AI on top of broken systems is the fastest way to chaos

At some point, somebody in the boardroom will ask a simple question: "What is AI actually doing for us?"

Not how many licences we have bought. Not how many emails a new agent can send before lunch. What changed for the customer, for the team and for the bottom line? If the answer is a dashboard full of activity and a shrug about revenue, the problem probably predates the AI.

AI on top of broken systems is the fastest way to chaos. More chaos, faster, at scale, while you pay six subscription fees for the privilege.

That is not an argument against AI. It is an argument against asking a tool to accelerate a journey nobody has bothered to inspect.

So what is AI actually doing for us? For most firms, the honest answer is that nobody knows, because nobody measured the journey before the tools were switched on. This piece is about what goes wrong without it, and how to find where your system breaks before you pay to accelerate it.

The spending is real. The proof is thin.

Comviva's 2026 Global CMO Survey found that 90% of the organisations it surveyed had increased AI marketing investment over the previous two years, but only 12% could prove that it worked. This was a survey of more than 200 senior IT and business executives in retail, ecommerce and telecoms worldwide, not a census of CMOs everywhere. That distinction matters. The distance between investment and evidence still deserves a boardroom question.

Gartner's 2026 CMO Spend Survey, conducted January to March among 401 CMOs and other marketing leaders in North America, the UK and Europe, found an average 15.3% of marketing budgets allocated to AI initiatives. Only 30% reported mature or fully developed AI readiness capabilities. These are two different surveys with different samples. Put them side by side as warning lights, not as a single equation for return on investment.

What I hear beneath both is familiar: "It's not the AI that's failing. You didn't measure anything before, so you don't know where the improvements are." If a lead disappears between a form and a salesperson, a better writing model cannot tell you whether the message, the hand-off or the follow-up was responsible. It may simply make the bad hand-off happen more often.

Ford learned what the system didn't know

Ford has just made the point from the factory floor. In June 2026, its vehicle hardware VP said the company had mistakenly thought that introducing AI and changing design requirements would produce a high-quality product. The knowledge of experienced engineers had not all made it into the automated systems; quality work was too fragmented. Ford says it hired, promoted or brought back more than 350 experienced engineers to rebuild that expertise and improve the data going into its tools. Different industry, same order of operations: find what the system does not know before asking it to move faster. Source.

That is the failure mode I would look for in marketing, too. A customer status, offer rule or hand-off that only one person understands may work while that person catches mistakes. Once a tool acts on that partial picture at scale, the gap becomes a customer-facing decision. The problem is not simply that the process is digital. It is that nobody checked what the system knew or what it was allowed to do with it.

Before an outbound agent sends a single message, I would want to see the suppression list, customer status, offer rules, contract terms, approval gate and sending-domain plan. Then I would send a small, reviewed batch and check the results against actual conversations and contracts. Volume is not proof of progress.

The day an SEO win broke the business

I learned a version of this lesson long before today's AI tools. An early SEO client of mine was a business insurance firm. For months, the owner saw no meaningful change in traffic. Then three of our target keywords reached positions one to three. The phone went from a few calls a day to hundreds. The server broke. Email went down. Three salespeople could not answer the calls quickly enough.

The owner rang me and asked me to "switch the SEO off". I couldn't. It wasn't paid advertising with a pause button. We had won the visibility the business had asked for, but the operation behind it was not ready to receive the demand.

That is the uncomfortable part of a successful campaign. If your website, inbox and sales capacity cannot take the result, more demand is not automatically good news. The same applies to AI. You can automate acquisition, qualification and follow-up, but every extra hand-off puts more weight on the parts of the business underneath. If those parts are weak, faster is not better.

The market is changing while you argue about the tool

There is also a real reason CMOs are in a hurry. The routes by which people find and visit a business are shifting, and the measurements are not interchangeable.

Ahrefs' February 2026 update, based on 300,000 keywords and Google Search Console desktop data for December 2025, found that the presence of a Google AI Overview correlated with a 58% lower average click-through rate for the top-ranking page against its comparison group. That is a comparison, not a randomized estimate of what happened to every site. A separaterandomized field study, run among 1,065 US desktop Chrome participants in January and February 2026, reported 38% fewer outbound organic clicks on queries where an Overview appeared. Different design, denominator and question; same reason to look past rank alone.

For publishers, Chartbeat data reported by Search Engine Journal in September 2026 showed Google Search referral pageviews across its publisher network down 40.2% year on year from July 2025 to July 2026. Its client base skews towards news and media. That is not a forecast for every B2B website. BrightEdge's September 2026 dataset showed ChatGPT referrals to the brands it tracks up 101% from January to August, reaching 95.1% of AI-generated referral traffic in August. The share is of AI referrals in that dataset, not of all website traffic.

A further figure needs even more care. In a publishers' court filing reported by PPC Land on 21 September 2026, Microsoft's data was described as showing click-through rates to the New York Times's sites 87-93% lower from Bing Chat than from conventional Bing web search. That is a comparison for particular publisher properties, presented in litigation, not an 87-93% loss of traffic for the whole web.

These figures do not add up to one universal AI-search number. They tell me that visibility, clicks, enquiries and revenue need to be measured separately. A page can be found without being visited. An AI referral can grow from a small base. A lead can arrive without becoming a sale. If you only measure the first thing that looks good, you can be very confident about the wrong outcome.

Measure the journey before you automate it

My starting point is a Diagnostic, not another tool recommendation. Take one real lead and follow it all the way through. Where did it come in? Which page, search, referral or campaign brought the person to you? Where did the enquiry go next? What did the CRM record, which person owned the reply, which tool or AI system touched it, and what happened when a quote or contract went out?

Then do the less glamorous work: compare timestamps and records, read the messages, speak to the people doing the job. Find where leads wait, where fields lie, where a customer is treated as a prospect and where two systems disagree about the same transaction. Establish a baseline: response time, qualified enquiries, conversion to sale, value and the true cost of the people and subscriptions involved. If the sample is too small or the attribution too messy to claim an improvement, say so.

Only then ask what to automate. Sometimes it is a repetitive task that AI can genuinely do faster. Sometimes the first fix is a field in the CRM, a routing rule, enough capacity to answer the phone or an offer that the sales team can actually deliver. The best case is finding out before the new tool renews.

Every other agency may be comfortable with vagueness; the subscriptions still get paid. I don't recommend anything before I've measured. That is not caution for its own sake. It is how you distinguish an investment from a demonstration.

You walk into the board confident what's working and why - or you hand the diagnostic to the agency that sold you the AI in the first place.