How you know it is working

The fastest way to find out whether an AI project is real is to ask two questions. What number did it move, and where does that number live?

Not what the model can do. Not how the demo went. What number, and where. If the number lives inside the vendor's own dashboard, be careful. If it lives in a system you were reading long before AI arrived, you are probably looking at something real.

That test sounds too simple to matter, so it is worth saying why it works and why most of the market fails it. In its 2025 State of AI in Business report, MIT found that about 95 percent of enterprise generative AI pilots produced no measurable impact on profit and loss (MIT NANDA, 2025). The word to sit with is measurable. Some of those pilots failed on integration. Plenty failed on something more basic: nobody agreed, before the work started, on which number would prove the thing was working. So the pilots ran, the dashboards glowed, and the business stayed exactly where it was. Measurement was not a detail those companies skipped at the end. It was the discipline missing from the start.

Activity is what vendors sell

Open the reporting screen of almost any AI tool and you will meet the same family of metrics. Tasks completed. Hours saved. Automations run. Conversations handled. Time reclaimed.

Notice what those numbers have in common. Every one of them measures activity, proof that the software did something, and none of them measures whether your business changed. That is not an accident. Activity is what vendors sell, so activity is what vendors report. It is also the kindest possible scoreboard, because activity only goes up. Nobody has ever opened a vendor dashboard and found it apologizing.

Here is the tell. You do not run your business on any of those numbers. You have never made a payroll decision because an automations counter ticked over, and you have never lost sleep over tasks completed. The numbers you actually run on live somewhere else, which is exactly why the vendor's numbers get offered to you. They are impressive precisely because you cannot check them against anything. Hours saved is an estimate you cannot audit. Tasks completed is a count of motion, not of consequence.

None of this means the tools are lying. It means the scoreboard answers the wrong question. The question is never whether the AI is busy. It is whether the function is running.

The numbers that count predate the AI

The numbers that count are the ones you were already reading before AI showed up. How fast leads hear back. What happens to a customer at nine at night. How old your receivables are. What moves through the week without someone chasing it.

Those numbers have three properties the vendor's dashboard can never have. They live in your systems, your CRM, your inbox, your accounting software, so nobody can curate what you see. They existed before the project, so there is a real before to compare against. And they connect to money and customers through short, obvious chains, so when one of them changes, you know what it means without anyone translating.

If a function is really running, it shows up in the boring numbers you already watch. If it does not show up there, it is not really running.

That is the whole philosophy, and it is deliberately unglamorous. No new dashboard to learn. No invented metrics. No ROI theater. The proof of an AI implementation should look exactly like the proof of a good hire: the ordinary numbers of the function, read in the ordinary place, looking different than they used to.

What to watch, function by function

Ensolve implements across six functions, marketing, sales, customer service, operations, finance, and HR, and each one has its own version of the boring numbers. Four of them cover most of what owners ask about.

In sales, watch response. How many leads got an answer, and how fast. Follow ups that went out versus follow ups that should have gone out, which your CRM has always been able to show you and almost nobody looks at. Speed to lead and follow through are the two numbers that decide whether the demand you already paid for becomes conversations, and both are sitting in the system your team opens every morning.

In customer service, watch coverage. What got answered after hours, when nobody was on the clock. Which questions stopped waiting until morning. And just as important, what reached a human that should have, because a service function is only healthy if the judgment calls still land with people. The goal is not fewer humans in the loop. It is nothing sitting in the queue that did not need to.

In finance, watch age and cadence. How old your receivables are, and whether the tail of old invoices is shortening. How long the close takes, and how much of it is a person retyping numbers between systems. Finance is where manual drag hides in plain sight. The average executive loses around sixteen hours a week to manual administrative work (ServiceNow, State of Work), and a good share of those hours sit near the ledger, chasing, reconciling, and rekeying.

In operations, watch what moves on its own. Work that advanced without a person pushing it. Handoffs that used to stall and no longer do. The category of things that used to fall through, missed confirmations, unlogged jobs, orders waiting on a reply, quietly getting smaller.

Every one of those numbers existed in your business before AI, which is the point. You do not need permission to check them, you do not need the vendor to export them, and nobody can dazzle you with them, because you already know what normal looks like.

One function at a time is what makes the numbers readable

There is a reason this measurement philosophy and our delivery model are the same idea. Ensolve implements one function first, proven in your own numbers, and then expands. That is usually described as a risk decision, and it is. It is also a measurement decision.

When ten initiatives launch at once, attribution dies. Something moved, nobody can say what moved it, and the debrief becomes a negotiation. The research suggests most companies are living in exactly that fog. MIT found that while over 80 percent of organizations had piloted tools like ChatGPT or Copilot, only around 40 percent deployed them into operations (MIT NANDA, 2025). A pile of pilots is not a strategy, partly because a pile of pilots cannot be read.

One function at a time makes the effect legible. Before the work starts, you and we look at the same numbers, in your systems, and agree on which ones this function should touch. Then the function goes live inside the tools your team already uses, with access you granted and can revoke, and the same numbers get read again. Not our export of them. Yours, at the source. This is why the forward deployed engineer model matters here: the engineer works inside your business, in your systems, so the readout happens where you can see it. And the judgment calls stay with your people the whole way, so what you are measuring is genuinely the function, not a black box you have been locked out of.

When the numbers do not move

Now the part most pitches skip. What if the numbers do not move?

We will not promise you they will. Anyone who guarantees a specific movement in your numbers before touching your business is guessing on your behalf, and you should treat the guarantee as a sales instrument, not a forecast. What we will commit to is the standard: the function is judged in your numbers, or not at all.

So if the numbers do not move, that is a conversation we have with you, in plain terms, looking at the same screens. Sometimes the readout says the function needs rework, and reworking it is our job, not a change order and not your homework. Sometimes it says the function was pointed at the wrong number, and the aim gets corrected. What never happens is the failure getting relabeled as success by switching to a friendlier metric. When the numbers disagree with the pitch, the numbers win. That rule binds us more than it binds you, and it is what keeps the whole engagement honest.

It also keeps the money conversation clean. Pricing is a direct conversation in the mapping call, scoped to the function, with no surprise scope, and it is never a wager on results. You are not buying a bet. You are buying a function, run properly, and the standing right to check it against reality.

Judged in your numbers, or not at all

Hold this standard over everyone, including us. Ask any AI vendor, any agency, any pitch that lands in your inbox: what number will this move, and where does that number live? If the answer is their dashboard, you have learned what you needed to know. If the answer is your CRM, your ledger, your queue, keep listening.

This is the measurement half of an argument we have made before: you are buying outcomes, not software, and an outcome is only real where your numbers live. If you want the operating half, what an implementation actually looks like walks through it step by step. And if you are wondering which boring numbers we would start with in your business, sales and finance are usually where the readout is sharpest, because the numbers there are the ones you already check first.