Category: Marketing Technology

I recently came back from a trip to Europe and realized that I had travelled very differently this time.

I used an LLM for almost everything.

Trip planning. Getting around unfamiliar cities. Public transit. Finding things to see. Understanding signs and menus. Dealing with the language barrier. Figuring out what was open, where something was located, and how to get there.

Things that once required jumping between Google, transit websites, travel blogs, maps and translation apps increasingly happened through one conversation.

It worked remarkably well.

But the occasional mistakes were actually more interesting to me than the successes.

In several cases, the problem wasn’t necessarily that the AI couldn’t understand my question. It was that the information it was working with was messy, incomplete, ambiguous or difficult to extract from a conventional web page.

And that got me thinking about something much bigger than travel.

We built the web for people

For decades, the basic contract of the web has been fairly straightforward.

A business publishes information in HTML. A browser turns that HTML into something a person can read and interact with.

We have become very good at this.

Modern websites can be beautiful. We have navigation systems, responsive layouts, accordions, sliders, pop-ups, interactive maps, JavaScript applications and sophisticated visual hierarchies.

But increasingly, the person visiting your website may not be the one actually reading it.

A machine may read it first.

Someone might ask an AI assistant:

Does this hotel have parking?

What is the cancellation policy?

Is this business open on Sunday?

Which of these products is appropriate for a small business?

Can I get from the airport to this location by public transit?

The AI then has to find, interpret and assemble the answer.

That changes the problem.

HTML isn’t going away. But HTML alone may not be enough.

HTML is an extraordinary technology. It has survived decades of enormous change to the web.

But the meaning that seems obvious to a person looking at a rendered page isn’t necessarily equally obvious to software.

This isn’t a completely new problem.

We’ve been adding machine-readable layers to websites for years. XML sitemaps help search engines discover URLs. Structured data helps machines understand that something is a product, event, organization, review or offer.

Schema.org, for example, has spent years developing a shared vocabulary for describing information on the web in machine-understandable ways.

What is changing is the importance of the machine audience.

For traditional SEO, machine readability could influence how a page was indexed or appeared in search.

In an AI-mediated environment, it can influence the answer itself.

That’s a much higher stake.

We may need two versions of the web

I increasingly think businesses should start viewing their content as having at least two presentations.

The first is for people.

It should be attractive, persuasive, accessible and easy to navigate.

The second is for machines.

It should be structured, explicit, concise and extremely difficult to misunderstand.

We’re already seeing experimentation in this direction.

The emerging llms.txt proposal, for example, suggests giving AI systems a concise, LLM-friendly guide to important website content and potentially linking to cleaner Markdown versions of pages.

Whether llms.txt itself becomes a lasting standard is almost beside the point.

The important idea is this:

A website may no longer be just a collection of pages. It may also need to become a reliable source of structured knowledge about the organization behind it.

Think about content as data

This leads to a different way of thinking about websites.

Consider a hotel.

Its website contains room descriptions, rates, check-in times, amenities, parking information, cancellation rules, restaurant hours, addresses and contact information.

Traditionally, we think about how to arrange all of that into pages that are useful and persuasive to visitors.

But underneath those pages is something else:

information.

The webpage is simply one presentation of it.

The same underlying information could potentially be presented as HTML for visitors, structured data for machines, feeds for other systems, clean text for AI tools or through an API.

That changes the model from:

page → information

to:

information → multiple representations

I think this distinction is going to become increasingly important.

Your CMS should eventually do this automatically

I don’t think marketing teams should have to maintain five separate versions of every piece of information.

Eventually, a good content management system should allow an organization to maintain a reliable source of truth and automatically produce whatever representations are required.

HTML for people.

Structured data for machines.

Clean text or Markdown where appropriate.

APIs and feeds where appropriate.

Perhaps formats we haven’t invented yet.

We’re not completely there.

So for many organizations, some of this will remain manual work for a while.

Fortunately, we’ve been through something similar before.

SEO forced organizations to think beyond what a webpage looked like. We learned to care about page titles, metadata, XML sitemaps, canonical URLs, structured data and information architecture.

The AI transition isn’t starting from zero.

But the implications could be considerably broader.

What businesses can do now

I wouldn’t recommend rushing out and creating an “AI version” of every page on your website.

Start with the information itself.

Identify the questions people most frequently ask about your organization.

Make the answers easy to find.

Make your terminology consistent.

Don’t bury important facts inside clever marketing copy.

Use appropriate structured data.

Separate core information from its visual presentation wherever your technology allows it.

And start asking a new question whenever you publish important content:

Could a machine understand this correctly without seeing the page the way I do?

If the answer is no, there is probably work to do.

The web isn’t becoming less important

There has been plenty of discussion about AI reducing search traffic and potentially making websites less relevant.

I wonder if we’re looking at it backwards.

Websites may become even more important.

But their role may change.

Your website won’t only be a destination people visit.

It will increasingly be an authoritative source of information that search engines, assistants and AI agents consult when they need to understand your organization.

For the last 25 years, we’ve spent enormous effort making websites easier for people to use.

The next challenge may be making sure machines understand them just as well.

And I suspect that will become a surprisingly important part of digital marketing.

AI can help a marketing team move faster. Speed is useful only when the team knows which work should accelerate and where a pause protects customers, brand and accuracy.

The boundary should be based on risk, not enthusiasm.

Good uses are bounded

Summarizing supplied material, producing variations, classifying records, checking consistency and drafting from approved facts can save time. The input is known, the output can be reviewed and the consequence of an error is manageable.

These jobs still need owners. “The model did it” is not a control.

Keep judgment at sensitive points

People should own claims about products, decisions affecting access or opportunity, responses to vulnerable customers, use of confidential data and public statements during uncertainty. Context and accountability matter more than fluency.

Brand voice also needs editing. A plausible paragraph can be wrong, bland or out of character.

Build verification into the workflow

The NIST Generative AI Profile treats risk across the AI lifecycle rather than as a final review. Marketing teams can apply the same logic: define approved inputs, restricted data, review depth, testing, records and escalation before a tool is widely used.

Higher-risk work deserves stronger evidence and human approval.

Measure quality, not output

Counting generated assets rewards volume. Track correction time, factual errors, rejected work, customer response and whether the team learned anything. Sometimes the fastest draft creates the slowest approval.

AI should remove avoidable labour and widen useful options. It should not become a reason to publish work nobody is willing to own.

Sources

• NIST: AI Risk Management Framework
• NIST: Generative AI Profile

A CRM can make an organization feel more organized before it has decided what it is organizing for. Fields are configured, stages are named and dashboards appear. The difficult questions remain untouched.

Who is the priority customer? What problem do we solve well? Why should that person continue the relationship? Software cannot choose those answers.

A system records a model of the customer

Every CRM contains assumptions: what counts as a lead, which interactions matter, when a relationship changes state and who owns the next action. If those assumptions are vague, the database becomes a tidy record of inconsistent behaviour.

Configuration should follow a lifecycle the organization can explain in plain language.

Segmentation requires a reason

Creating more segments is easy. Useful segmentation changes an offer, message, service level or next action. If two groups receive the same treatment, the distinction may not deserve operational complexity.

Collect data because a decision needs it, not because a field exists.

Strategy sits outside the platform

Salesforce’s own CRM strategy guidance spans business processes, customer needs and analytics. That is broader than software setup. The platform supports choices about relationships; it does not originate them.

The same is true of automation. A journey builder can execute a sequence only after somebody decides what a respectful, useful sequence should be.

Judge the CRM by behaviour

Do teams trust the data? Are handoffs clearer? Do customers receive more relevant responses? Can leaders see where relationships stall and change the process?

A system that answers those questions is valuable. A system celebrated mainly for the number of records it holds is an expensive address book.

Sources

• Salesforce: Creating a CRM strategy
• HubSpot: Workflow Automation Guide

The wrong automation question is, “What can the platform do?” The answer is usually far more than the team should deploy.

A better question is, “Which work is stable enough to encode, costly enough to repeat and safe enough to run without someone watching every step?”

Automate repetition, not uncertainty

Strong candidates have a clear trigger, predictable inputs, explicit rules and an output that can be checked. Form confirmations, task creation, routine routing, data normalization and stale-record alerts often qualify.

Weak candidates depend on context nobody has documented. A workflow that decides how to respond to a sensitive complaint or which complex account deserves attention may automate the appearance of judgment without the substance.

Start where errors are recoverable

Automation increases speed in both directions. A good rule runs quickly; a bad one can contact thousands of people before lunch. Begin with bounded volume, logs and a way to reverse or contain mistakes.

HubSpot’s workflow guidance emphasizes reviewing enrollment, timing and placeholders before publication, then watching history and errors. That operational discipline matters more than the number of branches.

Keep people at consequential boundaries

Human review belongs where brand trust, fairness, privacy or material customer impact is at stake. AI can summarize a record or suggest a draft. A person should own the decision to make a sensitive claim, deny a request or act on uncertain data.

That line will move as evidence improves. It should move deliberately, not because a vendor released a button.

Measure the whole cost

Time saved is only one side. Track exceptions, correction work, customer confusion and maintenance. An automation that saves five minutes per case but creates a weekly forensic exercise may be a loss.

Good automation is often boring. It removes small, reliable burdens and gives people more attention for work that needs context. If a process is unstable, fix it first. Encoding chaos only makes it arrive on time.

Sources

• HubSpot: Workflow Automation Guide
• Salesforce: Marketing Automation Guide

A new marketing tool often enters through a small door. One team needs a feature, the monthly price fits an operating budget, and implementation appears to require little more than adding a script and importing a list.

The invoice is real. It is rarely the full cost.

Every tool creates relationships

The platform needs identities, permissions, data inputs, outputs, consent rules and reporting definitions. Someone must decide which system is authoritative when records disagree. Someone else will answer support questions and remove access when staff leave.

These relationships remain even when the original use case ends. A cheap tool can create expensive dependencies.

Integration is not a one-time task

APIs change. Fields are renamed. Authentication expires. Business processes evolve. Each connection becomes a small product that needs an owner and a test. Multiply that by a crowded stack and teams spend more time keeping data moving than learning from it.

The hidden cost appears during change: replacing a CRM, revising consent, merging teams or answering a privacy request.

Fragmentation weakens measurement

When platforms calculate similar metrics differently, reporting becomes reconciliation. Teams debate whose dashboard is correct while customers move through journeys no single system can see.

Adding another analytics view can make the organization less informed if it introduces a new vocabulary without resolving the old one.

Use an entry test

Before approving a tool, name the capability it adds, the process it replaces, the data it touches, the system of record, the owner, the exit plan and the measure of success. Include implementation and retirement effort in the business case.

Then ask whether an existing platform already provides enough of the capability. “Enough” matters. The best standalone feature may not justify another vendor relationship.

A stack should not be judged by how many modern logos appear on its diagram. Judge it by whether people can operate it, change it and trust what it produces.

Sources

• Salesforce: Data, automation and transformation resources
• HubSpot: Workflow Automation Guide

Software can be configured in weeks. The arguments it exposes can last for years.

Who owns customer data? Which team approves a new field? Who can publish, export, integrate or delete? What happens when local speed conflicts with enterprise standards? These are governance questions, and no platform setting can answer them.

The hard work is social

Technology projects often frame governance as documentation produced near launch. Real governance is a working agreement about authority. It tells people who decides, who contributes, which standards apply and how exceptions are handled.

Without that agreement, teams route around the system. Duplicate tools appear, fields drift, integrations become brittle and reporting loses credibility.

Ownership must be specific

“Marketing owns it” is not specific. A named role should own platform health, another may own data definitions, and business teams may own processes that run through the technology. Shared responsibility needs explicit decision rights or it becomes nobody’s responsibility.

Governance also needs a route for change. A standard that cannot adapt will be ignored.

Standards reduce invisible costs

Naming conventions, lifecycle definitions, consent rules, access reviews and integration patterns can feel bureaucratic when considered one at a time. Their value appears later, when a team can change a campaign without breaking reporting or investigate an issue without interviewing five former employees.

Salesforce’s resource model spans data, automation and digital transformation. That breadth reflects the reality that platforms cross departmental boundaries. Governance must cross them too.

Start with decisions, not committees

A governance group should exist to make a bounded set of decisions. Publish those decisions in plain language. Record why they were made. Review them when the operating context changes.

The measure of governance is not the number of meetings or pages in a policy. It is whether people can make a routine change safely, know when approval is required and trust the data that comes out. Technology is visible. Governance is what keeps it useful after the launch team leaves.

Sources

• Salesforce: Business resources by topic
• Nielsen Norman Group: Ownership and oversight

Marketing automation demos are built on obedient processes. A person submits a form, the right data appears, the correct owner is known and a sensible message goes out. Real organizations are messier.

When automation disappoints, the software often gets blamed for faithfully executing rules that were never settled.

Automation is an unforgiving mirror

A manual process can survive ambiguity because people improvise. Someone recognizes a familiar name, fixes a field or messages a colleague. Automation cannot rely on that invisible repair work. It needs explicit triggers, states, owners and exceptions.

That is why implementation uncovers arguments about lead definitions, consent, territories, service levels and data authority. The platform did not create those problems. It made them harder to ignore.

Map the work before building the workflow

Start with a plain description of the customer event and the response it should cause. Identify required data, the system of record, the responsible person, timing and the conditions that stop the sequence. Then document exceptions. The awkward cases matter more than the happy path.

HubSpot describes workflows through triggers, actions and records. That technical model is useful, but each element needs a business decision behind it. “Deal stage changed” only works if people use stages consistently.

Good automation begins small

The best first candidates are frequent, stable and easy to verify: routing a completed enquiry, confirming receipt, creating a task or flagging a stale record. These jobs save time without pretending judgment has disappeared.

Complex nurturing journeys are tempting because they look sophisticated. They also multiply branches, content dependencies and failure points. Build them after the underlying lifecycle is understood.

Manage the operating system

Every live workflow needs an owner, a change log, a test method and a review date. Watch failures and unintended enrolments, not only completion counts. Retire automations that no longer serve a current process.

A new platform may eventually be necessary. But if the team cannot draw the process, name the owner and explain the exception rules, procurement is premature. The fastest route to better automation may be a whiteboard and an uncomfortable meeting.

Sources

• HubSpot: Workflow Automation Guide
• Salesforce: What is Marketing Automation?