How many hours does AI really free up in an SME? A Granby manufacturer’s numbers

“How many hours will AI save me?” It is a fair question, and most of the answers it gets are useless. An average percentage from an international study says nothing about your shop, your quotes or your bookkeeper.

The only honest answer is built role by role: which tasks fill each person’s week, how many hours they take, and which ones an agent can take over in whole or in part. That is the exercise the management of Les Industries Romy went through, and this is the count, as it appears in the case study.

Les Industries Romy fabricates and installs steel structures for heavy industry, the industrial sector and construction. Like many manufacturing SMEs, it relied on a few key people: estimating, project tracking and accounting all went through them, and information sat in email and network folders.

U.S. tariffs changed the equation. Demand from regular customers dropped and the company had to prospect. But the three project managers had no time to do it: their weeks went to supplier invoices keyed in by hand, follow-ups, coordination and internal tools that did not talk to each other.

Here are the weekly hours spent on the targeted tasks, before and after the agents went live, according to management’s estimate.

The drafter. An agent reads sketches and prepares shop drawings in DXF, with archived versions. The drafter no longer produces the sheets: they review, adjust and approve them. Nothing goes to fabrication without them. According to the case study, the drafter is now free to manage projects 80% of their time.

The project managers. Job tracking, follow-ups, quote preparation, coordination and written summaries are automated. Client relationships and judgment stay human, and that is where the three project managers reinvest their hours. The owner says he now sends roughly two to three times as many quotes. We cover that side in our article on construction quotes.

The table shows where the gain is concentrated: of the 152 weekly hours, 109.2 come from the project managers, roughly seven hours out of ten. That is no accident. Their administrative work was the most fragmented (a status check here, a follow-up there, a summary to write between two calls), and that is exactly the kind of work an agent does well, because it never tires of checking every file every morning. Accounting, by contrast, took only 12 hours a week: the gain is real but modest in absolute terms.

Accounting. Invoice entry disappears. The agent reads invoices as they arrive by email, attaches them to the right job and prepares them for the accounting software. Accounting validates instead of keying in, and job costing builds up as work happens rather than at month end.

Three caveats apply, and we repeat them every time we cite this case.

The share that stays human is not a leftover: it is by design. A price, a schedule or a commitment to a client is always approved by a person. An email to a client is saved as a draft, an invoice requires the pre-tax amount, and management sets the autonomy levels and the dollar threshold above which an agent must wait for approval.

That also explains the residual hours in the table. The project managers’ 10.8 weekly hours and accounting’s single hour are mostly review and approval. An AI project that promises zero residual hours is really promising that nobody will be left to check.

The method can be repeated. For each role, list the repetitive tasks, the hours they take each week and the share an agent could prepare. Then measure for two weeks before deploying anything: without a baseline, there is no measurable gain. Our article on measuring the ROI of an AI assistant lays out the approach.

The whole system runs on the company’s own server, connected to its accounting software, email and files: sovereign AI, with no client data leaving the shop. To see what the same approach would look like in your plant, visit our manufacturing page or get in touch.