Service · Operations
Make the repetitive work disappear
Every business runs on a set of tasks somebody does every week because nobody has had time to fix it. We find those, cost them honestly, and build the thing that removes them.
The test
“Does someone do this every week — and does it go wrong when they're on holiday?”
If the answer is yes twice, it's worth looking at. That's a more useful filter than any list of AI use cases, because it points at the work that is both expensive and fragile. Most companies can name three of these in about ninety seconds.
What we build
Four kinds of problem
Nearly everything we're asked to automate falls into one of these. The examples underneath each are real categories of work we've built, not a wish list.
01
Systems that don't talk to each other
Almost every operational headache is two systems holding different versions of the truth, with a person in the middle keeping them in sync by hand. We replace that person's copy-paste with a pipeline that runs on a schedule, retries on failure, and tells someone when it can't.
- Store ↔ ERP stock and pricing sync
- Orders pushed to a 3PL or carrier automatically
- Accounting exports assembled without a monthly export ritual
- Marketplace and feed data kept current from one source
02
Work that is repetitive but not simple
The tasks that resisted automation for years were the ones needing judgement — reading an email and deciding what it's about, summarising a thread, classifying a complaint, drafting a reply in your voice. Language models handle exactly that class of work now, reliably enough to put into production with the right guardrails.
- Inbox triage: classify, route and draft, human approves
- Support replies drafted from your own policies and order data
- Supplier documents read and turned into structured records
- Product descriptions and metadata generated at catalogue scale
03
Answers that live in five places
Your team already has the information; it's spread across a CRM, a drive, an inbox and somebody's head. An internal assistant grounded in your own documents answers the question directly, with a citation, instead of costing three people fifteen minutes each.
- Internal Q&A over policies, contracts and process docs
- Customer-facing assistant grounded strictly in your real content
- Onboarding that doesn't depend on one person's availability
04
Reports somebody rebuilds every week
If a person opens a spreadsheet on a fixed day to assemble the same numbers, that's not reporting — that's a scheduled job with a human runtime. We build the job, and add the layer that flags what changed and why it's worth looking at.
- Daily and weekly operating numbers, assembled and delivered
- Margin and profitability views across channels
- Anomaly alerts — the number moved, here's the likely cause
How it runs
Prove one, then scale
Nobody should sign a large automation programme on the strength of a slide. We build one thing, put it in production, and let it argue for the rest.
- 01
Automation audit
Typically 1–2 weeksWe sit with the people actually doing the work and map what they repeat. Every candidate gets an honest estimate: hours it consumes now, what it costs to automate, and how long until that pays back. Some things come out of that list marked 'not worth it' — that's a useful result.
- 02
Pilot
Typically 2–4 weeksWe build the single highest-value item first and put it into real use. You get evidence from your own operation rather than a promise, and we both learn how your data actually behaves before committing to a larger programme.
- 03
Roll out
OngoingWith one automation proven, the rest go faster — the plumbing, the credentials and the monitoring already exist. We work down the list in payback order, highest return first.
- 04
Keep it alive
Monitoring includedAutomations rot: APIs change, credentials expire, edge cases appear. Everything we build gets a heartbeat and alerting, so a job that has stopped is noticed in minutes rather than discovered at month-end.
Questions
The questions everyone asks
Is this just ChatGPT with extra steps?
No. A chat window is a tool a person has to remember to use. What we build runs on its own — triggered by an order, an email, a schedule or a webhook — and connects to your real systems, with logging, retries and monitoring around it. The language model is one component inside a system, not the product.
Where does our data go?
Wherever you're willing to let it go, and nowhere else. We scope data handling before anything is built: which data leaves your systems, which provider processes it, what's retained, and what stays entirely inside your own infrastructure. For sensitive material we can keep processing local or restrict it to providers with the contractual terms you need. Israeli privacy law obligations are part of that scoping conversation, not an afterthought.
What if the AI gets it wrong?
It will, sometimes — so nothing is designed around the assumption that it won't. Anything with real consequences is built with a human approval step, and everything is logged so a wrong output can be traced and corrected. Where a task genuinely can't tolerate error, we use deterministic code instead. Choosing correctly between those two is most of the job.
We're a small team. Is this overkill?
Usually the opposite. Small teams feel repetitive work harder because there's nobody to absorb it. The best candidates are almost always tasks one person does every week and that break when they're on holiday — and those exist in a five-person company as much as a fifty-person one.
How do you charge for this?
The audit is a fixed fee. Pilots are fixed-scope and fixed-price. Ongoing work is a monthly retainer covering new automations plus maintenance of what's already running. No per-seat pricing and no percentage of the savings.
What happens if we stop working with you?
You keep everything. Code lives in your repository, runs on your infrastructure, and uses your API credentials. We'll hand over documentation on the way out. Building things you can't leave is a business model we're not interested in.
Name three things you do every week
That's the whole first conversation. Bring three recurring tasks and we'll tell you which are worth automating, which aren't, and roughly what the worthwhile ones cost.