Underwrite AI like any capital expense

No investment committee would fund a new production line because it promised to make the plant more efficient. It would ask what the line costs, when it starts producing, what it returns, and what happens if it falls short.

AI line items often arrive with none of that. A license here, a pilot there, a budget for experiments. A year later the spend is real and the return is a story.

AI spend reaches EBITDA when each initiative is underwritten like capital, with a baseline, a mechanism, a measure, an owner, and a stop rule.

Stalled AI spend breaks in 3 familiar places.

Tools added on top. A license is bought and the old process keeps running beside it. Customer service gets an assistant that drafts replies, and the team keeps writing from scratch because the drafts cannot see order status. The company now pays for both, and nobody can name the work that went away.

Adoption assumed. The business case counts every seat as a user. By the third month a handful of people use the tool, and the usage report counts logins, which tell nobody whether any work moved.

No attribution. When a number improves, nobody can say whether the tool did it. Margin rises in the quarter a price increase lands, and the AI program claims a share. Without a baseline the gain is an opinion, and at exit a buyer will not accept an opinion as support for a run-rate adjustment.

Each break has the same cause. The spend was approved as an experiment, so nobody wrote down what it had to do. The most expensive AI program is the pilot that never ends.

The fix is a one-page memo, written before any money is spent, in the same terms as any other capital request.

The work it replaces. Name the work or the software that goes away. If nothing goes away, the initiative is a cost, and it should be judged as one.

Baseline. Measure the work before anything is built, over enough weeks to be fair, in the units the work runs on: hours, days, errors, dollars.

Mechanism. Describe how the work changes, step by step, and which system it runs in.

Measure. Pick one number that moves if the mechanism works, with a target and a date.

Owner. Name the person in management whose number it is. The owner can change the process and reports the measure at every review.

Cost. Count the build, the run cost, and the hours the company's own people will spend. Mark which costs recur, because a buyer prices those at a multiple.

Path to EBITDA. State how the gain reaches the P&L: more volume on the same cost, a price that is realized, or a cost that is removed. Hours saved are a means, and they count when they change one of those lines. A gain in inventory or receivables reaches cash, and the memo states it in cash.

Stop rule. Set a threshold and a date. If the measure misses the threshold by the date, the work stops, and no more money goes in.

The stop rule is the easiest field to leave blank, and it matters most. Agreed before launch, while nobody is attached to the result, it turns a hard conversation into a scheduled one.

Reviews. Put the review dates on the calendar before launch, with the owner presenting the measure against the baseline.

If nothing goes away, the initiative is a cost, and it should be judged as one.

Take Sample Co., a fictional distributor whose inside sales team answers quote requests by email. The case is illustrative, and its numbers are synthetic.

Over 8 weeks of baseline, the team sent 640 quotes a month. Each took a seller 45 minutes: the contract price from the ERP, stock and lead time from another screen, then a quote typed into the CRM. That is 480 seller hours a month spent on quoting.

The agent reads each request, pulls price, stock, and lead time from the ERP, and drafts the quote in the CRM. A seller checks and approves a draft in about 8 minutes.

Some requests are too unusual to draft, so the target is 15 minutes across all quotes, which assumes about 4 in 5 go out from a draft. At that rate the team gets 320 hours a month back, about 2 sellers' time.

That time reaches EBITDA as growth on the same cost base. Quote volume is rising, and this year's budget adds 2 inside sellers to keep up, at $190,000 a year of loaded cost. With the agent, the current team handles the volume, so sales grow while selling cost stays flat.

Net of $3,000 a month to run the agent, EBITDA runs $154,000 a year above the plan. The $85,000 build is a one-time cost, expensed as it is spent. Counting 120 hours of the team's time, valued at $9,000, the build pays back in about 7 months.

If the company would rather use the time to win more business, the memo says so, and the measure becomes revenue per seller.

The measure costs nothing to collect. Every draft and every approval is logged, so seller minutes per quote and the share of quotes sent from a draft come out of the system each week.

The stop rule is set before launch, with 2 tests at week 8. If fewer than 60 percent of quotes go out from an agent draft, the agent is switched off and the spending stops. The same happens if seller minutes per quote are above 30.

The floor sits where the case still pays. At 30 minutes a quote, the time returned covers 1 of the 2 hires, and the build still pays back inside 2 years.

Here is the whole case on one page, written before any money was spent.

In this case, the week 8 review shows 73 percent of quotes going out from agent drafts at 18 minutes a quote, which already returns 288 hours a month. Both numbers clear the stop rule, and the 15-minute target is due at week 12, so the work continues.

Had either number missed, the decision would already have been made, by people who agreed to it before they were attached to the result.

Before an operating partner approves any AI line item, the team should be able to answer 7 questions.

1. What work or software does this replace, and when does it go away?
2. What is the baseline, and who measured it?
3. Which number moves, by how much, and by when?
4. Who owns that number, and can they change the process?
5. How does the gain reach EBITDA or cash: more volume on the same cost, a price realized, a cost removed, or working capital released?
6. What is the stop rule, and who will call it?
7. Are the review dates on the calendar?

Each is a question an operating partner already asks of a new plant, a new hire, or a new location.

An initiative that answers all 7 is underwritten. One that cannot is a pilot, and a portfolio of pilots is a cost line with a good story.

Across a portfolio, the memos become a register: every AI line item, its owner, its measure, and its status, reviewed each quarter with the rest of the value creation plan. That is how AI spend shows up where a sponsor can see it, in EBITDA.