AI in the installation sector: where does the margin leak away?
In installation companies the money is rarely in purchasing and almost always in the feedback loop. Additional work that is delivered but not invoiced, projects that quietly overrun, and a post calculation that only tells you weeks later what happened. Those are not technical problems but information problems, and that is exactly where an AI agent can do something.
Why is the installation industry a hard case?
Three characteristics make it harder than in most sectors. The margin per project is thin, so a few percent of difference decides whether a year is good or mediocre. The work is done on site, so the information arises outside and comes in with a delay. And the utilization of technicians is the real constraint, not demand.
The result: by the time management sees that a project is running out of line, it has already happened. Steering then becomes explaining.
Where exactly does the margin leak away?
Four places, in order of how much money usually runs through them.
- Additional work that was not invoiced. The hours are written down, the order was given verbally on site, and nobody sends an invoice after it. This is almost always the largest item, and it only stands out in the post calculation.
- Projects that overrun. A project running a week behind costs you twice: your own hours that cannot be used elsewhere, and the risk of a penalty clause at handover.
- Utilization above ninety percent. Sounds good, is a warning. Above that line you have no room left to absorb setbacks and you start hiring in at day rates the estimate never foresaw.
- Price agreements that are not honored. Not out of bad faith, but because nobody puts the purchase invoices next to the agreed volume discounts.
Which systems are usually running?
At installation companies of twenty to two hundred people you usually find a combination from this list. The names matter less than the question of whether you can get at them yourself.
| Type | What you often see | What is in it for an agent |
|---|---|---|
| ERP and project administration | Exact, AFAS, Simplicate, Syntess, Vakware | projects, budget, costs incurred, invoices |
| Time tracking | often inside the ERP, sometimes separate | hours written per project and per technician |
| Scheduling | planning board, Excel, sometimes a separate package | who is where and when |
| Accounting | Twinfield, Exact Online, Moneybird, e-Boekhouden | revenue, costs, outstanding items, payment behavior |
| Estimating and quotes | own templates, sometimes a CRM | won and lost quotes with prices |
Most of these packages have a connection or an export. Where it goes wrong is rarely the technology and usually the question of who holds the key.
What does an agent actually do here?
The most valuable first case is almost always the same: put the hours written next to the signed orders and report the difference per project. That is one connection between two sources you already have, and the answer is a list of projects where work was delivered that was never charged.
After that usually come:
- a weekly alert as soon as the ratio between budgeted and written hours on a project crosses a threshold;
- utilization for the coming three weeks next to the schedule, with a warning before you have to hire in;
- payment behavior per customer, with the five customers that tie up the most working capital;
- lost quotes compared on price and lead time, so you see whether you are too expensive or too slow.
What an agent does not do here: send invoices, place orders or adjust prices. It flags and substantiates, a person decides. That is not only more comfortable, it is also what the governance agreements prescribe as soon as money is involved.
What does it deliver?
Honest answer: you only know after the baseline measurement, and that is exactly why you take one. What we do know is which figures matter in this sector.
- Share of additional work not invoiced. Measure it over the past quarter before you start. This is the number the business case stands or falls on.
- Days between signal and decision. How long does it take today before the leadership team knows that a project is overrunning?
- Hours of digging per monthly meeting. Often half a day from two or three people, and that is the item that disappears fastest.
- Average payment term against the agreement. On twenty million in revenue, the difference between thirty and thirty-eight days is half a million in working capital.
When is this not worth it?
Three situations in which we would advise against starting.
- When time tracking is structurally behind. If it runs three weeks behind, an agent reports old problems. First the rhythm, then the measurement.
- When additional work is not recorded. If orders are given verbally on site without anyone writing them down, there is nothing to compare against. That is a process agreement of one page, not an AI question.
- Below roughly twenty employees. At that size the overview still fits in the owner's head, and that is faster than any system.
How do you go about it?
The same order as in any sector, but with the accents of this industry. Start with the decision scan: ten questions that show whether your biggest bottleneck sits in margin, capacity, cash or decision speed. Then pick one application and prove it in 45 days against a baseline measurement. In this sector that is almost always the additional work alert, because it touches only two sources and the result can be expressed in euros right away.
What you arrange up front: someone who knows the challenge and has two to three hours a week free, access to the ERP and the time tracking (read only), and an agreement on when you stop. Nothing more is needed. The broader setup is in AI for business and the governance framework in AI for leadership.
Frequently asked questions
Does this work if we use subcontractors?
Yes, and it usually becomes more useful. With a fixed pool of freelancers the flow of hours runs through more parties, and that is exactly where additional work disappears in the gap between two administrations.
Do we have to switch packages?
No. We read out what you have and write nothing back into it. If your package has no export or connection, that is a conversation with your vendor, not with us.
What if our data is messy?
It almost always is, and it is no reason to wait. In the first week we measure how complete and how current your sources are. That is often the first time it is stated out loud, and useful in itself.
How long before we see something?
Halfway through the third week something is running that the leadership team can look at. On day 45 there is a decision.
Sources
- The four leak points, the system types and the measures come from our own approach and from the decision scan, which shows per company where the bottlenecks sit.
- The packages named are the names we come across most often in this sector. This is not market research and not a ranking.