Advertising, banking, legal work and commerce are starting to point in the same direction: AI is moving from helping us use software to becoming the layer through which we use it.
For most of the AI boom, the dominant product idea has been the copilot.
Open your existing software. Ask AI for help. Get a summary, recommendation, draft or analysis. Then go back into the application and do the work yourself.
That model is not disappearing. But another model is becoming much more visible.
The AI is beginning to sit between the person and the underlying business system.
Instead of Human → App → Action, more workflows are starting to look like Human → Agent → Systems → Action.
The idea itself is not brand new. Salesforce was already arguing in 2025 that agents could become a new user interface, and Microsoft has since described a world where some of software’s biggest users may not be human at all.
What feels different in 2026 is the evidence. Banking, legal work, advertising and commerce are arriving at surprisingly similar architectures at the same time.
McKinsey’s August 2026 State of AI survey adds useful context. Among respondents at organizations with more than $1 billion in annual revenue, 40% say their organizations are scaling AI agents, up from 27% the year before. That does not prove every agent program is creating value. It does show that this is moving beyond isolated demos.
From copilot to participant
A copilot helps you work inside a system.
An agent can begin participating in the workflow itself.
That distinction sounds small until the AI is allowed to read business data, use tools, call APIs, create work, change settings, initiate transactions or coordinate several systems on a person’s behalf.
The interface stops being only a screen full of menus and fields. Natural-language intent becomes part of the interface, while the software underneath becomes an execution layer.
That is the pattern worth watching.
Banking: the assistant becomes an interface into the account
On August 25, 2026, Scalable Capital announced Agentic Investing. The company says it is the first bank in Europe to open its platform to major AI assistants including ChatGPT, Claude and Grok.
Customers can connect an assistant to their Scalable account and work with areas such as investments, watchlists, market information, savings plans, price alerts and trading workflows through natural language.
The important part is not that a bank added a chatbot.
The assistant becomes another interface into the banking platform itself.
The customer can express intent in conversation, while the underlying platform handles the operational work. Scalable still keeps consequential actions behind approval. Trades and savings plans require confirmation before execution.
That detail matters because agentic systems do not necessarily remove the human. They can move the human to a different point in the workflow.
Instead of manually navigating every step, the person defines the intent, reviews the consequence and approves the action.
Legal work: Google is building the governed layer around the model
The same day, Google Cloud launched Gemini Enterprise for Legal in preview.
The interesting part is not that lawyers can use Gemini. Lawyers have been using generative AI for some time.
The important part is what Google has built around the model: purpose-built legal skills, specialist agents, connectors into legal systems, inherited permissions, citations and centralized governance.
Google describes the system as working across areas such as contract review, regulatory scanning, legal research, playbook creation and data-subject access workflows. Through MCP connectors, it can work with systems including document management, ediscovery, collaboration and research platforms while respecting the permissions those systems already enforce.
That is a fundamentally different proposition from pasting a contract into a general chatbot.
A generic assistant returns an answer. A governed agentic platform can participate in the workflow without pretending that the workflow has no rules.
In legal work, those rules are not optional. Matter permissions, confidentiality, source verification and human review are part of the product architecture.
Commerce: the storefront is becoming machine-operable
Commerce is moving in the same direction, although through several different pieces of infrastructure.
In late August 2026, select users of the Apple Store app began seeing an Early Preview of a virtual shopping assistant. Apple has not announced a broad rollout, so this should be treated as a limited test rather than a finished global product.
The assistant can help with product selection, existing orders, trade-in values, carrier offers, payment options and comparisons. Apple’s own privacy documentation also confirms that a Virtual Shopping Assistant exists where available and explains how chat and account information may be used to personalize the experience.
On its own, that could still look like a conversational shopping feature.
The more significant shift appears when you look at the infrastructure being built around agentic commerce.
Google’s Universal Commerce Protocol, developed with companies including Shopify, Etsy, Wayfair, Target and Walmart, is designed to give consumer surfaces, businesses and payment providers a common language for agent-driven shopping. It can work with APIs, Agent2Agent and MCP, and newer UCP capabilities can expose information such as pricing, variants and inventory to agents.
The traditional digital journey often looks like Search → website → category page → product page → checkout.
An agentic journey can compress parts of that into Ask → evaluate → choose → act.
That does not make websites irrelevant. It changes what the website and commerce stack may need to provide. Product data, pricing, availability, policies, identity, checkout and post-purchase actions increasingly need to be understandable not only by people, but by authorized software acting for people.
Advertising is already splitting into two kinds of AI execution
Advertising shows why it is useful to distinguish ordinary AI automation from agentic access.
Google’s AI Max for Search campaigns is an AI-powered optimization layer inside Google Ads. It can expand search-term matching, customize text and choose more relevant landing pages. Google says new Search campaigns now have AI Max selected by default.
That moves more execution and optimization inside the platform toward AI, but AI Max is not the same thing as an outside agent operating Google Ads on a marketer’s behalf.
TikTok’s Agentic Hub is closer to that second model.
Launched in June 2026, Agentic Hub is a marketplace for AI Skills built on TikTok for Business MCP. The skills can support campaign creation and management, creative generation, performance analysis, audience insights, catalog management and other advertising system workflows.
TikTok’s own Help documentation gives examples such as agents rotating creatives when performance declines, adjusting budgets and automating targeting based on campaign data.
This is a meaningful shift.
The old model was: the marketer learns Ads Manager and performs the work.
The emerging model is: the marketer defines what needs to happen, the agent can interact with the advertising system, and the marketer sets objectives, constraints and approval boundaries.
We may be moving beyond the dashboard era
For decades, software has been built around interfaces.
Menus. Tabs. Filters. Forms. Settings. Dashboards.
Learning the interface became part of learning the profession.
A Google Ads specialist knew where every setting lived. An ecommerce operator knew Shopify. A lawyer knew the research and document systems. A financial analyst knew the terminal. An operations team knew the ERP.
That knowledge is still useful. But what happens when an intelligent layer can interact with those systems for us?
The value of knowing exactly where the button is starts to decline.
The value of knowing what should happen starts to increase.
I do not think dashboards disappear. A more plausible outcome is that dashboards become the control plane, while agents become one of the interaction layers.
Humans define goals. Agents execute across systems. Humans inspect, approve, handle exceptions and intervene where judgment matters.
Agents make domain expertise more important, not less
It is tempting to look at all this and conclude that execution expertise becomes less valuable.
Some execution skills probably will become less scarce. But that is not the same as domain expertise becoming less important.
An agent can optimize an advertising campaign. What should it optimize for: revenue, contribution margin, new customers, qualified pipeline or lifetime value?
An agent can review a contract. Which clauses actually matter to this organization? What risk is acceptable? Which exception requires counsel?
An agent can help manage an investment account. What is the investor’s time horizon, liquidity need and risk tolerance?
An agent can recommend a product. Is the customer optimizing for price, durability, compatibility, delivery speed or total cost?
Execution capacity can grow dramatically while the quality of the objective becomes more important.
That shifts human value toward intent, context, economics, constraints and judgment.
The hard part is not the model
Another pattern appears when you compare these systems.
The difficult part is not simply getting access to a capable model. The difficult part is connecting intelligence safely to the systems where work happens.
That is why connectors, APIs, MCP, identity, permissions and orchestration are becoming such important parts of the agent story.
MCP was created as an open standard for connecting AI assistants to external data and tools. By mid-2026, it had evolved well beyond a niche developer experiment, with major vendors using it as an integration layer for business systems.
A useful agentic system needs at least seven things:
- Data: What information can the agent use, and can it trust it?
- Permissions: What is the agent allowed to see?
- Actions: What systems can it operate?
- Context: What does this organization actually care about?
- Guardrails: What must it never do automatically?
- Measurement: How do we know whether the result was good?
- Human escalation: When should the agent stop and ask?
The model is only one component. The business system around the model determines whether the agent is useful, dangerous or simply expensive.
Businesses may need to become agent-ready
We spent years making businesses mobile-friendly.
Then API-friendly.
Then search-friendly.
A new requirement is emerging: agent-ready.
Agent-ready does not mean adding a chatbot to the homepage.
It means asking whether an authorized AI system can understand the business well enough to act safely.
1. Is your core data machine-readable and trustworthy?
Products, prices, inventory, customer records, campaign data, policies and knowledge bases become far more useful when systems can retrieve them reliably. This is the same discipline behind real AI search and discoverability: bad data does not become good because an agent can access it faster.
2. Can systems expose useful actions, not just information?
Reading data is one thing. Creating an order, adjusting a campaign, opening a support case or changing a savings plan is another. Businesses need clean APIs, tools or standardized connectors around the actions they are actually willing to expose.
3. Are permissions granular enough for agents?
A human employee should not automatically have access to every system in a company, and neither should an agent. Identity, role-based permissions and auditability become part of the operating model.
4. Are approval boundaries explicit?
Scalable Capital’s model is instructive. The assistant can help prepare and work through the process, but consequential transactions still require approval. Every company needs its own version of that boundary.
5. Are business objectives defined clearly enough to optimize?
If marketing cannot agree on whether success means revenue, margin, qualified pipeline or new-customer growth, an agent will not solve that ambiguity. It may simply automate it.
6. Can agent actions be observed and audited?
You need to know what the system changed, why it changed it, what data it used and what happened afterward. Observability is not an engineering luxury once software can take consequential actions.
7. Is there an escalation path for uncertainty?
Good automation is not automation that never asks for help. It is automation that knows when confidence is low, risk is high or the situation falls outside the rules.
What this changes for marketing
For marketers, the implications go much further than campaign automation.
Consider a realistic future workflow.
Instead of opening five platforms, pulling reports and comparing them manually, a growth lead could ask:
“Find where we can add another $20,000 in acquisition spend next month without pushing blended CAC above our target. Consider sales quality, inventory availability, recent creative performance and geographic capacity.”
To answer that responsibly, the agent may need to work across Google Ads, Meta, TikTok, analytics, CRM, inventory, finance and creative systems.
That is not an ad-platform feature.
It is an operating layer across the growth system.
And notice what becomes more important as execution becomes easier: clean conversion signals, reliable tracking, business economics, inventory constraints, lead quality, attribution discipline and clear definitions of success.
This fits the same Growth Engine principle STRAVUM applies to marketing more broadly. Acquisition, conversion, measurement and learning create more value when the signal can travel between them.
Agents can make that loop faster. They do not make the loop unnecessary.
What should stay human?
The wrong question is whether a workflow can be automated.
The better question is what level of autonomy is appropriate for the risk.
A low-risk task might be safe to automate end to end.
A medium-risk task might be safe for an agent to prepare and recommend, with a human approving the action.
A high-risk task may require a human to remain the primary decision-maker while the agent provides research, evidence and options.
That spectrum will look different in advertising, banking, legal, healthcare, finance and customer service.
The point is not to maximize autonomy. The point is to place autonomy where it creates more value than risk.
A practical starting point for businesses
The easiest way to waste the agentic AI wave is to start with the question: which agent platform should we buy?
Start with the workflow instead.
Choose one bounded process where the outcome matters, the inputs are reasonably reliable and the action boundary can be defined.
Then map five things:
- What is the objective?
- What data is needed?
- What systems must the agent access?
- What actions may it take without approval?
- How will the result be measured?
That exercise usually reveals whether the real blocker is the model, the data, the integration, the permissions, the process or the fact that the business has never clearly defined success.
In many companies, the model will be the easiest part.
The bigger AI shift may not be another copilot
The developments in advertising initially made me think the marketer’s role was moving upstream.
Less platform operation. More economics, signals, measurement, experimentation and judgment.
I still think that is true.
But the same pattern is now appearing well outside advertising.
Banking. Legal work. Commerce. Enterprise operations.
That makes the thesis broader.
AI is moving from being something inside our applications to becoming a layer between us and our applications.
Apps are not disappearing tomorrow. Visual interfaces will still be better for many tasks. Regulations, security, liability, trust, error handling and user preference will slow the transition in important areas.
But the direction is becoming clearer.
One of the most important questions businesses will face is not simply: which AI model should we use?
It is: which parts of our business should an agent be allowed to understand, operate and optimize, and what must remain a human decision?
That is a much bigger transformation than adding another copilot.
Start with the workflow, not the model
If your business is exploring AI agents, identify one bounded workflow where the objective, data, permissions, approval boundary and measurement of downstream outcomes can be defined clearly. That is a better starting point than adding another AI tool to an already disconnected operating system.
Leave a Reply