Capabilities / AI and automation
AI and automation.Only where it earns it.
Somebody in your business spends every Friday afternoon moving the same numbers between the same two systems. That is worth automating. Most of what gets sold as AI is not.
Withheld at the client's request
Francis reads years of Odlings' own trading history to price and plan the parts of the business that used to run on judgement. What it works out for them is theirs, so that is as much as we will say about it.
Prove it on your data.Then decide.
Every AI claim sounds the same until it meets your actual paperwork. So it meets yours first, and you see the number before anybody commits to anything.
We take a real sample
Fifty of your own invoices, not a demo set. If it cannot read yours then it does not matter what it does on somebody else's.
We show you where it fails
Accuracy is a number and you should see it before you buy. The interesting part is always the small percentage it gets wrong, and what happens next.
We agree what a human still does
Anything that touches money or a customer gets checked by somebody. Automation nobody can overrule is a liability rather than a saving.
What we build here
Three jobs, worth doing properly.
There is a long list of things AI can nearly do. These three it can actually do, and they are the three businesses keep asking us for.
Document handling
Invoices, delivery notes, specifications and certificates, read and filed without anybody typing them in. The highest-volume, lowest-glamour job in most businesses, and the one with the clearest payback.
Agents that do the job
Given a task and the tools to finish it: raise the order, chase the missing certificate, reconcile two systems overnight. The model is the easy part - the harness around it is the work.
Estimating and pricing
Models trained on your own quoting history rather than on the internet. They are useful precisely because they have seen what you actually charge and what you actually won.
It has to be right.Or it has to say so.
The difference between something useful and something dangerous is whether anybody can tell when it is wrong. A system that is confidently mistaken costs more than the job it replaced.
Four things are in every build:
Measured, not claimed
Accuracy on your data, written down as a number, before anything goes live. If it is not good enough we will say so and you will not have spent the build.
It shows its working
Every answer comes back with where it came from, so somebody can check it in seconds rather than taking it on trust.
A person where it counts
Anything that moves money, sends something to a customer or changes a record gets a human in the loop. Speed is not worth an apology.
Your data is not the product
Nothing you give us trains a model anybody else benefits from, and we will tell you which provider is processing what before it does.
How it works
One job at a time.
A typical first piece of work looks like:
One call
Tell us what the repetitive thing is. Twenty minutes is usually enough for us to say whether it is worth automating, and quite often the answer is no.
A test
On your own documents or your own history, priced as a small piece of work. You get the accuracy number and the failure cases, whatever they say.
The first job live
The noisiest one, running in weeks, with a human still checking it. We measure what it actually saved before anybody talks about a second.
When you are ready
Tell us the repetitive bit.
Describe the job somebody does every week that nobody enjoys. We will tell you whether it is worth automating, including when it is not.
Technologies
No model loyalty.The cheapest thing that works.
Models change every few months and being tied to one is how you end up paying for last year's. We pick per job, we will move when something better arrives, and often enough the right answer turns out not to involve AI at all.
Where it runs
Where your data goes, and where it stops.
With this kind of work the first question is never how fast it is. It is what leaves the building, who processes it and whether any of it is kept.
AWS
Where models can run inside your own account and region, so the documents never leave infrastructure you control. The default for anything sensitive.
Krystal
UK based and renewable, for the parts that are ordinary software rather than inference. Most of any AI system is still just a system.
Vercel
For the interface people actually use. Fast to load and close to whoever is asking, which matters when somebody is checking an answer between jobs.
Before you ask
The questions we always get asked.
Will our data be used to train somebody else's model?
No, and you will be told which provider processes what before anything is sent. The enterprise tiers of the models worth using contractually exclude training on customer data, and where a job genuinely cannot leave your building we run it somewhere it does not have to.
What happens when it gets something wrong?
It will, which is why the design starts there rather than ending there. Anything touching money or a customer is checked by a person, low-confidence cases are flagged rather than guessed at, and every answer carries its source so a human can settle it in seconds.
Does our data need tidying up first?
Usually less than people fear. Reading messy documents is the thing this technology is genuinely good at, and waiting until the data is clean is how these projects stay permanently three months away.
Is this just ChatGPT with our logo on it?
If that is what a job needs, we will say so and it will be cheap. Most of what is worth building is not: it is a model pointed at your own documents and history, with the plumbing to get answers back into the system people already use.
What does it cost to run?
There is a per-use cost and you should see it modelled before you commit, because it is the part that surprises people. For most document work it is pennies per item, and we will tell you when the volume makes a cheaper approach the sensible one.
What if we want to turn it off?
Then you turn it off, and the process it replaced still exists because we do not delete it on your behalf. Anything that cannot be switched back off has not been built carefully enough.
No pitch, no deck
Tell us what isactually happening.
Describe the mess in your own words. We will say what we think is going on underneath it, and whether there is anything we could usefully do about it.
Unovo LimitedSteel City Stadium, Olympic Legacy Park
Worksop Road, Sheffield S9 3TL
Registered in England and Wales, company no. 16375453