Most businesses booking a first conversation about AI expect to be told what to buy. What they usually get instead is a list of things that are already broken, none of which need a model to fix.
That is not a criticism of the businesses. It is what an audit is for. You cannot sensibly choose between a voice agent, a CRM rebuild and a document pipeline until you know which one of those is sitting on top of the largest hole. And after enough of these, the holes start to look the same.
Here is what turns up, in roughly the order it turns up.
1. Calls nobody answers, and nobody counts
The first thing worth measuring is not a conversion rate. It is how many people tried to reach you and gave up.
Almost no small business knows this number. The phone system knows it (most VoIP providers will happily show missed calls by hour of day) but nobody has ever pulled the report. When we do pull it, the pattern is consistent: a cluster before 9am, a cluster at lunch, and a long tail from 5pm onwards that is often 20–30% of the day's inbound.
Those calls are not neutral. A caller who reaches voicemail at 5:40pm on a Tuesday is not waiting patiently until Wednesday. In trades and home services in particular, they are working down a list of three numbers from a search result. Whoever picks up wins.
What the audit produces: a count, by hour, of unanswered inbound over a representative month, and an estimate of what an average job from that channel is worth. That single pair of numbers decides whether voice belongs on the roadmap at all (and if it does, whether it belongs in quarter one or quarter three.
It also, quite often, kills the idea. A firm doing eleven inbound calls a week does not need a voice agent. It needs someone to answer the phone.
2. Follow-up that depends on who happens to be free
The second leak is slower and harder to see, because nothing visibly fails.
An enquiry arrives) web form, WhatsApp, an email to the general inbox. Someone sees it. Someone means to respond. In a good week they respond in twenty minutes. In a bad week, during a big job or a school holiday, it is Thursday.
Nobody logs the bad weeks. The business remembers its good response times and quotes those in meetings. The audit's job is to go and look: pull three months of enquiries, match each one to the first outbound reply, and plot the distribution rather than the average.
The distribution is always worse than the average. It is nearly always bimodal (a big cluster under an hour, and a stubborn tail measured in days. That tail is the money.
What the audit produces: actual time-to-first-response, spread rather than averaged, and a note on what triggers the tail. Sometimes it is volume. More often it is a specific person being unavailable, which is a design problem, not a staffing one.
3. Data being carried between systems by a human
This one is easy to spot because people apologise for it while they show you.
The phone system has the call. The CRM has the contact. The calendar has the appointment. The accounting package has the invoice. And somewhere in the middle, a person with a second monitor is copying a name, a postcode and a job type from one into the next, three or four times a day.
Two things make this worse than it looks. The first is that the copying is where errors enter) a transposed digit in a phone number costs a job. The second is that this person is the integration layer, so the integration has holidays.
What the audit produces: a map of every hand-off between systems and who performs it. Not a diagram of the ideal state (a diagram of what actually happens on a Tuesday. This is normally the single highest-value page of the whole engagement, and it very rarely requires AI to fix. It requires two systems being connected properly, which is workflow automation and nothing cleverer.
4. A process that exists only in one person's head
Every business has at least one. Usually it is the person who has been there longest, and usually it is quoting, or scheduling, or deciding which jobs are worth taking.
They are not hiding it. If you ask them, they will explain it) but the explanation takes forty minutes and contains eleven exceptions, and no two explanations are quite the same. It has never been written down because writing it down was never anyone's job.
This matters for AI specifically, because you cannot automate a rule you cannot state. Every failed chatbot project we have been asked to look at afterwards has the same root cause: someone tried to automate a decision that had never actually been defined, so the tool made a plausible guess and everyone lost confidence in it.
What the audit produces: the rule, written down, with its exceptions. Sometimes that is the entire deliverable for that area, and the business fixes it with a checklist rather than software. That is a fine outcome.
What the audit deliberately does not do
It does not produce a tool shortlist. A shortlist compiled before the leaks are measured is a guess wearing a spreadsheet.
It does not produce an ROI model with two decimal places. You can put a defensible range around unanswered calls and slow follow-up; you cannot put a defensible range around "productivity uplift", so we do not pretend to.
It does not produce a recommendation to replace your CRM. Sometimes the CRM genuinely is the problem, but that conclusion should survive a month of thinking, not arrive in week one.
What happens to the findings
The findings become the 12-month roadmap, ordered by impact against effort. The order matters more than people expect. A business that fixes its follow-up before it fixes its answering ends up with a faster response to fewer enquiries. A business that installs a voice agent before its CRM can receive a transcript ends up with a very articulate machine writing into a void.
Roughly:
- Quarter one goes to whichever leak is largest and cheapest to close. Frequently that is not AI at all (it is a routing rule, an integration, or a written-down process.
- Quarter two is usually where the first genuinely AI-shaped job lands, because by then there is somewhere for its output to go.
- Quarters three and four are for the work that only makes sense once the first two are stable, and for revisiting the assumptions, because volumes will have changed.
Build work) the actual construction of an agent, an integration or a pipeline: is scoped and quoted separately once a specific job is ready. That separation exists so the retainer stays advisory and does not quietly become an open-ended development contract with no fixed price.
The uncomfortable finding
The most common single conclusion from a first audit is that the business does not yet need AI for the thing it asked about, and does need to fix something dull first.
That is a difficult sentence to sell and an easy one to justify. A firm that automates on top of a broken process gets the broken process, faster and at scale. The audit exists to catch that before anyone spends money on it.
If you want the short version of this applied to your own operation, the AI Opportunity Assessment is free and takes about forty-five minutes. You do not need to pick a retainer tier to have it, and what comes out of it feeds the first paid day, so Discover starts from evidence rather than a blank sheet.