Most B2B organisations are accumulating AI capabilities one tool at a time.
Marketing gets AI for content creation. Sales gets automated prospecting and meeting intelligence. RevOps adds predictive scoring. Customer Success experiments with churn models. CRM vendors add copilots and agents. Specialist vendors add another layer of automation around them.
Each investment can make sense individually. But together they leave a more important question unanswered:
What would the commercial organisation look like if we designed it as one system?
That question is becoming increasingly practical.
Microsoft now has agents for lead qualification and opportunity management. Adobe is introducing agents for audience creation, buying groups and journey orchestration. 6sense can make buying-stage and intent intelligence available directly to external AI agents. Gong is developing an execution layer for governed revenue agents.
These are still different technologies at different levels of maturity. But together they point towards a different commercial architecture.
Start with the commercial process, not the AI tools
I would map the future B2B commercial engine across eight connected domains:
- Sense the market
Continuously detect customer needs, competitor activity, market developments, intent, company events and emerging commercial opportunities. - Choose where to play
Define markets, ICPs and segments, construct account universes and buying groups, score opportunities and allocate commercial resources. - Create demand
Develop propositions, campaigns and content; activate paid, owned and account-based channels; personalise digital experiences and continuously experiment. - Identify and engage demand
Recognise accounts and people, interpret intent, enrich data, prioritise prospects, research accounts, conduct inbound and outbound engagement, qualify interest and arrange meetings. - Develop opportunities
Prepare and analyse conversations, map stakeholders, assess opportunity health, recommend next actions, identify deal risk and manage pipeline and forecasts. - Design and close the deal
Configure solutions, develop business cases, determine pricing, answer RFPs, create proposals, negotiate terms, manage approvals and contracts, and complete the transaction. - Grow customer value
Transfer sales context into onboarding, monitor adoption and customer health, predict renewal risk, identify expansion opportunities and develop advocates. - Learn and optimise
Connect wins, losses, conversations, customer behaviour, campaign performance, profitability and experiments back into propositions, targeting and commercial strategy.
None of these domains is new.
What is changing is how much of the work between them can be connected.
The real architecture is signal to action
Consider a job change at a target account.
Today, that event might appear in a sales intelligence platform. A salesperson may notice it, research the new executive, decide whether it matters, find previous account interactions, develop an outreach hypothesis, write an email, remember to follow up and update the CRM.
In an increasingly agentic model, the sequence could become:
signal detected → account identified → commercial relevance assessed → buying group updated → account reprioritised → research performed → proposition selected → message prepared → policy checked → outreach initiated → response interpreted → meeting scheduled → CRM updated
The interesting technology is not any individual step.
It is the connection between them.
That suggests a different way to think about the commercial technology stack. Instead of organising it primarily around applications, I would organise it around six questions:
What can we sense?
What context do we have?
What decision needs to be made?
Who or what should make it?
What action can be executed?
How do we learn from the outcome?
AI makes each of these layers more powerful. Agents potentially connect them.
Not every decision should become autonomous
This is where I think the discussion about AI automation becomes too simplistic.
Technical feasibility is not the same as desirable autonomy.
Automatically enriching a company record is very different from determining the concession strategy for a strategic contract. Generating a meeting summary is different from interpreting the politics inside a buying committee. Automatically nurturing thousands of low-intent prospects is different from contacting the CEO of your largest account.
I therefore find it more useful to think in levels of commercial autonomy.
At the lowest level, people execute the work and AI provides assistance.
At the next level, AI recommends an action but a person decides.
Then AI can prepare an action for human approval.
Beyond that, agents can decide and execute within explicitly defined boundaries.
Eventually, some workflows may become adaptive: the system observes results and changes its approach within those boundaries.
The highest level should not automatically be the ambition.
A lead-routing decision may be almost completely autonomous. A major pricing exception probably should not be. And in a complex enterprise negotiation, human attention may itself be part of the value delivered to the customer.
The design question is therefore not simply:
Can we automate this?
It is:
What autonomy do we want to delegate here, under which conditions, and who remains accountable?
This also changes commercial roles
If this architecture develops further, marketing and sales work does not simply move from humans to AI.
The human role changes.
I see at least five roles becoming more important.
The strategist determines markets, propositions, objectives and trade-offs.
The relationship owner understands trust, organisational politics, ambiguity and the context that is difficult to capture in systems.
The expert contributes judgement where product, industry, legal or commercial complexity exceeds the agent's mandate.
The exception handler takes over when value, uncertainty or risk crosses a defined threshold.
And increasingly, the agent supervisor or workflow designer determines how commercial agents operate, measures their performance and improves their instructions, context, tools and boundaries.
That last role deserves more attention.
If an SDR agent can research and qualify thousands of prospects, somebody still needs to decide what good qualification means. If an agent recommends next actions on an opportunity, somebody needs to understand when those recommendations systematically fail. If a campaign agent can create and optimise journeys, somebody needs to decide what it is allowed to optimise for.
Automation does not remove operating-model design. It makes it more important.
From martech stack to commercial operating system
Underneath this future commercial process sits a different technology architecture.
CRM and other systems of record remain important. So do customer and account data, identity resolution, external intent and company intelligence, knowledge bases, predictive models, decision engines, content systems, marketing automation, CPQ and contract platforms.
But new layers are appearing between them.
Agents need access to reliable context. They need permission to use tools. Multiple agents need orchestration. Their actions need to be observable. Organisations need controls over identity, permissions, escalation and cost.
We can already see this architecture emerging.
6sense, for example, now makes predictive buying stages, account qualification and keyword intent accessible to external agents through MCP. Adobe's Agent Orchestrator is designed to coordinate AI-powered workflows across customer experiences. Gong has introduced an execution layer intended to govern and connect custom agents across the revenue cycle.
The application is no longer necessarily where the intelligence lives.
And the person using the application is no longer necessarily the entity initiating the action.
That is a meaningful architectural shift.
A better checklist for commercial leaders
This also suggests a different way to assess AI maturity.
I would not start by asking:
Do we have an AI SDR? Do we have generative content? Do we have predictive lead scoring?
Instead, take the end-to-end commercial process and assess every important capability against the same questions:
- Business value — what customer or commercial outcome are we trying to improve?
- Signal — what tells us an intervention may be required?
- Decision — what judgement needs to be made?
- Context — what data and knowledge are necessary to make it well?
- Agent role — what can AI reliably analyse, recommend or execute?
- Human role — where do judgement, relationship, expertise or accountability remain important?
- Desired autonomy — should AI assist, recommend, prepare, execute within boundaries or adapt?
- Control — what approval, monitoring and escalation are required?
- Technology — which existing systems and new capabilities enable the workflow?
- Learning — how does the outcome improve the next decision?
That produces a very different roadmap from a catalogue of AI use cases.
It also exposes gaps that a tool inventory misses.
An organisation may own excellent intent technology but have no process for turning a signal into a decision. It may have a powerful content agent but insufficient customer context to make its output relevant. It may deploy an autonomous prospecting agent while sales and marketing still disagree about what constitutes a valuable account.
More AI can then automate the fragmentation rather than solve it.
The end-state is not maximum automation
My current view is that the AI-first B2B commercial organisation will not be defined by how many agents it has.
It will be defined by how effectively it connects signals, context, decisions, actions and learning across the customer lifecycle.
Some of those decisions will belong to traditional software. More will be delegated to agents. Others should remain deliberately human.
The hard part is designing those boundaries as one commercial system.
That is a more useful ambition than maximum automation.
It is also a much better starting point for deciding which technology you actually need.