Increasingly capable AI agents can already research markets, generate content and act across revenue systems. The larger question is whether today's fragmented go-to-market stack can give them the context, authority and control required to operate reliably.
Before the revenue team begins work, hundreds of specialised processes have already run.
Some have reviewed overnight changes across target accounts. Others have analysed sales calls, product usage and support conversations. They have identified possible buying signals, prepared account briefings, proposed audience changes, drafted campaign variants and flagged opportunities requiring human attention.
Most would not be persistent digital employees. They would be temporary agent invocations: software given an objective, a set of tools and permission to complete a bounded task.
That distinction matters. The near-term future of go-to-market (GTM) is unlikely to be one omniscient AI running the revenue organisation. It is more likely to involve networks of specialised agents operating across customer data, content, communication and workflow systems.
The strategic constraint will not simply be whether a model can write an email or research an account. It will be whether the organisation can give large numbers of agents reliable context, bounded authority and measurable feedback without losing control of the customer relationship.
The AGI Question
The practical threshold is operational
Public forecasts about artificial general intelligence create the impression that businesses are waiting for one recognisable technological event.
They are not.
In September 2024, Sam Altman wrote that superintelligence could be possible within a few thousand days, while accepting that it might take longer. Dario Amodei has argued that the kind of “powerful AI” he envisages could appear as early as 2026, although he also says it might take much longer or never arrive in that form. Demis Hassabis has generally discussed a five-to-10-year horizon for AGI rather than one fixed date.
The predictions also refer to different things. OpenAI has described AGI as systems generally smarter than humans. Google DeepMind has used a definition based on matching human capability across most cognitive tasks. Other definitions emphasise adaptability, economic usefulness, autonomous learning or breadth across domains.
These are forecasts by leaders whose organisations are investing heavily in the outcome. They deserve attention, but not the status of a planning calendar.
The more useful question for CEOs, CROs and CMOs is operational: when will agents become capable and reliable enough to change how revenue work is organised?
That transition has already begun, although it remains uneven. Gartner reported in December 2025 that 81% of 413 marketing technology leaders surveyed earlier that year were piloting or had implemented AI agents. The statistic measures reported activity, not successful autonomy, sustained adoption or return on investment.
Current agents can research accounts, enrich records, analyse conversations, generate content and initiate bounded actions. They cannot yet be assumed to manage complex revenue functions independently. A 2025 study covering 306 practitioners found that 68% of production agents performed no more than 10 steps before requiring human intervention, while 74% relied primarily on human evaluation. Reliability remained the leading development challenge. The study covered multiple domains rather than GTM specifically, but its finding is a useful warning against equating technical demonstrations with operational autonomy.
GTM does not need AGI to change. It needs agents that can perform enough linked work to make the current division between people, processes and applications inefficient.
The Point-Solution Era
Specialisation created a coordination problem
The current GTM stack emerged through specialisation.
Customer relationship management platforms organised accounts, contacts and opportunities. Marketing automation systems managed campaigns and lead flows. Sales engagement platforms handled sequences. Customer data platforms unified profiles. Revenue-intelligence tools analysed calls and forecasts. Enrichment providers supplied external information.
Each category solved a legitimate problem. Collectively, they produced overlapping data, duplicated workflows and fragmented customer context.
A marketing platform may regard someone as a qualified lead while the CRM records an inactive opportunity. A product system may show intensive usage while customer success records unresolved dissatisfaction. An intent provider may flag an account that has already opted out of contact. A person can sometimes interpret those contradictions. An agent requires them to be resolved through data authority, policy or explicit escalation.
This is why the point-solution problem becomes more serious under agentic operations. Fragmentation is not merely inefficient. It creates ambiguity about what the system knows, which source it should trust and what it is allowed to do.
The market is showing some appetite for consolidation. Gartner's 2025 martech research highlighted composable stacks and capability-focused operations, while its 2026 CMO guidance noted active interest in consolidation as vendors added overlapping AI functions. That does not prove that one platform will absorb the entire stack. It shows that the cost of coordination is becoming a more visible executive concern.
AI may accelerate consolidation, but it could also produce the opposite result. If agents can operate tools through standard interfaces, organisations may combine more specialist systems without requiring employees to navigate each one directly.
The likely battle is therefore not simply between consolidation and fragmentation. It is over who controls the coordination layer.
The Architectural Shift
An AI interface is not an AI-native operating model
The criticism that incumbents are placing conversational interfaces over static systems contains some truth, but it is becoming outdated.
Salesforce now provides agent APIs, an agent scripting language, command-line development, software development kits, programmatic testing and trace export. It can expose configured Agentforce agents as Model Context Protocol tools, allowing external assistants to invoke Salesforce capabilities while retaining user permissions, sharing rules and field-level controls. Some Salesforce Model Context Protocol command-line functions remain in Developer Preview.
HubSpot offers a remote Model Context Protocol server with authorised read and write access to CRM data, alongside a local developer server. In June 2026 it released an Agent CLI in public beta for background and headless work, including searching, creating and updating CRM objects, pipelines and workflows.
Clay combines external signals, enrichment, AI-assisted research, message generation, routing and sequencing. Its product materials illustrate how GTM workflows are beginning to merge data, timing and execution, although its performance and accuracy claims remain vendor-reported.
These developments are more than interface design. They show established and emerging vendors making their systems operable by agents.
Systems of record are also not becoming irrelevant. Agents need dependable records for contracts, customer permissions, opportunity state, account ownership and previous activity. As autonomous action expands, structured records may become more important because they provide control points and audit evidence.
The limitation is not that CRM stores records. It is that no single record system necessarily has the context or authority to coordinate the whole revenue environment.
An AI-native operating model is organised around goals, context, state, tool selection, action and feedback. It still uses records and deterministic rules, but it does not require a person to configure every path in advance.
GTM Infrastructure
The control environment agents require
“GTM infrastructure” is not yet an established software category. A genuine category would require more than funding announcements and vendor language. It would need:
- A consistent problem that existing categories do not solve
- An identifiable buyer and independent budget
- A repeatable product boundary
- A recognised set of providers and evaluation criteria
The market has not clearly passed those tests.
The underlying architectural requirement is more credible. Organisations deploying multiple GTM agents need a shared control environment connecting seven capabilities.
Direction and control translate commercial intent into objectives, budgets, constraints, approval thresholds and escalation routes.
The interface might be conversational, visual or programmatic. The important change is that users define desired outcomes and boundaries rather than configuring every individual step.
Knowledge and context give agents access to product information, positioning, pricing rules, customer history, brand guidance, competitive intelligence and previous campaign outcomes.
Access alone is insufficient. The system must know which source is authoritative, when it was updated, who may use it and what to do when sources conflict. This creates a form of governed GTM institutional memory.
Data, signals and operational state include customer records, product usage, intent signals, conversations, campaign performance and external events.
Operational state adds a different question: what is happening now? Which agent is already acting? Which approval is pending? Which contact was recently approached? Which budget has been committed?
Without shared state, several individually competent agents can create a collectively incoherent customer experience.
Reasoning and orchestration interpret objectives, decompose work, select tools, assign tasks, preserve state and handle exceptions.
This is the most plausible strategic control point because it determines which system acts and why. It may be embedded in an incumbent platform, supplied by a new vendor or built internally.
Actions and execution provide approved interfaces through which agents can research, enrich, generate, publish, contact, update and trigger.
Actions should be classified by risk:
- Recommendations
- Drafts requiring approval
- Reversible actions
- Limited autonomous actions
- High-impact or irreversible actions
Permission should follow the potential consequence, not the apparent simplicity of the command.
External platforms and models provide CRM, email, advertising, communication, analytics, enrichment and model capabilities.
GTM infrastructure is therefore more likely to operate above these systems than replace all of them. Its value lies in coordinated use, not in reproducing every specialist capability.
Governance, evaluation and learning establish agent identity, least-privilege access, traceability, quality evaluation, rate limits, budgets, incident response and lifecycle ownership.
This layer also creates the conditions for improvement. A genuine learning loop requires a defined objective, trusted feedback, controlled experimentation, rollback and authority to change behaviour.
An agent observing its own output is not organisational learning.
From Workflows to Loops
Adaptation must remain controlled
Traditional workflow automation follows a predetermined sequence. When an event occurs, the system performs configured steps.
Agentic orchestration can choose among different routes. It may gather more information, select a specialist tool, request approval or revise its plan when circumstances change.
An emerging practice sometimes described as loop engineering focuses on the whole cycle:
- Observe the environment
- Receive or infer a bounded objective
- Retrieve relevant knowledge
- Develop a plan
- Execute permitted actions
- Measure the outcome
- Evaluate performance
- Revise the next action
- Repeat
This differs from prompt engineering because it includes state, data, tools, controls and measurement. It differs from conventional automation because the route may adapt. It differs from reinforcement learning because most enterprise deployments will not continuously retrain model weights.
Improvement will usually remain a designed process. Teams will collect traces, label failures, update prompts, alter tool definitions, revise retrieval sources and run controlled evaluations before releasing a new version.
Current model platforms are being built around this requirement. OpenAI's guidance, for example, recommends inspecting end-to-end traces during development and moving to repeatable datasets and evaluation runs when teams need systematic comparison and regression testing.
The commercial danger is local optimisation. An agent can increase open rates by using more provocative subject lines while damaging trust. It can book more meetings by lowering qualification standards. It can improve short-term conversion by making claims the business cannot sustain.
The objective function must represent customer value, margin, risk and long-term performance, not only the metric easiest to measure.
The Operating Model
GTM roles will be redesigned around decision rights
Agentic systems redistribute work before they remove roles.
Sales development representatives are likely to delegate more account research, enrichment, prioritisation and message preparation. Their work can move towards qualification, conversation, play design and exception handling.
Account executives may spend less time preparing meetings and updating records. Their differentiation will lie increasingly in stakeholder judgement, negotiation, trust and the ability to recognise when the apparent next action is commercially wrong.
Marketing teams may supervise systems that create, adapt and test content. As generation becomes cheaper, the scarce capabilities become evidence, originality, brand judgement, audience understanding and the design of useful experiments.
Customer-success agents can monitor health signals, prepare interventions and maintain routine communication. People remain essential when the situation involves emotion, ambiguity, contractual discretion or a damaged relationship.
Revenue operations may undergo the deepest change. Its remit can expand from configuring systems and reporting activity to designing agent tools, permissions, evaluation criteria, state models and escalation logic.
Technology, data, security and legal teams will also move closer to revenue execution. When an agent can alter customer communication or commercial records, architecture and governance are no longer back-office concerns.
The central operating-model question is not which department “owns AI”. It is who defines the objective, who grants authority, who evaluates performance and who remains accountable when the system acts.
The Platform Battle
The most credible future is federated
Five market outcomes remain plausible.
A giga-platform could combine data, workflows, agents and governance. It would reduce integration complexity, but increase lock-in and make the organisation dependent on one provider's data model and roadmap.
An agentic control layer could coordinate objectives and policies above existing systems. This is the clearest architectural expression of GTM infrastructure.
A composable agent ecosystem could connect specialist agents through APIs and protocols such as the Model Context Protocol. It would maximise choice, but place more responsibility on the enterprise to manage identity, state, observability and failure.
Incumbent reinvention could absorb much of the control layer into Salesforce, HubSpot, Microsoft, Adobe or another established environment. Data gravity, installed distribution and mature permissions give incumbents substantial advantages.
Finally, limited autonomy could remain the dominant model. Security, data quality, customer expectations and regulation may restrict autonomous action to narrow, reversible work.
The most credible near-term outcome is federated: established systems of record remain in place, specialist agents perform bounded work and an orchestration layer coordinates context, permissions and evaluation across them.
That layer may become a standalone product category. It may also become an architectural function distributed across several platforms.
The Risk Surface
Greater autonomy requires stronger controls
An agent that proposes an email creates a review problem. An agent that selects the recipient, sends the message, updates the CRM and increases campaign spend creates an operational-risk system.
Failures can include unauthorised outreach, hallucinated claims, privacy breaches, discriminatory targeting, damaged deliverability, incorrect records and uncontrolled spend.
Connected agents also introduce technical threats. Malicious information can manipulate tool selection through prompt injection. Excessive credentials can expose customer data. Compromised integrations can turn routine agent access into account takeover. Several agents acting against inconsistent objectives can create runaway campaigns or repeated customer contact.
Agent identity therefore matters. Each agent or workload needs an attributable owner, scoped credentials, action-level permissions, rate limits and budget constraints.
The Model Context Protocol ecosystem is developing authorisation mechanisms for machine-to-machine services and enterprise-managed identity, but support must be implemented by both clients and servers. The existence of a specification does not establish secure deployment.
A production control model should include:
- Least-privilege access
- Separate identities for agents and services
- Human approval for high-impact actions
- Data and consent controls
- Full action and tool traces
- Quality and policy evaluations
- Rate and budget limits
- Rollback or compensating actions
- Incident response
- Named ownership and retirement processes
Governance must operate at the moment of action. A quarterly committee cannot supervise a system making decisions continuously.
Immediate Actions
Understand the current GTM environment
Map the current GTM stack, including duplicated data, brittle integrations and unofficial workflows.
Identify authoritative sources for customer, product, commercial and consent information.
Select tasks where agent assistance can be measured without granting broad autonomy. Begin with recommendations, research and drafts.
Classify potential actions according to sensitivity, reversibility, customer impact and financial consequence.
Define outcome measures before launching the pilot. Faster activity is not sufficient evidence of better GTM.
Capability Building
Create the foundations for governed execution
Create governed access to organisational knowledge, with provenance, effective dates and ownership.
Improve identity resolution, semantic consistency and operational-state management.
Develop reusable agent tools rather than one-off connections.
Give agents distinct identities and scoped permissions.
Capture traces and build evaluation datasets from realistic GTM situations, including failure and refusal cases.
Maintain an agent registry recording purpose, owner, model, tools, data access, risk class, version and lifecycle status.
Strategic Actions
Design the AI-native GTM operating model
Define the target AI-native GTM architecture and the continuing role of each system of record.
Decide where cross-platform orchestration should sit and which capabilities must remain under organisational control.
Assess suite, control-layer and composable options against lock-in, flexibility, governance and total operating cost.
Redesign roles around task movement, decision rights and accountability rather than assuming whole functions disappear.
Build GTM engineering as a shared capability spanning revenue operations, data, technology and governance.
The decision does not depend on predicting the date of AGI. Even if progress slows, these capabilities improve the coherence of today's fragmented revenue environment.
The future of GTM will not be determined by who deploys the most agents. It will be determined by who can turn their actions into governed, repeatable learning.
