The problem with AI-generated leads is how convincing they look
Why finding companies is easy, finding genuine commercial opportunities is difficult, and the difference matters more than most AI-generated prospect lists suggest.
Ask ChatGPT or Claude to find 20 potential customers for your business.
Give it a description of your services, your ideal customer profile and perhaps a geographical market.
Within minutes, you may receive an impressive list.
Companies. Industries. Decision-makers. Recent developments. Potential needs. Reasons to make contact. Sometimes even a suggested opening message for each prospect.
It looks remarkably like market intelligence.
And that is precisely what makes it dangerous.
The AI may be much better at constructing a convincing explanation for why a company could need your services than at establishing whether that company actually needs them.
I have been thinking about this distinction for some time. Recent work on identifying commercial signals has made the problem much more concrete.
The difficulty is not generating plausible leads. We can do that almost effortlessly.
The difficulty is establishing which of those leads represent real, timely and commercially relevant opportunities.
Those are two very different tasks.
A good story is not evidence of a buying opportunity
Imagine asking an AI assistant to identify companies that could need legal advice.
It finds a company expanding internationally.
The reasoning sounds sensible:
International expansion creates regulatory complexity. Regulatory complexity may require legal advice. The company could therefore be an attractive prospect for a law firm with relevant expertise.
Nothing about that reasoning is inherently wrong.
But consider what we still do not know.
- Is the company actually expanding, or is the information outdated?
- Which legal entities are involved?
- What legal work, if any, does the expansion require?
- Has that work already been completed?
- Does the company use internal lawyers?
- Does it have an established external adviser?
- Who would influence or make the decision to appoint someone new?
- Why would this particular law firm have a better chance of winning the assignment than its competitors?
The AI has identified a plausible connection between an event and a professional service.
It has not necessarily identified a buying opportunity.
A plausible need is not a verified need. A verified need is not a buying decision. And a buying decision is not the same as having a Right to Win.
That distinction gets lost surprisingly easily when the output is presented as a professional-looking prospect report.
What happens when we try to verify the evidence?
In our work on commercial signal detection, we have been testing what it actually takes to identify relevant developments systematically.
One early, deliberately restricted test produced an interesting result:
966 website pages discovered. Three pages actually read. Zero qualifying signals.
At first glance, that looks like a disappointing result.
I think it is an important one.
The system did not pretend to have examined 966 pages. It distinguished between pages it could identify and pages it could actually access, read and evaluate.
It did not invent signals to make the exercise look successful.
That is what responsible behaviour looks like when the available evidence is insufficient.
The test was small, and the restrictions were substantial. It tells us nothing reliable about the overall frequency of commercial opportunities in the market.
But it demonstrates a problem that is easy to overlook when working with general-purpose AI assistants.
Discovering the existence of information is not the same as having access to its contents.
And having access to the contents is not the same as understanding their commercial significance.
Seven things must work before an event becomes a useful signal
Identifying a reliable commercial signal requires several conditions to hold at the same time.
First, the source must be accessible under the applicable technical, contractual and legal conditions.
Second, its relevant content must actually be retrieved.
Third, the content must be interpreted correctly.
Fourth, the event must be connected to the correct company or legal entity.
Fifth, any collection and processing of information must comply with applicable privacy requirements.
Sixth, the event must be sufficiently recent to have commercial relevance.
Seventh, there must be a defensible connection between the event, a potential business need and the services being offered.
Each step introduces uncertainty.
An error early in the process can invalidate everything that follows.
And none of this establishes, by itself, that the company intends to buy anything.
This is why a system that identifies commercial opportunities cannot simply be a language model producing persuasive summaries.
It needs a reliable process for collecting evidence, validating it and preserving the distinction between what is observed and what is inferred.
The information we need is often the information we cannot see
Some commercial developments are relatively easy to observe.
A public procurement has been announced.
A company has entered bankruptcy proceedings.
A formal corporate announcement has been published.
Other developments are much harder.
A management team may be considering restructuring long before it becomes public.
An acquisition may be discussed privately for months.
A company may be dissatisfied with its current adviser without publishing anything that reveals the problem.
A new strategic direction may create professional-service needs that are not immediately visible in any public source.
Even publicly available information is not necessarily available for unrestricted automated collection and reuse.
Some publishers block automated access. Others require subscriptions or licensing agreements. Many websites depend on JavaScript, employ security challenges or expose information in fragmented formats.
Public-sector information can be spread across different systems with different access conditions.
This matters because the commercial opportunities we are most interested in may emerge well before a company announces that it needs external assistance.
By the time a purchasing decision becomes publicly visible, the buyer may already have selected a supplier.
The earlier we want to identify an opportunity, the less direct evidence of buying intent we are likely to have.
That is not an argument against early signal detection.
It is an argument for being much more disciplined about what we claim to know.
Why AI makes weak evidence look stronger
Language models are exceptionally useful at identifying relationships, interpreting text and constructing explanations.
But those same capabilities can create a problem in commercial research.
Suppose we ask for 20 qualified prospects.
The requested output is a list of 20 prospects.
Unless the task has strict evidence requirements, explicit rejection criteria and a genuine willingness to return fewer results, the assistant may produce a list of companies supported by varying levels of evidence.
Some may be excellent.
Others may have little more than a plausible industry profile and an interesting recent announcement.
Yet the explanations can look equally polished.
A weak lead with an articulate explanation may appear more valuable than a strong lead supported by a short, factual observation.
The presentation can disguise the difference in evidence quality.
This is not necessarily a traditional AI hallucination. The company may exist. The announcement may be genuine. The suggested business need may be perfectly reasonable.
The problem is that the conclusion can be more confident than the evidence permits.
A carefully worded hypothesis can begin to look like a confirmed opportunity.
And once it enters a CRM system, a sales pipeline or a management report, that distinction may disappear entirely.
The missing question: Why should this company choose us?
Even when a genuine need exists, we are still missing the most important commercial question.
Why should the buyer choose us?
A company requiring legal advice may already have excellent advisers.
It may have internal capabilities.
It may prefer a competitor because of an existing relationship, specialist reputation, geographical presence, price or previous experience.
The existence of demand does not establish our ability to win it.
This is where Right to Win becomes essential.
I think useful commercial intelligence must examine at least three different questions:
- Is something happening? What can we document about the company and its situation?
- Could it create a relevant need? What is the commercial hypothesis, and what evidence supports or weakens it?
- Why could we win? What gives this particular supplier a credible advantage, and how might it reach the people involved?
The third question is frequently absent from automated prospect lists.
A company can be a perfect match for our services and still be a very poor sales opportunity.
Conversely, a less obvious prospect may be highly attractive because we have a unique capability, a relevant relationship or a particularly strong understanding of its situation.
That cannot be established through industry matching alone.
The answer is not to stop using AI
I do not think the lesson is that ChatGPT or Claude should not be used for prospecting.
Quite the opposite.
They can be extraordinarily useful for exploring markets, interpreting developments, generating hypotheses, examining business situations and preparing more relevant conversations.
But I increasingly think we should stop asking them to deliver a fixed number of supposedly qualified leads as if qualification were simply a research output.
A more useful assignment is:
Find companies with documented developments that could create a need for our services. Preserve the original sources and dates. Explain what is observed, what is inferred and what remains unknown. Identify the evidence that would strengthen or weaken the hypothesis. Assess our potential Right to Win. Reject opportunities that do not meet the evidence requirements.
And accept that the answer might be three companies rather than 20.
Or none.
That is not an unsuccessful search.
It may be the most commercially valuable result the system could produce.
What changed
Our recent work on commercial signal detection has made the limitations of automated lead identification much more visible.
We have been mapping potential business needs to observable developments, examining source accessibility and distinguishing between structured public data and information that requires interpretation of company websites and other text sources.
The scale of the challenge is significant.
In one working catalogue, we have 5,842 potential signal relationships. Our current analysis suggests that only around 1% can be covered through the structured feeds considered so far, while the large majority require access to less structured information.
Those are estimates within our current source architecture, not universal measures of market coverage.
The operational work has also exposed practical problems with company identification, source permissions, machine readability, outdated information and privacy safeguards.
Most importantly, we have started treating an event, a potential need and a credible Right to Win as separate things.
We have not established that our current system can consistently identify confirmed, winnable opportunities. That remains work to be done.
But we have become much better at describing exactly what the system has observed, what it has not observed and what it is permitted to conclude.
That is meaningful progress.
What I think matters
I believe one of the biggest risks of AI-assisted business development is that companies will confuse the ability to generate impressive prospecting material with the ability to generate commercial opportunities.
Those are not equivalent.
A list of 100 convincing prospects can be less valuable than five companies supported by documented developments, relevant business hypotheses and a credible reason to initiate contact.
The consequences are practical.
Salespeople spend time approaching companies with no identifiable need.
Marketing teams build campaigns around assumptions that have not been validated.
Management teams see activity levels and pipeline values that suggest commercial progress, even when the underlying opportunities are weak.
The organization becomes more efficient at producing the appearance of business development.
Not necessarily at winning business.
AI reduces the cost of producing commercial hypotheses. It does not remove the cost of validating them.
That is the distinction I want to preserve.
For me, the ambition is not to build a system that produces the largest number of leads.
It is to build one that helps people identify developments worth investigating, understand why those developments might matter and decide where they have a realistic opportunity to win.
That also changes how success should be measured.
The number of leads produced tells us very little on its own.
I would rather measure how many hypotheses survive verification, how many lead to relevant conversations, how many develop into genuine buying processes and how many ultimately produce profitable assignments.
A credible commercial system should be capable of learning from every stage, including why apparently attractive opportunities never materialized.
What remains uncertain
We still have substantial questions to answer.
We do not yet know how much relevant commercial activity can be observed reliably across the available sources.
We do not know which combinations of signals are the best predictors of an emerging need.
We have not proved that detecting earlier developments results in a higher probability of winning business.
Nor have we established how much human verification can be reduced without compromising evidence quality, privacy or commercial judgment.
There is also a more fundamental limitation.
Some of the most important information about a buying decision may never be publicly available.
Relationships, internal priorities, supplier preferences, budgets and political considerations inside an organization may be invisible until a conversation takes place.
No amount of automated research can legitimately claim to know what it has not observed.
And the comparison with ChatGPT and Claude deserves a qualification: their results depend on the tools, sources, instructions and verification processes they are given. A general chat task is not the same as a rigorously designed research workflow using those models.
We still need controlled comparisons to quantify the difference.
But I am increasingly convinced that the most useful commercial AI systems will not be those that always provide an answer.
They will be those that can explain why an opportunity might exist, show the evidence behind that conclusion and acknowledge what remains unknown.
The real achievement is not making a lead look convincing. It is making the reasoning behind it possible to verify — and being willing to conclude that the lead is not good enough.
That is where I think genuinely useful AI-assisted business development begins.