INSIGHTS

AI for water treatment and hygiene firms: where it actually helps

AI for water treatment and water hygiene firms: practical ways to improve job notes, reports and office handovers without handing over technical judgement.

By Harrison Jones · · 7 min read

Picture an engineer finishing a water hygiene visit. The work is done, but the office still needs the job notes, a photograph needs matching to the right asset, and the customer wants an update. Someone has to turn those loose ends into a usable record.

That is where I would start a conversation about AI for water treatment businesses: with the work between the visit, the report and the invoice. An autonomous treatment plant can wait.

This article is about improving that administrative flow. It is not advice on treatment settings, water safety decisions or technical assessment. My position is straightforward: let technology help people organise the evidence, and leave deciding what it means to the people qualified to judge it.

AI for water treatment starts with the handover

For a practical starting point, imagine a water hygiene contractor whose engineers use a job app, whose office keeps an Excel tracker, and whose customer emails land in a shared Outlook inbox. This is an illustrative scenario; the business is made up.

Follow a completed visit through that business. Where does someone copy information? Where do they stop because a reference is missing? Who decides that the report is ready to send?

I would separate the possible fixes into distinct jobs:

  • Process change: agree what the engineer hands over and who checks it.
  • Ordinary automation: move an approved record, send a reminder or update a status using a defined rule.
  • AI assistance: suggest a summary, extract information from a document or sort an incoming message when its wording varies.

Treat these as design choices you make for your own office. If a required field would stop the chasing, use the required field. If the office cannot tell which job a photograph belongs to, fix the reference before adding another tool.

My preferred starting question is: what would let the next person do their job without asking the previous person to explain it again?

Turn job notes into a draft that keeps the engineer's account

In the illustrative contractor, suppose an engineer writes a rough note about access, the work completed and an outstanding customer query. A useful trial would ask AI to organise that text into a consistent draft for the office to review.

There is an existing product example of the underlying capability: Microsoft documents work order summaries in Dynamics 365 Field Service, drawing on work orders and related records, including notes and asset information. It shows a summarisation feature exists. Keep your current job system unless you have other reasons to change it.

For a trial, I would give the draft a fixed structure:

  • Work recorded as completed.
  • Access problems or other limitations recorded by the engineer.
  • Outstanding questions.
  • Proposed office follow-up, clearly separated from the original observations.

The instruction should be to organise what is present and flag what is missing, leaving the gaps visible.

Keep the original note available beside the draft. Ask the reviewer to look for changed meaning as well as spelling. “Unable to access” must not become “checked”; “customer asked about a return visit” must not become a booked appointment.

My rule is simple: I would never treat fluent wording as proof that a job record is accurate. Keep the engineer's observations as the source, and have a named person approve the customer-facing version.

Extract document details without losing the source

For another illustrative use, imagine the office receiving supplier invoices as PDFs and manually transferring the supplier name, invoice reference and purchase order reference into a tracker.

Microsoft's AI Builder document-processing documentation describes extracting information from documents and using a trained model within Power Automate or Power Apps. That makes document extraction a reasonable capability to investigate. It does not establish that your particular forms will be read correctly.

I would start with a narrow document type and define the fields before testing anything. For an invoice, those might be the supplier, invoice reference and purchase order reference. Keep each extracted value linked to the document it came from.

Then agree what happens when the result is uncertain. My preference is a visible review queue, so nothing gets guessed and entered quietly into the live system. If the purchase order reference cannot be matched, stop that record for a person to check.

I would keep this first trial away from interpreting laboratory results or making technical recommendations. Those are separate decisions with their own risks, and they shouldn't creep in because the extraction tool is already open.

Also ask whether extraction is needed at all. If a supplier can provide a usable structured export, I would test that before building an AI step around a PDF. The aim is less handling.

Use inbox triage to suggest the next owner

Consider a shared mailbox with booking requests, invoice queries, document requests and messages about incomplete visits. For this hypothetical business, I would test whether suggested categories help the office find the right next action.

Microsoft lists text classification and information extraction among AI Builder's Power Automate capabilities. Those are the relevant building blocks here. Whether a ready-made system understands your customers or your service priorities is something you would still have to test.

Start with a small set of categories the office actually uses. Give each category an owner and an expected action. Include an “unclear” route rather than forcing every message into a confident label.

During the trial, show the original message and the suggested category together. Let the office accept or correct it before anything is routed. I would keep outgoing replies under human approval and leave existing urgent-contact arrangements untouched.

Judge the trial on whether messages reach the right person with less handling. Do not judge it on the quality of a generated reply. A polished answer to a message nobody owns leaves the operation exactly where it was.

Fix the record before buying the assistant

Before spending money, I would check whether the business can reliably connect a customer, site, asset and visit. For our illustrative contractor, that means agreeing how those references are created and carried through the job.

Ask the team to walk through an awkward completed visit together. Open the job record, find the attachments and trace the report back to the engineer's notes. Make a list of every point where someone has to explain what a field or filename really means.

Microsoft's summary feature provides references back to source records and requires specific setup for UK environments. I would take both details seriously when evaluating it: test the route back to the evidence, and check availability and setup rather than assuming a feature is already included and enabled.

My proposed minimum for any pilot is a consistent job reference, an identifiable source record, a visible review status and a named owner for exceptions. Add a way to correct the result without overwriting the original evidence.

If those basics are missing, start there. Our Assess stage is about understanding that operational friction before deciding what to change. The AI readiness check for non-tech firms is a useful companion if you are still deciding whether AI belongs in the answer.

Run a bounded trial and count the checking

This week, choose a routine administrative handover. Keep it small and away from the most technically sensitive work. A draft office summary from completed job notes is a sensible candidate to evaluate.

Use approved test material and keep the existing process running while you compare results. Write down what a correct output looks like before trying the tool. Include incomplete notes and awkward wording alongside the clean examples.

For each attempt, record the preparation, review, correction and re-entry work. Note whether the next person could proceed and whether anything important was omitted or changed. These are proposed measures for your trial; they don't promise any result.

Agree in advance what would make you stop. I would stop a trial that repeatedly changes technical meaning, hides missing information or creates more checking than it removes. Give someone responsibility for reviewing it after launch, too; approval to start doesn't cover running unattended indefinitely.

My view is that AI has to earn its place at the handover. A clearer form, a dependable connection between systems or a better ownership rule may be the better answer. Use whichever makes the work easier to complete correctly.

If your engineers have finished the visit but the office is still piecing the job together, let's talk about that handover.

This article is general information, not legal, tax or regulatory advice. Check the linked sources before making a decision.

Want a second pair of eyes on your operation?

Thirty minutes with a founder. Tell us how the work runs and where it gets stuck, and we will tell you honestly whether we can help.

Book a conversation