Automation that clears the work,
until a person should.
We automate the high-volume half of a workflow and staff a trained seat on the half that needs a person. You decide where the line between them sits.
- You set the confidence threshold per workflow. We staff the queue below it.
- Money leaving the business and anything a customer is upset about never clear automatically, whatever you set.
- Every action is reconstructable afterwards — input, score, decision, who approved it, when.
AI does not rescue a broken process. It accelerates it.
Most automation that stalled did not stall on the technology. It stalled on one of these three, and all three are diagnosable in an afternoon.
- The wrong step got automated
The constraint was somewhere else, so the process got faster at the one point that was never holding it up. Nothing downstream moved.
- It worked until an input was unusual
Rule-based automation has no behaviour for the case nobody wrote a rule for. It either stops and waits, or proceeds and is confidently wrong.
- The exception queue became someone's job
The work moved rather than left. Nobody decided who owns the leftovers, so they found a volunteer — usually your most senior person.
Four controls, and you hold all four.
Everyone in this market says there is a human in the loop. Far fewer say where that human sits, what wakes them, or how you check the machine's work afterwards.
Volume, matching, extraction, routing — the same decision made the same way ten thousand times.
Exceptions, consequence, tone, and every call you would want a name attached to.
The confidence threshold
Each item gets a score, and you set the line it has to clear. Above the line it goes through. Below it, a person picks it up. Start the line high and lower it as the record earns it — that direction is much easier to defend than the other one.
The always-hold list
Categories that go to a person no matter how confident the score is, because the cost of being wrong is not proportional to how often it happens. This list is agreed before go-live and it is not ours to edit.
The escalation path
Who gets a held item, in what order, and what happens when nobody answers inside the window. Written down before go-live rather than discovered during the first incident, and the last hop is always a named person on your side.
The audit trail
Input, score, decision, actor and timestamp, for every item — the ones a person touched and the ones that cleared on their own. If a decision cannot be reconstructed six months later, it was never really a control.
None of the four is a feature we can turn off to hit a number. They are the reason this is sellable at all: automation you cannot inspect is a liability with a good quarter behind it.
Walk through these on a callFive services, one control surface.
The four controls apply to every line below. What changes is how much we build and how much we then staff.
AI Automations
One workflow you name: we automate the genuinely repetitive part and put a threshold and a staffed queue around the rest. Built on the tools you already run.
- Document and invoice capture, coded and filed
- Inbound triage, tagging and routing
- Record updates from a known template
- Reconciliation, with the mismatches surfaced
What you getThe repetitive volume off your team's day, and a written record of every exception a person had to make a call on.
AI Implementation
You bought the tool and it is not landing. We do the configuration, process design and training that turns a licence into something people use.
What you getThe tool in production, with the people around it trained.
AI Agents & Chatbots
Answers your documentation already contains, handled automatically. Anything outside it handed to a person mid-conversation, not looped back to the menu.
What you getRoutine questions answered, and a real handover for the rest.
Workflow & Systems Integration
A step is usually still manual because two systems do not speak. We connect them, and where no integration exists we staff the gap rather than pretend it closed.
What you getOne flow end to end, and honesty about the seams.
AI-Assisted Back Office
Our back office seats with the volume already automated underneath them — so the seat spends its day on judgment rather than typing.
What you getFewer hours billed for the same work handled.
One workflow first. Always one.
Nobody should buy an automation programme. Buy one workflow, watch it for a month, then decide about the second.
Map it
We follow one item end to end and write down where the time goes, which steps are automatable, and which should not be. You keep the map whether or not you buy anything.
Pilot it
The threshold starts deliberately high, so most items still route to a person. The staffed half runs from day one, which is what makes the automated half safe to widen.
Widen it
Lower the threshold as the audit trail earns it, then scope the next workflow on its own merits. No platform commitment, and no bundle where a weak workflow hides behind a strong one.
The volume and the exception, function by function.
The split is the same everywhere. What differs is which items you cannot afford handled by a confident guess.
Customer support
Routing, tagging, order-status and delivery questions, and anything your help centre already answers.
Anyone already unhappy, anything owed money back, and any conversation where the next reply decides whether they stay.
Finance operations
Invoice capture and coding to known vendors, statement matching, and the reminder schedule nobody runs.
Releasing any payment, a mismatch with no obvious cause, and anything an auditor would later ask you to explain.
Back office
Data entry and validation, document classification, and record updates following a known template.
A document that does not match its template, and any record where two sources disagree about the truth.
Sales operations
Lead routing and enrichment, CRM hygiene, meeting logistics, and follow-up that stops when a human replies.
Pricing, anything a prospect pushed back on, and anything that changes a commitment you have already made.
The questions worth asking before you automate anything.
What happens when the model is wrong?
It will be, which is why the threshold exists. An item the model is unsure about never clears on its own — it goes to a seat. For the items that do clear, the audit trail is what makes a wrong one findable and fixable, and a pattern of them a reason to raise the line. We would be more worried about a provider who answered this question with an accuracy figure.
What will you refuse to automate?
Money leaving the business. Anything a person is legally required to attest to. Anyone who is already upset. Any decision about somebody's employment. Those are on the always-hold list at the start of every engagement, and raising a threshold does not reach them.
Is our data used to train anything?
No. Your data is used to run your workflow. Access is least-privilege and scoped per client under ISO 27001 controls that are independently audited rather than self-declared, and where a workflow runs through a third-party model we will tell you which one and what its retention terms are before you approve it.
Does this replace people on our team?
It replaces a category of task, and in our experience the task it replaces is the one nobody wanted. What we will not do is tell you a headcount number, because the honest answer depends on what your team does with the hours back — and that is your decision, not an output of our model.
What if the automation breaks, or a vendor goes down?
The workflow falls back to the staffed queue, which is the same queue already handling every held item. That is a real advantage of a staffed loop over a pure software one: there is no state where the work simply stops, only a slower state. The fallback is documented and tested before go-live rather than after.
How will we know it paid for itself?
Phase one gives you the baseline — how long the workflow takes now and who spends the time. After that, the volume that cleared automatically and the hours the seat billed are both in the record, so the comparison is arithmetic rather than a claim. If it does not clear, that shows up in month one and the pilot was deliberately small.
Do we need to buy new software?
Usually not. Most workflows we are asked about are already sitting on tools the client pays for and half-uses. Where something genuinely is missing we will say so and you will own the licence directly — we do not resell tooling, which keeps the recommendation honest.
Name the workflow. We will map it for free.
One workflow followed end to end, and a written map of where the time goes and which steps should never be automated. You keep the map either way.