Volume × minutes decides whether a saving exists; whether anyone can state the rules decides whether it can be built. How to rank four candidates in an afternoon.
The stages, what each one costs, what actually moves the price — and the costs that are not on our invoice but belong in your budget.
Measure the baseline before you build, count only hours that leave someone workload, subtract what the automation costs to run, and write the assumptions next to the numbers.
MIT found 95% of enterprise GenAI pilots showed no P&L impact. The four failure modes behind that, all of them decided before any code is written.
The arithmetic on 800 invoices a month, why straight-through rate matters more than extraction accuracy, and the volume below which we would tell you not to.
Deflection and assistance are two different savings that behave differently. How to work out yours from six months of ticket history before spending anything.
If people know which document they need, that is a search problem. RAG earns its cost when the answer spans documents — and citations, not accuracy, are the product.
What actually leaves your systems, the five questions to ask any vendor, the GDPR points that come up on nearly every project, and what we do by default.
Stable rules and structured inputs mean plain automation wins on cost, speed and trust. A model earns its place where the input is messy language — which is usually one step, not the whole process.
RPA drives the screen when there is no API. A model reads what a rule cannot. Most processes need neither first — they need the integration nobody built yet.
Support automates in layers, not all at once: classify and route first, draft second, answer alone last. The order is what decides whether customers notice for the right reason.
Reading from a CRM is easy. Writing to one is where projects fail — on permissions, on field discipline and on the reps who stop trusting what they see.
Four to eight weeks to production is realistic for one process. What stretches a timeline is almost never the build — it is access, undecided exceptions and a missing owner.
Ask what they would refuse to build, what happens when the model is wrong, and who owns the code. The answers separate engineers from demo builders faster than any case study.
A quoted accuracy figure means nothing without the field, the document mix and the definition. What you should ask for is the straight-through rate on your own documents.
A review step that shows everything and asks a person to confirm is not oversight — it is a rubber stamp. Design the human step around the cases the system flags, not all of them.
Model calls are usually the smallest line. What actually costs money after launch is maintenance, monitoring and the exception queue — and only one of those appears on an invoice.
Comparison pages written by one of the parties are usually worthless. These have a column for when the other option is the right one, with real content in it — that is the only thing that makes the rest of the page worth reading.
The crossover is a utilisation number, not a headcount number. Where each option wins, and the hybrid that usually beats both.
Two different shapes of supplier, not two grades of quality. When programme scale genuinely requires the larger one — and when it is overhead.
We build on both. Where per-operation pricing and a forty-node canvas start costing more than code, and the prototype-then-harden sequence.
Most of the time the honest answer is buy. The cases where building wins are specific, and they usually come down to the seams between systems rather than the systems themselves.
Your volume, your minutes per instance, your hourly cost. It shows the annual saving and the payback period, with every assumption on screen and adjustable.
{{ toolBody }}
We use cookies for analytics — to see which pages bring enquiries. Nothing else, and nothing before you agree. Cookie Policy