AI automation. Pointed at the work that eats your week.
We find the tasks worth automating, build the agents, put them into production and measure the time actually recovered. No data team to hire, no transformation programme, and a baseline recorded before we start so the gain can be proven.
Show me something running in production, and tell me who maintains it
Anyone can demo. A pilot that runs for three weeks is not much harder. What is rare is a system still alive six months after launch, with someone watching it, correcting its drift and making it evolve. Most AI projects in smaller companies never reach that point, which is why the question sorts vendors faster than any credentials slide.
It would be dishonest to ask it without answering it. We build and operate two of our own products. One audits websites for compliance, security and accessibility, then writes the corrections rather than merely listing problems. The other watches sites and catches the failures a simple availability check misses, because a page can return a perfectly healthy status code and still be broken for the person using it.
Both taught us the same lesson twice over. On the monitoring product, one morning the system raised forty-seven alerts and forty-three of them were wrong. A tool that misfires that often gets switched off within a week, and then it protects nothing. We spent far more effort silencing false alarms than detecting real ones, and that is the part nobody sells you.
Four stages, measured at each one.
Not a transformation programme. A map, a priority order, waves of automation with the return measured every time. You keep control throughout.
- 01
Map what actually eats time
Half a day with the people doing the work, not only the people commissioning it. We list the tasks that consume hours, rank them by return, and write down where the data actually lives. You leave with a prioritized plan, whether or not you continue with us.
Half a day, with your operators
- 02
Build one wave, measure it
Two or three agents on the highest-return tasks, wired into the tools you already run, put into production. We record the starting point before building, because without it no gain can be demonstrated and the project gets abandoned for lack of evidence.
The baseline is recorded first
- 03
Stabilize before widening
We watch the edge cases in production, tune what misfires, and only move to the next wave once the current one holds. A system that cries wolf gets switched off within weeks, which is worse than having no system at all.
False positives are the enemy
- 04
Hand over the keys
Documentation, a runbook for when something breaks, and one person on your side trained to operate and adjust. The arrangement can continue as light maintenance or stop entirely. Either way you are not stuck.
Your team keeps control
The hard part is never the demo
A demo lives in the plausible. Production demands the correct. Most of the effort on a serious engagement goes into closing that gap, and the projects that fail are almost always the ones that started from the wrong end.
Three patterns account for most of it. Starting from the technology rather than a named task, so nothing is measurable and the budget runs out before value appears. Feeding a model a body of data that is inconsistent and scattered, which no amount of model quality compensates for. And treating adoption as a consequence rather than part of the work, which produces tools that run correctly and get ignored.
The way around all three is unglamorous: one precise task, a baseline recorded before building, a human kept on the decisions that carry consequences, and the next wave only after the current one holds.
Four questions with straight answers.
-
Which tasks does AI actually handle well today?
Four families hold up in production. Sorting and routing what comes in: messages, forms, scanned documents. Drafting first versions of quotes, standard replies and meeting notes that a human then reviews and signs. Summarizing long documents such as contracts, reports and regulations. And searching a body of knowledge in plain language, so the right document surfaces even when nobody remembers its name. Outside those, results get thin fast.
-
What should never be automated without a human in the loop?
Anything that commits the company: accepting a return outside policy, granting a discount, approving a refund, rejecting an application. Those calls weigh customer satisfaction against margin, and a model has no business context to make them alone. It can propose, with its reasoning and its confidence level, and a person approves. That takes under a minute and prevents the expensive kind of mistake.
-
How do you prove the time saved is real?
By recording the starting point before building anything: how long the task takes today, who does it, how often. Without that number the promise cannot be verified, and unverifiable projects get quietly dropped. We then report three figures monthly: volume handled automatically, accuracy against a target set during scoping, and hours recovered compared with the baseline.
-
Where do our data go?
Wherever you decide, and it is a decision made at the start rather than discovered later. Three levels exist: public interfaces with a contractual framework for non-sensitive content, models hosted inside the European Union for personal or confidential business data, and open-weight models running on your own infrastructure for genuinely sensitive material. We document which model handles what, where the servers sit, and what never leaves your side.
Name one task that eats your week.
Twenty minutes to check whether it is a good candidate. If the honest answer is that a simple script would do it, we will say so.