. Neither amount establishes November's software-connection economics.

Better historical coverage is separate from November activity

The registry first added 104 reviewed records from November 1 through this snapshot: the 44 November announcements and 60 with earlier dates. Of those 60, 53 came from the January 2025–June 2026 historical backfill published November 18; the other seven are October announcements recorded in early November. Counting from the October 21 snapshot, 177 records were added in all: 73 from October 22–31 and these 104. The reconciliation is 119 + 73 + 104 = 296.

Among records already present on October 21, one was revised: the Google News AI publisher pilot, SS-AI-2026-154749D9. Its October 27 review adds Reach plc and supporting evidence to the existing program. That revision is not counted among the 177 additions.

Those additions improve the history available for analysis. They are not evidence of a November surge in deal activity. Likewise, growth from 119 arrangements in the October 21 snapshot to 296 in this snapshot measures expansion of our registry, not a market growth rate.

Follow-up on our existing forecasts

The following are interim assessments. Original wording, issue dates, horizons and resolution tests remain unchanged.

ForecastEvidence added by this reviewInterim status
SS-FC-2026-001: controlled access and learning arrangements versus ownership transfersNone of the 44 November records has classifiable ownership terms (all are undisclosed), so November adds no observation toward the test's five-observation minimum. The test resolves at the September 16, 2027 horizon.Unresolved
SS-FC-2026-002: two unrelated industrial model-learning arrangementsInfor adds manufacturing workflow evidence. Its recorded training rights are undisclosed, so it does not establish a qualifying learning-rights arrangement.Unresolved
SS-FC-2026-003: another disclosed proprietary data-learning-rights price reference beyond TempusThe October Tempus filing adds two arrangements, but neither meets the test: oncology consideration is bundled and unallocated; neurology fees are per delivered record with training rights undisclosed. Both are Tempus agreements, while the forecast asks for a reference beyond Tempus. November adds none.Unresolved
SS-FC-2026-004: measured execution outcomes across three industriesThe 27 records that permit action execution identify follow-up candidates. Classification as enabled does not establish a numeric result, defined metric, measurement period or attribution under the original test.Unresolved

The separate Frontier Value Capture forecast and its six industry predictions also remain unresolved in this issue. More access does not automatically mean downstream revenue participation. VerticalScope's November 10 disclosure of an October agreement does not count toward that test: there the frontier lab pays the content owner usage-linked fees, while the forecast tracks frontier labs earning downstream economics beyond tokens. The usage-linked detail for that agreement also rests on management's earnings-call comments, not a published contract. Record: SS-AI-2026-4A687F13.

These are interim observations, not completed horizon-end searches or failure verdicts. This issue introduces no new scored forecast.

What we will track next

Three developments would materially advance the analysis: providers disclosing how they charge for assistant-mediated work; customers showing repeat use and measured results; and terms showing who controls access when an assistant or software relationship changes.

Our working conclusion is that moving the interaction into an assistant can redistribute bargaining power without eliminating the underlying software provider. The economic result depends on whether that provider remains necessary—and whether it can charge for the role it retains.


Evidence basis: frozen canonical archive 000045, archived November 19, 2026 at 16:57:47 UTC; ontology 2.0.0; data-rights release 1.0.44; capability release 1.0.264. The comparison uses archive 000012, frozen October 21. The issue deliberately uses the November 19 frozen state, not a live latest alias. An independently established complete market-coverage cutoff is unknown. Sources support their stated claims; our bargaining-power conclusions are analysis.

Companion interpretation: The interface can move. The dependency can stay. · Distinct analysis by the same author; not independent corroboration.

Related research: October outlook · Original forecast register · AI Access Registry

Published by exmxc. Not investment advice.

Original forecast definitions and assessment history

4 dated forecasts

These forecasts retain their original issuance dates, observation windows, wording and criteria. Dated assessments below append evidence without reissuing the forecasts. These are Strategic Signal analysis, not disclosed facts; confidence is qualitative.

SS-FC-2026-001 · moderate confidence

The next sufficiently observable cohort will favor learning licenses, controlled access and co-development over outright data ownership transfers.

Mechanism: Continuing owners can monetize or benefit from learning while retaining operational access and protecting confidential source data.

Scope: Newly announced, subsequently human-approved transactions within the horizon; exclude historical backfill and updates to existing deals.

Assumptions: Meaningful training permission can be separated from ownership. Public disclosure will describe enough rights to classify at least five future observations.

Alternative: Distressed archive sales or exclusive dataset acquisitions become the more observable transaction form.

Confidence: All four opening cases fit the mechanism, but the sample is selected and small.

Confirmation: At horizon end, at least five approved new observations have classifiable ownership terms and more than half explicitly use licenses, local learning or controlled access without a disclosed transfer of underlying data ownership.

Invalidation: With at least five classifiable observations, half or fewer meet that test.

Unresolved: Fewer than five observations have sufficiently clear rights. Unspecified ownership is not counted as retained ownership.

Supporting records: SS-AIDT-2025-0001 · SS-AIDT-2025-0002 · SS-AIDT-2026-0004 · SS-AIDT-2026-0007

Originally issued 2026-09-16 · Observation window 2026-09-17–2027-09-16 · Original forecast and criteria

SS-FC-2026-001 · 2026-09-30 · unresolved

Interim unresolved. OneMedNet is one conservative eligible new classifiable observation. Exclude Harell under the corrected September 15 source-stated first-disclosure treatment, GuideAI as a renewal, and pre-horizon backfill. Threshold of at least five and horizon-end majority test are not met.

Supporting records: SS-AIDT-2026-0011 · SS-AIDT-2026-0013 · SS-AIDT-2026-0014 · SS-AIDT-2026-0009 · SS-AIDT-2026-0012

SS-FC-2026-001 · 2026-10-13 · unresolved

Interim unresolved: OneMedNet, the antibody federation and OpenAI-Synopsys supply retained-control learning evidence, but enough new observations with classifiable ownership terms and the horizon-end majority are not established. Ironclad and Biohub require conservative treatment.

Supporting records: SS-AIDT-2026-0014 · SS-AIDT-2026-0015 · SS-AIDT-2026-0016 · SS-AIDT-2026-0017 · SS-AIDT-2026-0018

SS-FC-2026-001 · 2026-10-21 · unresolved

The October 15 Google publisher licensing record establishes paid governed content access but does not disclose underlying ownership transfer or model-training rights. The required eligible ownership-term cohort and horizon-end majority test remain unresolved.

Supporting records: SS-AIDT-2026-0019 · SS-AIDT-2026-0014 · SS-AIDT-2026-0016

SS-FC-2026-001 · 2026-11-19 · unresolved

None of the 44 November records has classifiable ownership terms (all are undisclosed), so November adds no observation toward the test's five-observation minimum. The test resolves at the September 16, 2027 horizon.

SS-FC-2026-002 · low confidence

At least two unrelated industrial data owners will announce qualifying learning-rights arrangements involving manufacturing or equipment-maintenance histories.

Mechanism: Expensive failures and repeated interventions can provide task-specific learning examples that generic public text does not reproduce.

Scope: Manufacturing process, quality/yield, industrial equipment or maintenance corpora; exclude software-only archives, ordinary deployments and two announcements from the same owner group.

Assumptions: Owners can link records to interventions or outcomes. At least some owners can grant usable learning rights and disclose the arrangement.

Alternative: Learning stays internal or agreements disclose only software deployment, leaving no qualifying observable rights transaction.

Confidence: An adjacent energy precedent supports the logic, but the approved baseline has no direct manufacturing precedent.

Confirmation: By horizon end, two qualifying announcements within the horizon from unrelated owner groups are approved with explicit model-learning use.

Invalidation: After a documented horizon-end sector search, fewer than two qualifying announcements are found. This misses the observable-deal prediction, not proof that private arrangements do not exist.

Unresolved: Evidence or final sector search remains incomplete; no automatic success based on deployment announcements.

Supporting records: SS-AIDT-2026-0007

Originally issued 2026-09-16 · Observation window 2026-09-17–2027-09-16 · Original forecast and criteria

SS-FC-2026-002 · 2026-09-30 · unresolved

Interim unresolved. Zero qualifying approved manufacturing or equipment-maintenance arrangements from unrelated owner groups in this snapshot. No horizon-end sector search has occurred; do not infer failure or absence of private deals.

Supporting records: SS-AIDT-2026-0011 · SS-AIDT-2026-0013 · SS-AIDT-2026-0014 · SS-AIDT-2026-0009 · SS-AIDT-2026-0012

SS-FC-2026-002 · 2026-10-13 · unresolved

Interim unresolved: there are not two qualifying new model-learning deals from unrelated manufacturing or equipment-maintenance data owners. The historical provisional Caterpillar record is excluded; engineering-tool and biology arrangements do not meet this forecast scope.

Supporting records: SS-AIDT-2026-0014 · SS-AIDT-2026-0015 · SS-AIDT-2026-0016 · SS-AIDT-2026-0017 · SS-AIDT-2026-0018

SS-FC-2026-002 · 2026-10-21 · unresolved

Additional aviation maintenance and industrial workflow access appear in the canonical capability registry, but no two unrelated, newly announced and human-approved industrial model-learning-rights arrangements satisfy the original confirmation test. The 2027 horizon remains open.

Supporting records: SS-AIDT-2026-0016 · SS-AIDT-2026-0007

SS-FC-2026-002 · 2026-11-19 · unresolved

Infor adds manufacturing workflow evidence. Its recorded training rights are undisclosed, so it does not establish a qualifying learning-rights arrangement.

SS-FC-2026-003 · low confidence

At least one newly announced transaction will disclose consideration specifically allocated to proprietary corporate data learning rights, adding a pricing reference beyond Tempus.

Mechanism: Explicit licenses create payment obligations that may appear in filings or transaction announcements even when larger programs bundle services and compute.

Scope: A new owner/buyer arrangement announced within the horizon; disclosed amount, currency, consideration scope and payment form must be distinguishable. No minimum dollar threshold.

Assumptions: Some counterparties disclose the license leg separately. A monetary allocation reflects a data right rather than merely a total program budget.

Alternative: Material rights remain hidden inside co-development, equity or service agreements without separate pricing.

Confidence: One clear historical allocation establishes possibility, not its future frequency or price level.

Confirmation: At least one approved transaction meeting scope is supported by a filing, contract or company disclosure before horizon end.

Invalidation: No such transaction is found after a documented horizon-end pricing search. A bundled headline or pending bid does not satisfy the test.

Unresolved: Source access or final search is incomplete. A subsequently discovered pre-horizon deal is historical backfill and cannot count.

Supporting records: SS-AIDT-2025-0002 · SS-AIDT-2026-0007

Originally issued 2026-09-16 · Observation window 2026-09-17–2027-09-16 · Original forecast and criteria

SS-FC-2026-003 · 2026-09-30 · unresolved

Interim unresolved. OneMedNet supplies directional support for separately material data-license consideration but the reviewed release does not state explicit currency or settlement form. Do not add a new exact-price or named-buyer requirement: neither is stated in the original forecast. Silencio and Tempus–Recursion are excluded pre-horizon/historical arrangements.

Supporting records: SS-AIDT-2026-0011 · SS-AIDT-2026-0013 · SS-AIDT-2026-0014 · SS-AIDT-2026-0009 · SS-AIDT-2026-0012

SS-FC-2026-003 · 2026-10-13 · unresolved

Interim unresolved: OneMedNet does not disclose currency or payment form, Synopsys does not disclose allocated amounts, and Biohub's commercial payments are tied to a bundled scientific program, not a separately priced qualifying data license. Historical pricing references cannot confirm a post-issuance forecast.

Supporting records: SS-AIDT-2026-0014 · SS-AIDT-2026-0015 · SS-AIDT-2026-0016 · SS-AIDT-2026-0017 · SS-AIDT-2026-0018

SS-FC-2026-003 · 2026-10-21 · unresolved

Google News AI pays publishers under licensing arrangements but discloses neither allocated monetary amounts nor payment form. It cannot satisfy the original requirement for a comparable disclosed data-learning-rights allocation; the test remains unresolved.

Supporting records: SS-AIDT-2026-0019 · SS-AIDT-2026-0014

SS-FC-2026-003 · 2026-11-19 · unresolved

The October Tempus filing adds two arrangements, but neither meets the test: oncology consideration is bundled and unallocated; neurology fees are per delivered record with training rights undisclosed. Both are Tempus agreements, while the forecast asks for a reference beyond Tempus. November adds none.

SS-FC-2026-004 · low_moderate confidence

At least five new governed AI-access arrangements across three distinct industries will publicly report measurable operational outcomes from AI-enabled workflow execution, not merely access or availability.

Mechanism: When AI is permissioned to act on domain-specific systems, some providers should progress beyond capability announcements to measurable work completed, cost or cycle-time changes, or independently auditable task outcomes.

Scope: Arrangements announced October 23, 2026 through October 22, 2027; require later human review, explicit governed access to operational systems, reported AI-enabled workflow execution, and attributable outcome measurement. Backfills, vendor demos, generic adoption claims, retrieval-only use, and previously announced arrangements are excluded. Industries must be source-supported at evaluation, not guessed from subject_type.

Assumptions: Some providers will disclose outcome methods and meaningful operational baselines rather than only claiming productivity. The registry will capture source documents with enough specificity to distinguish measured results from vendor demonstrations.

Alternative: Most integrations remain discovery, retrieval, or promotional capability claims; outcomes are undisclosed or impossible to attribute to the AI workflow.

Confidence: The 119-record registry contains 78 action-execution classifications, but evidence quality is dominated by vendor-described enablement or intent and independently demonstrated outcomes remain rare.

Confirmation: By 2027-10-22, five newly announced, subsequently reviewed arrangements across at least three source-confirmed sectors each disclose a numeric operational outcome, explicit metric definition and measurement period, and an identified AI-execution workflow with governed access. The primary source must explain the basis for attributing the result to that workflow; an independent study is stronger but not required.

Invalidation: After a documented horizon-end review and search, fewer than five arrangements meet all conditions, or fewer than three sectors qualify.

Unresolved: If primary-source evidence is insufficient to assess workflow execution, attribution, measurement period or sector, keep the relevant observation ineligible and the aggregate prediction unresolved until the final documented search.

Supporting records: SS-AIDT-2026-0016 · SS-AIDT-2026-0017 · SS-AIDT-2026-0014

Originally issued 2026-10-22 · Observation window 2026-10-23–2027-10-22 · Original forecast and criteria

SS-FC-2026-004 · 2026-11-19 · unresolved

The 27 records that permit action execution identify follow-up candidates. Classification as enabled does not establish a numeric result, defined metric, measurement period or attribution under the original test.

Versioned outlook JSON · Frozen measures and evidence hashes