Scroll through the release notes for any software your firm uses and you will find the same word appearing with suspicious frequency. CRM platforms. Financial planning tools. Back-office systems. Practice management software. All of them, in the last eighteen months, have shipped something they are calling AI. Some of it is genuinely useful. A lot of it is a search box with a new name.
This is not a cynical reading. It is the predictable result of a market where “AI-powered” has become a sales requirement rather than a product description. Vendors know that buyers are under pressure to show they are adopting AI. So they label existing features as AI, attach them to a premium tier, and wait. Only about 12% of companies with AI tools achieve actual business impact, according to recent industry data. The gap between feature announcements and changed workflows is real, and it is wide.
The question worth asking is not whether a tool uses AI. Almost everything does, in some sense. The question is whether the AI feature changes a real workflow outcome in a way you can measure.
What AI-washing actually looks like
The most common form is rebranded search. A feature that used to say “search your client records” now says “ask your data anything.” The underlying mechanism is the same keyword lookup it always was. The interface has changed; the capability has not.
The second form is surfaced summaries. A dashboard that used to show you a table now shows you a paragraph above the table. The paragraph says something like “your top five clients by revenue are…” and lists what was already visible. This is a generative model processing data you already had access to, to produce text you did not ask for and do not need.
The third form is the assistant that only works in demos. You have seen this: a vendor walkthrough where the AI produces a clean, accurate output from a well-structured prompt. Then you try it on your actual data and it hallucينates a client’s pension value, misreads a date range, or simply declines to answer because the context window ran out. The capability is real in controlled conditions. It is not real in yours.
The honest test is not “does this feature exist?” It is “has anyone in my firm changed how they spend their time because of it?”
A plain-English checklist for evaluating AI features
Before you upgrade a licence tier for an AI feature, or before you dismiss one as worthless, run through these questions.
Does it replace a step I currently do manually, or does it add a new step? Useful AI removes friction. AI-washing adds a new interface you have to manage on top of the old one. If using the AI feature requires you to export data, paste it somewhere, review its output, and then re-enter it into the original system, that is not automation. That is more work.
Can I describe the before and after in concrete terms? “Before: I spent forty minutes each Monday pulling together the weekly client review list from three systems. After: that happens automatically and takes four minutes to check.” That is a real workflow change. “Before: I searched for things. After: I also search for things but using natural language” is not.
Is the AI operating on my actual data, or on a general model that doesn’t know my firm? This distinction matters a great deal in regulated environments. A general-purpose assistant that can answer questions about pensions legislation is different from a system that can answer questions about your specific client’s pension, drawing on the actual file your firm holds. The latter requires grounded AI, where the model’s responses are constrained to a specific knowledge base. Many vendor “AI features” are the former dressed up as the latter.
Who owns the liability if the output is wrong? This is the question most platform vendors would prefer you did not ask. If the AI suggests a course of action that turns out to be incorrect and a client is harmed, the answer is almost certainly not the vendor. The FCA and Bank of England issued a joint statement earlier this month making clear that AI governance for UK-regulated firms is moving from voluntary best practice toward formal expectation. “The platform said so” will not be a sufficient answer.
Is this feature solving a problem I actually have? Vendors design AI features around what is technically feasible to build, not around what your firm’s operations actually need. A financial planning platform might ship an AI that drafts client emails. If your bottleneck is actually the suitability report sign-off process, that feature is irrelevant no matter how well it works.
How to apply the L1/L2/L3 lens here
When a vendor AI feature passes the checklist above, it is essentially functioning as a Level 1 (education) or Level 2 (integration) solution that someone else has built for you. That is not a criticism. If a tool you already pay for genuinely removes manual work from a real process, using it is the right call. It costs nothing extra, and you should start today.
The problem comes when a vendor AI feature fails the checklist and you are still being charged for it. Or when the feature almost solves the problem, but not quite, and you are now considering a separate Level 2 (integration) project to paper over the gaps in a tool that was supposed to make your life simpler.
The more expensive mistake is assuming that because your software vendor has shipped AI, your firm’s AI requirements are met. They are not the same thing. What your platform can do and what your specific workflows need are separate questions, and conflating them is how firms end up paying for three AI features across different tools, none of which connect, and none of which have changed how anyone spends a Tuesday afternoon.
What good looks like
A useful AI feature has an owner. Someone in your firm can name the feature, describe what it does, and tell you when it last saved them time. If nobody can do that, the feature is not in use, regardless of what the vendor’s renewal pitch says.
A useful AI feature has a measurable before and after. Four hours saved per week. Twelve fewer manual exports per month. One step removed from a process that used to take six. If you cannot name a number, the feature is either too new to evaluate or not actually being used.
A useful AI feature fits the actual architecture of your work, not a generic version of it. That means it operates on your data, in your systems, in the way your firm actually processes information. The closer a vendor’s AI feature gets to that specificity, the more seriously it is worth taking.
The rest is marketing. And there is a lot of it around at the moment.
If you are trying to work out which AI features in your current stack are worth keeping and which are licence-tier padding, a discovery call is a straightforward way to start.